In case you missed it: There’s an announcement at the end about what’s coming next for Card Catalog.

Maybe it starts with a link someone sends, or a documentary that autoplays after the one we meant to watch. Maybe it’s a thread that shows up in our feed at midnight when we’re still scrolling. The subject could be anything: the pharmaceutical industry burying research, the government hiding evidence about UFOs, a mass shooting that doesn’t add up, the financial system operating in ways the public was never meant to understand.

The specifics almost don’t matter, because the experience of encountering certain kinds of conspiracy content follows a remarkably consistent pattern. Not all conspiracy theories work this way, and some are easy to dismiss on contact. But the ones that spread farthest and hold on tightest tend to share a specific architecture: they mix documented, verifiable information with conclusions that are unprovable or unsupported, and that mixture is what makes them so difficult to evaluate.

At first, the claims track with things we can check. Congressional hearings and court documents that can be verified. Documented institutional failures covered by mainstream news outlets and sitting in public archives. The tone is measured, the sourcing looks solid, and the early arguments make the kind of sense that doesn’t require any leap of faith at all. We find ourselves nodding along, maybe even feeling a little surge of discovery, that satisfying click of puzzle pieces fitting together.

And then, at some point we can rarely pinpoint afterward, the ground shifts. The sourcing gets thinner and the connections get longer, and the claims start requiring us to accept things we can’t verify, built on the credibility earned by the things we could. The narrative becomes cleaner and more all-encompassing in its explanations, and the questions we’d normally ask (What’s the counterargument? What evidence would disprove this?) start to recede, because the framework has built enough momentum that questioning any part of it can feel like rejecting the parts we already agreed were true.

The boundary between “that’s a fair point” and “that’s a stretch” can be difficult to locate in the moment, especially when the early parts of the framework were well-sourced. Conspiracy content can leave behind a specific kind of ambiguity: the sense that some of what we encountered was solid, without a clear method for determining which parts were and which parts weren’t. That ambiguity is where information literacy becomes most relevant, because the skills for sorting verified claims from unverified conclusions are the same skills that apply to evaluating any information source in a complex environment.

What makes something a conspiracy theory

We tend to recognize conspiracy theories when we see them, but articulating what separates a conspiracy theory from a legitimate investigation or a proven fact requires more precision than a gut feeling can provide. A widely cited academic definition comes from philosopher Brian Keeley, who describes a conspiracy theory as a proposed explanation of events that attributes their cause to the secret actions of a small, powerful group. What makes something a conspiracy theory specifically is that the explanation is circulating, often with a framework of supporting evidence assembled around it, before the claims have been independently confirmed.

Conspiracy theories tend to be deeply social. They circulate through communities that build shared frameworks for interpreting events, creating networks of people who feel they have access to knowledge that the mainstream is missing or suppressing. This social dimension can also become isolating, because people outside the community may dismiss the theorist’s concerns entirely - which can push the theorist deeper into the only group that takes their questions seriously.

What makes the concept particularly slippery is that sometimes these proposed explanations turn out to be right. Powerful people do sometimes coordinate in secret, and the truth does sometimes get suppressed. Researchers have identified several structural features that show up consistently across conspiracy content and that distinguish conspiratorial frameworks from other forms of skepticism or investigation:

• The theory ties together events or observations that appear unrelated on the surface and presents them as part of a single, coordinated pattern.

• The mainstream or official explanation is framed not just as incomplete but as part of the concealment. The institutions providing the official account are cast as either complicit or compromised.

• Counter-evidence doesn’t weaken the theory. Instead, it gets absorbed as further proof of the cover-up. The absence of evidence is treated as evidence of how well the secret is being kept.

• The theory relies on the idea that a small group of actors is secretly orchestrating events that affect large populations, and that this orchestration is being deliberately hidden.

The difficulty is that these same patterns have also shown up in claims that were later confirmed as true. The structural features of a conspiracy theory and the structural features of an accurate accusation of institutional deception can look nearly identical before the evidence surfaces. In several well-known cases, people who made claims matching every feature on the list above were dismissed before the evidence proved them right:

• MKUltra: Throughout the Cold War, rumors circulated that the CIA was conducting psychological experiments on unwitting subjects. These claims were dismissed for decades before declassified files and the 1977 Senate hearings confirmed that the CIA had run a covert mind-control program.

• COINTELPRO: Civil rights activists and political organizers in the 1960s reported being surveilled, harassed, and infiltrated by government agents. Despite these documented accounts, the FBI denied the existence of any such program, and the claims were broadly dismissed by mainstream institutions until stolen FBI files and a subsequent congressional investigation confirmed that the bureau had run a systematic program to disrupt domestic political organizations.

• The Tuskegee syphilis study: Long-held suspicions in Black communities about government medical experimentation were confirmed in 1972 when investigative journalism exposed a 40-year U.S. Public Health Service study in which Black men with syphilis were deliberately left untreated and misled about their care. Tuskegee remains one of the most cited reasons for institutional distrust in American healthcare.

• The tobacco industry cover-up: Public suspicion that tobacco companies knew smoking caused cancer and were hiding the evidence circulated for years before internal documents released during litigation confirmed coordinated, decades-long campaigns to suppress and discredit the science.

• NSA mass surveillance: Claims about the scope of government surveillance of citizens were widely dismissed before Edward Snowden provided documentation in 2013 confirming the extent of the National Security Agency’s domestic surveillance programs.

In each of these cases, the institutions being accused actively denied the allegations, and the confirming evidence was classified, sealed, or otherwise inaccessible until it wasn’t. These cases gave conspiracy theories a foundation of precedent to draw on: institutions have coordinated in secret and denied it before, so the claim that they’re doing it again doesn’t sound unreasonable on its face. Each confirmed case reinforces the idea that conspiracy theories can be true, which lends plausibility to the category as a whole, even when any individual theory lacks supporting evidence. This functions like a shift in the Overton Window, a concept from political psychology: historical precedents act as proof of concept, moving claims that would otherwise sound extreme from the category of “impossible” into the category of “historically demonstrated to be possible.”

Why the confusion runs so deep

Some conspiracy theories are entirely fabricated, and those tend to fall apart quickly under scrutiny. But the ones that gain traction often contain verified, checkable information alongside conclusions that aren’t supported by the evidence. Once we’ve confirmed that some of the claims in a framework are true, something shifts in how we evaluate everything else alongside them.

When we verify that claim A is accurate and claim B checks out, claim C arrives carrying the credibility that A and B earned, even if C is a completely different kind of assertion with a completely different evidence base. A could be a documented historical event. B could be a verified piece of congressional testimony. C could be a speculative interpretation of what A and B mean when we put them together. But because C is sitting right next to A and B, and because A and B were solid, C inherits a sense of reliability it hasn’t earned on its own merits. In psychology, this is driven by the halo effect, where our positive evaluation of one part of a framework influences our perception of the rest, and what researchers call source credibility transfer: the trust earned by verifiable facts spills over and artificially shelters the unverified speculation sitting alongside them.

That slide from “these parts are verified” to “so the whole framework must be valid” is what makes this kind of conspiracy content so much harder to navigate than straightforward misinformation. The natural logic feels airtight: If these things are true, wouldn’t the rest of the story be true too? In formal logic, the assumption that a valid conclusion must follow from a set of true premises, when the conclusion doesn’t logically follow, is called a non sequitur. But verified facts don’t guarantee that the conclusion drawn from those facts is also correct, because the conclusion is an interpretation of the facts (and the interpretation can be wrong even when the facts are right). But that principle is much easier to understand in the abstract than to apply in the moment, because our brains process narrative coherence as a signal of accuracy. Psychologists call this the narrative fallacy: when a story hangs together well, we’re inclined to treat the coherence itself as evidence that the story is correct, even when the coherence comes from editorial choices about what to include and what to leave out.

Conspiracy content is fundamentally different from a simple lie. A lie can be caught. A framework built on documented evidence, with unproven conclusions woven in among proven facts, asks us to do something cognitively much harder: to hold two things at once. Some of this is true AND the conclusion doesn’t follow. The evidence is documented AND the interpretation is unsupported. Most information we encounter doesn’t ask us to do this kind of sorting. Conspiracy content demands it constantly.

And the confirmed conspiracies listed above deepen this confusion further. Each one functions as a domino: once we learn that the CIA did conduct secret mind-control experiments, or that tobacco companies did coordinate campaigns to suppress cancer research, the ground shifts under future dismissals. If that was true, what else could be true? The confirmation of one conspiracy doesn’t just validate that specific theory; it can widen the field of plausibility for theories that follow. The result can be an ever-widening circle of openness to conspiratorial explanations, where each confirmed precedent makes the next unconfirmed theory feel more plausible because the pattern of “they lied to us before” is documented and repeated. The challenge is that the pattern is documented without every new application of the pattern being supported by evidence, and holding those two things simultaneously is precisely the kind of cognitive work that conspiracy content makes difficult.

The proximity problem and the narrative fallacy aren’t the only cognitive tendencies at work. Information literacy as a discipline studies how people evaluate sources, weigh evidence, and reach conclusions, and the research has identified several specific ways that process can break down when the information environment is complex or deliberately constructed to persuade. The conspiracy frameworks that gain the most traction tend to engage multiple tendencies at once, which compounds their persuasive effect:

• Confirmation bias in source evaluation. When we’ve already reached a conclusion, we evaluate supporting evidence generously and contradicting evidence harshly. Conspiracy frameworks take advantage of this by providing a pre-built conclusion and surrounding it with evidence pre-selected to confirm, so everything that contradicts the conclusion has already been filtered out before we ever encounter the framework.

• Pattern completion. Our brains are wired to find structure, a cognitive tendency known as illusory pattern perception. When a handful of data points line up, we instinctively connect them into a pattern and assume the pattern is meaningful. But data points can always be linked if we’re selective about which ones we include, a logical error known as the Texas Sharpshooter Fallacy: clustering random data points together to invent a pattern, then drawing a bullseye around them. The completeness of the pattern is an illusion generated by aggressive cherry-picking. A conspiracy framework that presents only the evidence supporting its conclusion can create a pattern that feels overwhelming precisely because everything contradicting it has been left out.

• Narrative coherence registering as truth. Good research often produces unsatisfying answers full of uncertainty. The most compelling conspiracy theories produce clean narratives where every piece fits. Our brains can mistake narrative consistency for factual truth, registering that coherence as a sign of truth - even though tidiness in an explanation for a complex event is more often a sign that something has been edited out to make the pieces fit.

• Source credibility confusion. Conspiracy thinking often elevates individual voices (the independent researcher, the whistleblower) over peer-reviewed, institutionally vetted sources, precisely because institutional sources have been wrong before and have lost public trust for documented reasons. Questioning institutions is a reasonable response to institutional failure. Replacing the entire evidence hierarchy with no hierarchy at all is where the response stops being productive.

These tendencies are universal, but they don’t activate with equal intensity in every situation. When someone is feeling secure and well-informed, the pull of a clean conspiracy narrative is easier to examine critically. When someone is anxious, uncertain, or feeling shut out of the institutions that are supposed to provide reliable information, these same tendencies can become significantly harder to counteract, and conspiracy frameworks can become significantly more appealing.

Why some of us are more susceptible than others

Information literacy provides consistent principles for evaluating information, but research shows that our ability to apply those principles can vary based on our emotional state and the conditions we’re operating in. What we’re feeling, what we’re experiencing, and what needs the information is serving all affect how rigorously we evaluate what we encounter. Research suggests that certain psychological and situational factors can make some people more receptive to conspiracy explanations than others. A major meta-analysis published in Psychological Bulletin, drawing on 170 studies, found that conspiracy belief tends to be driven by several interconnected motivations, including the need to understand and feel in control of our environment (especially when events feel random or threatening) and the need to maintain a positive sense of ourselves and the groups we belong to.

In practical terms, this means that people experiencing higher levels of anxiety, uncertainty, or a sense of powerlessness may be more drawn to conspiracy frameworks, not because they lack intelligence, but because conspiracy theories offer something that uncertainty doesn’t: a clear explanation for why things are the way they are and who is responsible. Research in cognitive psychology has found that people with a stronger tendency to detect patterns in random data (like finding order in a random sequence of coin flips) are also more likely to endorse conspiracy beliefs. The drive to find structure is deeply embedded in human cognition, and conspiracy theories provide structure in situations where the evidence-based explanation is messy, incomplete, or simply unavailable.

Feelings of disenfranchisement play a role too. Surveys have consistently found that people who feel their side in politics is losing, or that they’ve been left behind by economic or social changes, are more open to explanations that name a specific group or force as responsible for their circumstances. A conspiracy theory can provide comfort by naming a specific cause and making the world feel more controllable: if these people weren’t doing this, things would be better. That framing is more psychologically manageable than the alternative, which is that complex systems produce outcomes that no single group controls and that bad things sometimes happen without anyone orchestrating them.

None of this means that susceptibility is a fixed trait or that believing a conspiracy theory is a sign of a psychological deficit. The research consistently emphasizes that these are normal human tendencies that become more pronounced under specific conditions: a chaotic information environment, eroded institutional trust, economic anxiety, and social isolation.

How the information environment compounds all of this

The cognitive tendencies described above (confirmation bias, pattern completion, narrative coherence, source credibility confusion) have always been part of human thinking. What’s changed is the information environment in which they operate. The skills required to evaluate information effectively have always been specialized, but the environment has shifted to make those skills more necessary than ever while simultaneously making them harder to practice.

Conspiracy content circulates more visibly now than at any point in recent history, even though research suggests that overall levels of conspiracy belief have remained remarkably stable over time. A 2022 study published in PLOS ONE examined belief in 46 conspiracy theories across data spanning decades and found no systematic increase in conspiracism. What’s changed is the volume and velocity of conspiracy content, not the underlying susceptibility, and that distinction points directly to the environment itself.

The current information landscape has several features that interact to make conspiracy content more effective than it would be in a healthier ecosystem:

Institutional trust has eroded, often for documented reasons.

Public trust in government, media, and corporations has declined significantly over the past several decades, as major polling organizations have consistently tracked. And institutions have often earned that erosion through their own documented failures: suppressed research, misleading public statements, regulatory capture, and a long track record of prioritizing institutional interests over public transparency. When the institutions that are supposed to serve as reliable sources of information have a credible track record of deception, the entire foundation of “trust the experts” weakens, and conspiracy frameworks step into the vacuum with alternative explanations that feel more credible precisely because they don’t ask us to trust the same institutions that have let us down before.

Algorithms amplify the most compelling version of any claim, not the most accurate.

Social media platforms and content recommendation systems are engineered to maximize engagement. Content that provokes strong emotional responses (outrage, fear, the thrill of hidden knowledge, moral certainty) gets amplified over content that is measured, nuanced, and carefully sourced. This means the most emotionally potent version of a conspiracy theory reaches more people than the boring, careful debunking of that same theory. The information environment actively promotes conspiracy content over the kind of evidence-based analysis that would serve as a counterweight.

Primary sources are inaccessible to most of us.

Government databases, court filings, academic research papers, and regulatory records are the raw material of evidence-based analysis, and they’re largely locked behind paywalls or formatted in specialized language that requires professional training to interpret. When primary sources are this inaccessible, secondhand interpretation becomes the version that circulates. And when secondhand interpretation is all we have access to, we’re dependent on the interpreter’s framing, biases, and editorial choices, whether that interpreter is a journalist, a content creator, or the architect of a conspiracy framework.

Traditional gatekeepers have lost persuasive authority.

Newsrooms, universities, and scientific publishers have been credibly shown to have their own biases, financial pressures, and editorial blind spots. This doesn’t mean their output is unreliable, but it does mean that the argument “experts say so” carries less automatic weight than it once did. When gatekeeping institutions can’t reliably be trusted to be neutral, the space for alternative narratives widens, and conspiracy frameworks can fill exactly that kind of space.

The cumulative effect of these forces is an information environment where conspiracy thinking can become a predictable response to institutional failure and information chaos. When institutional sources have a documented track record of deception, when algorithms surface the most emotionally charged version of every story, when primary evidence is locked away behind paywalls and jargon, and when the gatekeepers themselves have been caught with their own biases on display, the reasonable question becomes: what can we trust?

Conspiracy theories can spread because they step into that vacuum and function as a compass through the disarray. In a landscape where every official source feels suspect and the information environment feels chaotic and unnavigable, conspiracy content does something that institutional sources often fail to do: it acknowledges the distrust. It says, in effect, “you’re right to feel like you’re being misled.”

That acknowledgment, that validation of a feeling many people carry based on lived experience, is extraordinarily powerful. People share conspiracy content not because they’ve abandoned critical thinking but because, in an environment where nothing feels reliable, the conspiracy framework at least acknowledges that the official story has holes. The framework’s alternative explanation may also be wrong, but it earns trust by starting from a place that feels true: something is broken, and the people in charge aren’t telling us the whole story. A conspiracy theory can offer something that hedged, cautious institutional communication rarely does: clarity, even if that clarity is false.

Evaluating the claims

Information literacy as a discipline has been mapping these dynamics for decades. Evaluating conspiracy content draws on the same information literacy skills we’d use to evaluate any claim, applied under conditions that make those skills unusually hard to deploy.

The first question to ask about any claim within a conspiracy framework: does this reference something specific that can be independently checked? If the claim cites a particular event, document, study, or data point, it’s testable. If it describes a diffuse, unfalsifiable pattern (”a secret group is controlling everything”), it falls outside the reach of evidence-based evaluation.

For testable claims, the same tools we’d use for any factual assertion apply, starting with the most efficient step and moving deeper as needed:

1. Check whether someone has already done the work. Fact-checking organizations employ researchers who trace claims back to their original sources. A few starting points by region:

   • International: AP Fact Check, Reuters Fact Check, and AFP Fact Check operate across multiple countries and languages.

   • North America: Snopes, PolitiFact, FactCheck.org

   • Europe: Full Fact (UK), Maldita (Spain)

   • Africa: Africa Check (South Africa, Kenya, Nigeria, Senegal)

   • South Asia: BOOM (India)

   • Southeast Asia: Rappler (Philippines)

   • Latin America: Chequeado (Argentina)

   • Global directory: The International Fact-Checking Network maintains a searchable directory of verified fact-checking organizations worldwide, which is the fastest way to find a credible fact-checker in any country.

2. Find and read the original source. When a theory cites a study, hearing, or document, locate the original rather than relying on the theory’s characterization. These databases can help:

   • Academic research: Google Scholar indexes peer-reviewed research globally.

   • North America: Congress.gov (US congressional records), PACER (US federal court documents), FOIA.gov (US Freedom of Information Act releases), Open Government Portal (Canada)

   • Europe: Parliament.uk and GOV.UK (UK), EUR-Lex (EU legislation and court rulings)

   • Asia-Pacific: Right to Information Act portal (India), National Archives of Australia

   • Global directory: SearchSystems.net maintains an international directory of public records databases across 70+ countries. Over 130 countries have some form of freedom of information law.

3. Evaluate the source and compare it to the claim. When a conspiracy theory cites a study, a few questions can reveal how much weight that study should carry:

   • Who funded the research? A study funded by an industry group with a financial stake in the outcome has a built-in conflict of interest that independent research does not.

   • How large was the sample? Studies with small samples are more likely to produce results that don’t hold up when tested on larger populations.

   • Is the journal peer-reviewed? Peer-reviewed journals send submissions to independent experts in the field who evaluate whether the methodology is sound before the study gets published. Research published without that review process hasn’t been vetted by anyone outside the authors’ own team.

   • Have independent researchers replicated the findings? A single study is a starting point, but a finding confirmed across multiple independent research groups carries significantly more weight.

   • Does the theory accurately represent what the source says? A documented hearing or study could be cited, but the theory’s interpretation may not match the transcript or the findings. Reading the source directly often reveals a gap between what was said and what the theory claims was said.

4. Check the broader coverage. A claim that’s being independently reported by multiple unrelated outlets carries more weight than one circulating only within a single ecosystem. Ground News aggregates local and international coverage of the same story from multiple perspectives, making that distinction easier to see. For evaluating the outlets themselves, Media Bias/Fact Check rates over 10,000 sources globally on political bias and factual reliability, and NewsGuard provides credibility ratings for over 35,000 sources across North America, Europe, and Australasia.

For claims that can’t be tested, no amount of research will produce a definitive answer. A claim like “a secret cabal controls the global economy” can’t be verified or disproven by any available evidence. When we encounter claims like these, we can be clear with ourselves that accepting them is a choice based on belief rather than evidence, and we can evaluate the rest of the conspiracy theory’s claims independently rather than letting the unverifiable parts carry the whole narrative.

Seeing the architecture

The information environment we’re navigating produces more questions than credible answers. Institutional trust continues to erode and primary sources remain behind paywalls and specialized language, while algorithmic systems keep surfacing the most emotionally charged version of every story. As long as these conditions persist, conspiracy frameworks will continue to step in with explanations that feel coherent, even when the evidence behind them is incomplete or nonexistent.

Conspiracy theories gain their persuasive power from being evaluated as a whole. When we accept or reject an entire theory based on how the complete narrative makes us feel, we’re responding to how the theory is constructed rather than the evidence behind it. The shift that information literacy offers is the ability to evaluate individual claims within a conspiracy theory independently, which often reveals that a theory’s persuasive power comes from its narrative architecture rather than the strength of what it’s presenting.

Evaluating claims individually rather than accepting or rejecting entire conspiracy theories at once is a skill that sharpens with use. Each time we practice separating what's documented from what's been built around it, the structure of conspiracy content becomes easier to read. What once felt like a seamless, convincing narrative starts to reveal where the evidence stops and the speculation begins. From there, we get to decide for ourselves what to do with what we find inside it.

You might also like:

• What “Do Your Own Research” Actually Means: A step-by-step methodology for evaluating any claim and exploring any subject.

• The Hierarchy of Sources: A Cheat Sheet: A guide to evaluating information sources in the AI age.

• Steel-Manning Can Be More Useful Than Fact-Checking: How engaging with the strongest version of an opposing argument builds a deeper defense than verification alone. A practical guide with a downloadable worksheet.

Special announcement!

I’ve been building something I’m excited to share. On June 23, the Card Catalog Classroom opens!

The Classroom modules are the heart of the project. Every teachable article you’ve read here will become a structured lesson you can move through, with prompts and exercises a standalone piece can’t carry. This will extend both the actionability and the usefulness of what we discuss in the articles. Future recorded workshops will live there too.

Alongside the lessons is the Reference Collection: a browsable directory of every archive, database, history, and tip I’ve ever pointed to in a Note. Instead of scrolling through my feed to find that one thing I mentioned, you’ll be able to search the entire collection in a more organized and accessible way. Every new Note on those topics folds in automatically, so the collection keeps growing as we go. I firmly believe that good information is findable information, and this is how I’m putting that into practice.

One thing to mention: Substack does not yet natively host things like courses or directories, so these paid subscriber benefits will be hosted on Circle. Substack stays exactly where it is for the free articles and Notes, but Circle gives me the structure to organize everything properly. And when the time comes to open up an information village, Circle is where that will live too.

Paid subscription pricing is also shifting when the Classroom launches. Right now it’s $8 a month/$80 a year, and those rates will move to $10 monthly/$100 annual. Anyone who becomes a paid subscriber before June 23 will be grandfathered in at the current rate of $8 monthly/$80 annual, permanently. Even if there’s another increase down the line, you’ll always be locked in at what you signed up at. It’s how I want to say thank you to readers who’ve been here since the beginning!

Nothing in the current paid tier is going away. The existing benefits (AI Briefings, Librarian Hotline, Research Packets, Founding Member Report) will continue to be published on Substack as they always have, and they’ll also be featured and findable in the Circle library along with everything else. And all the new services for paid subscribers (the Classroom modules and the Reference Collection) are going to live in Circle, organized in a way that’s far easier to use, learn from, and reference. A dedicated email is coming soon with the full breakdown of what to expect - more coming soon. ☺️

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

This week:

• India partnered with the UAE to put AI supercomputers on its own territory, bypassing Amazon, Microsoft, and Google to build infrastructure it owns and controls.

• The CEOs of OpenAI, Anthropic, Google DeepMind, Microsoft, and Meta set aside their rivalry to co-sign a public letter warning Congress that their own models could help someone build a biological weapon.

• Google launched a feature that tells us when a phone call may be AI impersonating a voice we trust.

India and the UAE created a blueprint for countries that want to own their AI, not rent it from American companies

What Happened

For most governments that want AI capability, the path has been to rent computing power from Amazon, Microsoft, or Google. These companies operate the data centers, own the hardware, and govern the infrastructure that most of the world’s AI currently runs on. India already had at least $45 billion in commitments to all three. This month, it signed a different kind of deal. G42, an AI company based in the UAE and backed by Abu Dhabi’s government investment fund, agreed to deploy 64 supercomputers made by US chipmaker Cerebras on Indian soil. (Cerebras builds chips designed for speed in running AI applications, which is different from the Nvidia processors India uses for training models.) The choice reflects India’s focus on deploying AI across healthcare and public services rather than building large foundation models from scratch.

The machines will sit on Indian territory under Indian governance, with G42 handling operations. India is the first country to join what G42 calls the Intelligence Grid, a global network of AI computing facilities it builds and operates for governments. G42 is already in discussions with other governments about the same model. Cameron Kerry, former acting secretary at the US Department of Commerce, reflects that India’s “pragmatic approach to AI sovereignty, using the power of its scale to adapt what’s available from other countries to its own needs.” G42 brings hands-on experience to the arrangement: it built and now operates supercomputer facilities in the US and became Cerebras’s largest customer before expanding into this government-focused model.

Why It Matters

Until this deal, a country that wanted serious AI computing capability had to route that capability through American infrastructure, on American companies’ terms, and subject to their supply chain and regulatory decisions. India’s arrangement with G42 introduces a different option: machines physically on Indian soil, data governed by Indian rules, managed by a partner outside the American technology ecosystem. This is what AI sovereignty looks like in practice. The Intelligence Grid model is designed to replicate this arrangement across other countries, and the fact that G42 is already in discussions with other governments means India’s deal may be the first of many.

What It Means for Us

The AI tools available to us today are products of a specific moment in which a small number of American companies control most of the world’s AI infrastructure. That control shapes what gets built, for whom, and on what terms. If the Intelligence Grid model spreads, the next generation of AI development happens in a more distributed world, with more governments owning the infrastructure that serves their populations and making decisions about how it operates. The tools that emerge from that world will reflect a wider range of priorities than the ones currently shaping what we have access to now.

Bottom Line: India’s deal turns AI sovereignty into a budget question. Building this kind of infrastructure used to require years of national investment in domestic chip production and data center capacity. G42 has packaged it into a commercial service. India just secured what previously required a national-scale construction effort, which lowers the practical barrier to AI independence significantly.

OpenAI, Anthropic, Google, Microsoft, and Meta jointly warned Congress that their own chatbots could help someone make a biological weapon

Read more

I’m glad you’re here for this one. There’s an announcement at the end about what’s coming next for Card Catalog.

In 1967, the New England Journal of Medicine published a review of the scientific literature on diet and heart disease. Three Harvard scientists examined the evidence and reached a conclusion that would echo through American kitchens for the next fifty years: dietary fat, not sugar, was the primary driver behind rising rates of coronary disease. National dietary guidelines shifted accordingly. Food manufacturers raced to reformulate their products as low-fat, loading them with sugar to compensate for the lost flavor. Grocery store shelves filled with fat-free cookies and reduced-fat peanut butter, products stamped with the word “healthy” that were anything but.

Nearly fifty years later, Cristin Kearns, a researcher at the University of California, San Francisco, uncovered internal documents from the Sugar Research Foundation showing that the organization had paid the Harvard researchers $6,500 (the equivalent of around $65,000 in 2026) to produce the review. The foundation had selected which studies to include and steered the conclusions toward exonerating sugar. The funding was never disclosed, and at the time the journal didn't require its authors to reveal who was paying them.

For half a century, the dietary choices of hundreds of millions of people were built on conclusions the sugar industry had purchased. The industry knew its product was implicated in heart disease. The rest of the world was deciding what to feed their families based on a review designed to make sure they’d never find that out.

Economists call this information asymmetry: a situation where one party in an exchange holds far more relevant knowledge than the other, and that gap gives the more-informed party the power to shape the outcome. The concept operates at every scale, from a corporation redirecting an entire nation's dietary policy to a contractor charging for repairs a homeowner can't evaluate. The term entered modern economics in 1970, when George Akerlof published his analysis of the used car market and demonstrated that when sellers know whether a vehicle is reliable but buyers can only guess, the market itself degrades. Buyers assume the worst and offer less; sellers of good cars refuse those prices and walk away. Akerlof's insight, which won him a Nobel Prize in 2001, was that the knowledge gap between parties doesn't just create unfairness in individual transactions; it corrodes the integrity of entire systems.

The sugar industry’s campaign illustrates something larger than one company’s deception. It shows information asymmetry functioning as an ongoing investment in controlling what the public could know, sustained across decades because the gap itself is a profit engine. A gap that valuable never closes from the inside; ending it took an outsider willing to spend years of research the industry assumed no one ever would undertake.

Underneath that campaign, and underneath every case like it, sits a single question: who benefits from my not knowing? That question changes what we look at. Instead of asking whether a claim sounds “right”, we ask who’s telling us and what they get if we believe them. Asked of the 1967 review, the question turns straight to who paid for the research, which is exactly what the sugar industry had kept hidden. But a question like that only occurs to us once something feels off, and the best-built gaps make sure nothing ever does. Their targets get no flicker of doubt to work with. The hardest asymmetries are the ones engineered so well that we never sense them at all.

This is not a new problem. The same pattern runs through every society that ever kept records, long before anyone called it information asymmetry.

A pattern as old as civilization

The pattern shows up in some of the earliest writing humans produced. In ancient Sumer, a professional class of scribes mastered cuneiform, the wedge-shaped writing system pressed into clay tablets to record grain stores, property boundaries, debts, and legal agreements. Most Sumerians had no practical need to read these records, though. The majority were farmers and herders, so the scribes served a functional role: translating administrative information for people who needed the results but not the process (much the way an accountant processes tax records today).

The scribes weren’t hoarding secrets, and the farmers had no reason to learn cuneiform - so asymmetry grew out of specialization rather than any intent to exploit. And yet over generations, the scribal class accumulated influence because they controlled the interpretation of records that governed everyone’s economic life. A system that began as a practical division of labor hardened into structural authority, and the people who could read the records came to shape what those records said. The Sumerian case demonstrates that information asymmetry doesn’t require a villain; the pattern can emerge from perfectly reasonable arrangements and still concentrate power on one side.

In later examples, the link between knowledge and power is harder to miss. Medieval European monasteries preserved much of the classical learning that survived in the Latin West, and monks who could read Latin occupied positions of authority in a largely illiterate society. Colonial administrators in British India maintained legal codes in English that ruled local populations speaking dozens of other languages, a linguistic barrier that forced reliance on those who could speak the governing language. Across much of the antebellum South, laws made it a crime to teach enslaved people to read, because slaveholders understood what literacy would do to the existing power structure. These cases range from specialization hardening into privilege to deliberate suppression designed to preserve control, but the common thread persists: wherever knowledge concentrates on one side, the other side loses ground.

Telling a helpful gap from a harmful one

Strung together, those cases make information asymmetry look like nothing but a threat, which misses half of what these gaps do. Many knowledge gaps serve us rather than disempower us, and complex societies couldn't run without them. We rely on exactly this kind of gap every time we trust an expert to know what we don't. For example, a dentist spends years learning to spot hairline tooth fractures we’d never see in a mirror, or a pilot reads cockpit systems we’d need months of training to follow. We’re glad to stay on the uninformed side of both, since closing either gap would cost years of study we have no reason to take on.

What separates a functional asymmetry from an exploitative one is whether the person holding more knowledge is using that advantage to serve the relationship or to exploit the gap. For example, two doctors can see the same symptom. One explains the treatment options and walks the patient through the alternatives and risks. The other orders unnecessary tests that funnel revenue to an imaging lab they co-own. The knowledge gap between doctor and patient is identical in both scenarios, but what changes the patient’s experience entirely is whether the doctor treats that gap as a responsibility or as an opportunity for extraction.

This cuts both ways, because we’re rarely only on the receiving end of a knowledge gap. Every one of us holds an information advantage somewhere: as the parent who understands the family finances, the manager who knows why a decision was really made, the friend with a medical background, or the colleague who’s been at the company longest. When we’re the one who knows more, the question turns inward: are we sharing what we know so the other person can decide for themselves, or holding it back to keep them dependent on us?

On the other side of that gap, where we’re the one who knows less, the hard part is judging how far to trust the one who knows more. Refusing to trust any expert would cut us off from knowledge we need, and taking everything on faith leaves us exposed the moment someone exploits the gap. What protects us is the ability to tell an expert who informs us from one who exploits us, and reading a situation closely enough to know which we’re facing.

Who profits when the gap stays open

The exploitative kind gives itself away in one respect: the party benefiting from the gap spends money and effort to keep it open. When a company lobbies against ingredient disclosure laws, or a financial institution buries fee structures in legally dense documents, the asymmetry isn’t a byproduct of complexity. These are calculated investments in keeping the public less informed, because an informed public would make different choices.

Food packaging is where many of us run into this first. In December 2024, the FDA finalized a rule tightening the criteria food manufacturers must meet to call a product “healthy,” the first revision to that standard since the 1990s. The original criteria set no limit on added sugars, so a sugary cereal could qualify for the label while foods like nuts and salmon could not. A loose standard served the companies using it, because a shopper reading “healthy” on a package had no easy way to know how little the word guaranteed. The same opacity shapes pharmaceutical pricing, where rebates and negotiated discounts are buried across a complex supply chain, leaving patients and even physicians unable to determine what a medication truly costs or why one option is priced far above another.

The pattern extends into political life as well, where the connection between campaign funding and policy positions is the kind of information that would change how voters evaluate their representatives. Dark money channels exist specifically to prevent voters from drawing that connection. According to the Brennan Center for Justice, spending by groups that don’t disclose their donors reached a record $1.9 billion in the 2024 federal election cycle. The legal structures that allow this spending were designed to keep the distance between what politicians know about their own funding and what voters can discover as wide as possible.

Financial services follow a similar logic, with banks and credit card companies designing fee structures that are technically disclosed in fine print but practically illegible. And social media platforms know far more than we do about how their recommendation algorithms shape what we see and how our behavioral data feeds the engagement machine, while we encounter only a feed. The privacy policies and terms of service governing these exchanges are written to go unread; a Carnegie Mellon study estimated that reading the privacy policy of every website a person visits in a year would consume roughly 244 hours.

And a more concentrated version of this gap is taking shape now: in systems built on artificial intelligence. The companies training large language models know what data the models learned from and how the systems are tuned to respond, while the rest of us see only a confident answer in a chat window. There’s no source attached and no easy way to tell whether a given response reflects the weight of evidence or the preferences of whoever built the tool. As these systems move into research, healthcare, hiring, and lending, the distance between what their makers promote and what the rest of us can see keeps widening.

The sectors differ, but they share one aim: making the answer so costly to dig up that the question of who profits from our not knowing rarely gets asked.

Four questions for any information exchange

Asking that question anyway, and the ones that follow from it, is the everyday work of information literacy. Information never reaches us in a neutral state: every study, news report, product claim, and policy brief is shaped by someone’s choices about what to include and what to leave out, and the skill is reading those choices, not only the content. Librarians have spent a long time turning that skill into questions anyone can use. Four of them do most of the work, each coming at a claim from a different angle to surface who shaped it and how their motives influence what we end up seeing.

Who benefits from this, and what do they gain?

Every piece of information reaches us through someone’s effort, and that someone has reasons for producing the content. A clinical trial funded by the manufacturer of the drug being tested carries different weight than one conducted by independent researchers. A policy paper from a think tank funded by the fossil fuel industry operates differently than the same research produced by an academic institution with no industry ties. Funding doesn’t automatically invalidate findings, but it shapes what gets asked and how the results get framed. The same logic carries into personal situations, where the question becomes who stands to gain if we accept what we’re being told without pushing back.

How to find out: Most studies and reports include funding disclosures at the bottom or in an acknowledgments section. A web search for an organization’s name plus the word “funding” or “donors” will often surface what their website doesn’t. For personal situations, ask ourselves who stands to benefit if we take the recommended action.

What’s being emphasized, and what’s been left out?

Spotting an omission is harder than catching a false claim, because we have to notice something that isn’t there. The move is to ask what a transparent version would have to include, then see what’s missing. A credit card advertises “0% APR for one year” in large type while the rate it later jumps to sits in the fine print, and a cereal stamped “good source of fiber” says nothing about the sugar that dwarfs the fiber. What a message leaves out often tells us more about its intent than what it puts forward.

How to find out: Compare how multiple sources cover the same topic. If several sources mention a detail that the one we’re reading omits, that absence tells us something about the priorities of the omitting source. For products, compare what the marketing says with what independent reviewers report.

What would someone with a different stake in this say?

The move here is what fact-checkers call lateral reading: rather than reading a single source more closely, we leave it to see what independent sources with different motives say about the same claim. If a supplement company says its product boosts immune function, we can search for what the National Institutes of Health says about the same ingredient. If a landlord calls an apartment's rent the market rate, we can look up what comparable units in the same zip code go for.

How to find out: Start a fresh search on the source itself, the company or author behind the claim, and read what independent outlets and reference databases say about them. In personal situations, ask a trusted friend or advisor who has no stake in the outcome.

Have we gone to the original, or stopped at the summary?

Most claims reach us at second or third hand (for example, a headline summarizing a study or a campaign ad describing a record), and each retelling is a chance to bend what the original said. Going to the source lets us see whether the summary was fair, and lets us turn these first three questions on the original itself. Keep in mind, this doesn’t make a primary source trustworthy on its own; the 1967 review was a primary source, and the results had been bought. But the value here is the chance to weigh the evidence ourselves instead of accepting a secondhand version.

How to find out: Follow a claim back to its primary record - the actual study or filing - rather than the version someone summarized for us. Scientific papers are searchable worldwide through PubMed, and many governments keep public registers of court records, company filings, lawmakers’ votes, and public spending, though the databases differ from one country to the next. When the trail is hard to follow, a reference librarian can point us to the right one.

None of this is a checklist to run in order. Asked at the right moment, any one of these questions can tip the power balance of a lopsided exchange back toward ourselves. The hard part, though, isn’t necessarily knowing which question to ask; it’s asking that question out loud, and following through to find the answer. The discomfort we feel doing that is sometimes what the more-informed party is counting on.

When we don’t know we’re in the dark

All of this assumes we can feel the gap in the first place. The hardest asymmetries give us nothing to work against, arriving as consensus and the ordinary background of what everyone “already knows”. The work then shifts from reacting to a bad exchange toward questioning the things we never thought to question.

Since the gap won’t announce itself, we turn the question on our own certainties. What am I taking as fact, and where did it come from? Comfortable, unchecked beliefs that feel like tradition often started as someone’s sales pitch. One of these is the rule that an engagement ring should cost two months’ salary, which a De Beers ad campaign invented to sell more diamonds. When a belief leads back to a marketing line instead of evidence, we’ve found the gap that was hiding inside the consensus.

The De Beers story is easy to trace because its history is on the record. Most concealment leaves no such trail, and we catch it instead as a mismatch between sources that should agree. When everything we read shares the same interests, nothing ever contradicts anything else and a hidden gap stays hidden. Reading sources whose interests clash, such as the trade journal alongside the consumer group report, is what surfaces the disagreement that finally makes us stop and ask.

But even these methods have a limit. Some gaps are buried too deep for any one of us to reach, no matter how carefully we look. Uncovering the low-fat story took researchers years in the archives, well beyond what one reader could manage. That is what investigative reporters and regulators are for: they can spend months on an investigation and force a company to disclose what it hides. So we don’t have to carry this alone. Our part is to support that work and to spread what it finds, so a gap one person uncovers becomes something all of us know.

The pattern in plain sight

Information asymmetry will always be a part of life. Specialization guarantees that someone will always know more than the person across the table; that’s a feature of complex societies rather than a flaw. But what changes is how we hold that fact. We can stop treating knowledge gaps as fixed conditions we’re stuck with, and start treating them as distances we can question and close.

One question travels everywhere information reaches us shaped by someone’s motive: who benefits from my not knowing? When no gap is visible at all, we turn the same question on our own settled beliefs: what are we treating as fact, and who benefits if we never check where it came from? We don’t have to answer either one perfectly. We only have to remember to ask, and then do the work the question points us toward - whether that means looking up who funded a study or comparing a claim against a source with nothing to gain.

Every question we ask and every source we verify narrows the distance between what we know and what the other party assumed we’d never find out. What one of us uncovers and shares narrows that distance for everyone else. The sugar industry counted on fifty years of no one looking. The companies and platforms that profit from information asymmetry today are counting on the same thing. The tools for checking have never been more widely available, and they’re strongest when we use them together.

You might also like:

• The Rise of the Grifter Guru: Inside the self-help industry’s plagiarism problem, and what it teaches us about evaluating any source.

• What to Do When Information Is Skewed, Censored, or Silenced: Skills that outlast the systems trying to suppress them.

• The information diet isn’t enough. What we need is an information village.: What changes when we go beyond optimizing our consumption and start building shared infrastructure.

Special announcement!

I've been building something I'm excited to share. On June 23, the Card Catalog Classroom opens!

The Classroom modules are the heart of the project. Every teachable article you’ve read here will become a structured lesson you can move through, with prompts and exercises a standalone piece can’t carry. This will extend both the actionability and the usefulness of what we discuss in the articles. Future recorded workshops will live there too.

Alongside the lessons is the Reference Collection: a browsable directory of every archive, database, history, and tip I’ve ever pointed to in a Note. Instead of scrolling through my feed to find that one thing I mentioned, you’ll be able to search the entire collection in a more organized and accessible way. Every new Note on those topics folds in automatically, so the collection keeps growing as we go. I firmly believe that good information is findable information, and this is how I’m putting that into practice.

One thing to mention: Substack does not yet natively host things like courses or directories, so these paid subscriber benefits will be hosted on Circle. Substack stays exactly where it is for the free articles and Notes, but Circle gives me the structure to organize everything properly. And when the time comes to open up an information village, Circle is where that will live too.

Paid subscription pricing is also shifting when the Classroom launches. Right now it’s $8 a month/$80 a year, and those rates will move to $10 monthly/$100 annual. Anyone who becomes a paid subscriber before June 23 will be grandfathered in at the current rate of $8 monthly/$80 annual, permanently. Even if there's another increase down the line, you'll always be locked in at what you signed up at. It’s how I want to say thank you to readers who’ve been here since the beginning!

Nothing in the current paid tier is going away. The existing benefits (AI Briefings, Librarian Hotline, Research Packets, Founding Member Report) will continue to be published on Substack as they always have, and they'll also be featured and findable in the Circle library along with everything else. And all the new services for paid subscribers (the Classroom modules and the Reference Collection) are going to live in Circle, organized in a way that’s far easier to use, learn from, and reference. A dedicated email is coming soon with the full breakdown of what to expect - more coming soon. ☺️

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

We’re encountering both events through the same channels: trending names, live blogs, a scrolling feed of updates that adds information faster than it adds clarity. We can recite death tolls and country names within minutes of opening a news app, but the question underneath all the others remains stubbornly out of reach. How worried should we be? And how would we know if that changes?

The institutions investigating these outbreaks publish their own findings in structured documents that address both of those questions. They explain the reasoning behind each assessment and describe what new evidence would cause the assessment to shift. These documents are public and free, updated as new information arrives, but they’re not widely known and few of us have been walked through how to find or read them.

Reader Question

“I keep seeing headlines about Ebola and hantavirus at the same time and I can’t tell how scared I should be about either one. The news gives me numbers and quotes but I don’t know how to make sense of any of it. A friend told me to just look at the WHO website, but when I got there I had no idea what I was reading or where to start. Where do I go to find information I can actually understand and trust?”

— Mara K.

A note before we begin: Thank you for being here, whether you’re a free or paid subscriber. This post covers where to find primary source health documents and what makes them different from news coverage (available to everyone). The full walkthrough of how to read a WHO Disease Outbreak News page, how to decode a CDC Health Alert Network notice, a guide to national and regional health agencies around the world, and a complete reference list of sources for any future health emergency is available for paid subscribers.

Librarian Answer

The World Health Organization (WHO), national agencies like the US Centers for Disease Control and Prevention (CDC) and the Public Health Agency of Canada (PHAC), continental bodies like Africa CDC and the European Centre for Disease Prevention and Control (ECDC), and their counterparts around the world publish their investigative findings as they go. These documents are freely accessible online, and they contain the answers to the questions news coverage often compresses away: how certain is the assessment, what evidence supports it, what’s still unknown, and what would change the picture. Once we know where to find them and what to look for on the page, we can use them for any health event.

What Makes These Documents Different from News Coverage

News coverage of a disease outbreak draws on a smaller body of original evidence. Somewhere upstream, an institution is collecting samples, sequencing the virus, interviewing patients, and tracing contacts. That institution publishes its findings in a structured document, and in library science, that document is called a primary source: the original material that every subsequent account is built from. For disease outbreaks, the most comprehensive primary source regularly available to the public is generally published by WHO in a format called the Disease Outbreak News, or DON.

A news headline might tell us “WHO says global risk is low.” A DON tells us why WHO reached that conclusion, what factors informed the assessment, how much of the evidence is laboratory-confirmed, and what new data would cause the conclusion to change. News coverage compresses that structural depth out because of how it operates: the need for speed and the pressure to distill complex assessments into short sentences. The primary source preserves that depth, and that’s what makes the difference between receiving a conclusion and being equipped to evaluate one.

For both the Ebola and hantavirus outbreaks, these primary source documents are available right now. WHO has published Disease Outbreak News updates for both events, and the CDC has issued Health Alert Network notices for each. For the Ebola outbreak specifically, Africa CDC has declared a Public Health Emergency of Continental Security and is publishing its own situation reports and technical guidance. The pages are public and free, and the structural features that set them apart from news coverage become clear the moment we open one.

How to Read a WHO Disease Outbreak News Page

Read more

Research has confirmed that concern on one specific point: when we write things down, we do remember them less precisely than when we hold them in memory alone. The oral cultures Socrates belonged to had developed extraordinary memory practices, and writing did erode those practices over time. But Thamus’s prediction that writing would produce shallow thinkers turned out to be wrong about the trajectory. Writing made possible an entirely new kind of thinking that oral memory couldn’t achieve: it accumulated knowledge across generations and gave individuals the ability to reason at a scale no single mind could sustain.

Major information technologies since have tended to follow a similar arc. The printing press was blamed for spreading dangerous ideas and destabilizing institutional authority, and it did both of those things while also making widespread literacy possible. Calculators were going to destroy mathematical competence, and mental arithmetic skills did decline among people who relied on them, but calculators also freed people to focus on higher-order mathematical reasoning at scales no human could compute by hand. Nicholas Carr’s “Is Google Making Us Stupid?” argued in The Atlantic in 2008 that the internet was rewiring our brains for distraction and shallow thinking, and attention research has found evidence consistent with that claim, even as the internet made it possible to synthesize information across vast numbers of sources and collaborate in ways that were previously unimaginable. The alarm was grounded in a true cognitive cost every time, and the full picture consistently turned out to include capabilities that offset or exceeded that cost.

We’re in the early chapters of that arc with AI right now.

The “AI is making us dumber” narrative

Since the widespread adoption of AI tools beginning in late 2022, a specific media narrative about AI and cognition has hardened into something close to conventional wisdom. “Is AI Making Us Stupider?” ran in Psychology Today. “AI May Be Making Us All Dumber” ran in Inc. Teachers reported declining comprehension among students who overuse AI. Some UK schools began experimenting with “no AI” zones designed to encourage independent thinking.

When a study from MIT’s Media Lab appeared to confirm these concerns with EEG brain scans, the MIT brand carried the story into CNN, The New Yorker, and countless LinkedIn posts. By the time the findings reached mainstream media, the nuance the researchers had built into their work had been stripped out almost entirely. On their own website, the MIT researchers asked journalists not to describe their findings using words like “stupid,” “dumb,” “brain rot,” “harm,” or “damage,” because those words misrepresented what their data showed. Their data showed something the coverage didn't reflect: that the cognitive effects they measured were tied to a particular way of engaging with AI, and that different modes of engagement produced different outcomes.

What the MIT study found (and what the coverage got wrong)

The MIT Media Lab study, led by Nataliya Kosmyna and published as a preprint (a paper shared publicly before undergoing peer review) in June 2025, tracked 54 participants across four sessions spread over roughly four months. Each participant was assigned to one of three conditions: writing essays with ChatGPT, writing with a search engine, or writing with no external tools. All participants wore EEG headsets, which measure electrical brain activity in real time, while they wrote SAT-style essays.

The ChatGPT group showed measurably weaker brain connectivity than either of the other two groups. Their alpha and beta brain waves, which are associated with working memory and concentration, showed significant decline. Their essays were more generic. When asked to recall their own work minutes later, 83% of the ChatGPT users couldn’t quote from the essays they’d just written. The researchers coined the term “cognitive debt” to describe what they observed: reduced neural engagement that accumulated over sessions and persisted even after participants stopped using AI.

These findings are consistent with decades of research on cognitive offloading, a well-documented pattern in which the brain reduces its own engagement when an external tool handles a task it would otherwise do. Research has shown this with GPS and spatial memory, and with typing versus longhand notes and retention. When we offload a cognitive task, the neural pathways responsible for that task weaken from disuse. The MIT study captured one version of that pattern clearly.

What the headlines didn’t report was the nuance the researchers found in their own data. In the fourth session, the team ran a crossover: participants who had been writing without any external tools were given access to ChatGPT for the first time, and ChatGPT users were asked to write without AI. The participants who had spent three sessions building their own writing muscles and then switched to AI maintained stronger neural engagement than the participants who had been relying on AI from the start. An independent cognitive baseline, built through those prior sessions of unaided effort, appeared to provide meaningful protection against the disengagement effect.

The researchers also observed, through interviews and behavioral analysis, that the steepest neural decline within the ChatGPT group was associated with the most passive use: accepting AI output without questioning or engaging with the material. They saw enough complexity in their findings that they cautioned against treating the results as a blanket indictment of AI. The cognitive disengagement the study found was specific to one mode of use - passive acceptance of AI-generated text - rather than a universal consequence of interacting with AI.

The coverage that reached the public reflected almost none of that complexity. The media cycle that carried the MIT study into mainstream awareness operates on simplification: outlets compete for attention, and attention rewards alarm over precision. A finding about specific EEG patterns under specific conditions doesn’t generate the same engagement as a sweeping claim about cognitive decline. By the time the study had been summarized, re-summarized, and shared across platforms, the qualifications the researchers considered essential had been stripped away almost entirely.

The research on collaborative AI use

Henry Shevlin, a British philosopher and AI ethicist at Cambridge’s Leverhulme Centre for the Future of Intelligence, offered a frame that helps clarify what the MIT study’s data showed versus what the coverage claimed. He compared passive AI use to handing over a daily commute to a self-driving car: convenient, but over time, we forget how to drive. The erosion happens because the skill goes unexercised, and the tool itself isn’t the source of the damage. AI can help us think and clarify our ideas, Shevlin argued, but if it starts thinking for us, the risk of cognitive disengagement becomes real. Dan Levy at Harvard Kennedy School, co-author of Teaching Effectively with ChatGPT, draws a similar line: some tasks build our capacity to handle what comes next, and some tasks are mechanical work that wouldn’t develop any skill regardless of who does them. Christopher Dede at Harvard’s Graduate School of Education captures the implication through a metaphor from Greek mythology: Athena, the Greek goddess of wisdom, is traditionally depicted with an owl on her shoulder. The owl extends her perception, but the wisdom is hers. AI should function like the owl: a tool that enhances what we can see and process, while the understanding and the judgment remain ours. When we let AI take over the thinking and reduce ourselves to the role of observer, we lose the ability to recognize when the tool gets it wrong.

All three researchers are describing the same core distinction, and a growing body of peer-reviewed research has been testing what happens on each side of that divide.

In June 2025, a randomized controlled trial published in Scientific Reports (a peer-reviewed journal in the Nature portfolio) tested an AI tutor built on evidence-based pedagogy against in-class active learning. Students using the AI tutor learned significantly more in less time, with a large effect by education research standards (0.73 to 1.3 standard deviations above the comparison group). These tutors work nothing like ChatGPT used as a text generator. They ask questions and prompt the learner to work through problems step by step, withholding answers until the student arrives at understanding through their own cognitive effort. The AI structures the conditions for deeper thinking rather than replacing the thinking entirely.

In 2023, a study from Harvard Business School, MIT Sloan, Wharton, and the University of Warwick tested 758 consultants at Boston Consulting Group across 18 realistic work tasks. The AI-using group completed 12% more tasks and finished 25% faster, with 40% higher quality ratings. The consultants who produced the highest-quality work were those who maintained active cognitive engagement throughout, dividing tasks between themselves and AI while retaining control of the analysis and judgment.

And in an October 2025 cross-country experiment with 150 participants spanning Germany and Switzerland as well as the UK, researchers tested the mechanism directly. Participants who used AI without guidance showed cognitive offloading and no improvement in reasoning, a finding consistent with the MIT study’s results. But the guided group, who were prompted to question AI outputs and build on the AI’s responses, showed significantly improved critical reasoning. They described the experience as being challenged in a seminar or arguing with a tutor. The tool was identical in both conditions, but the cognitive outcome depended entirely on whether participants engaged passively or actively with what the AI produced.

Passive AI use vs. Active AI use

The research points to a consistent finding: the cognitive effect of AI depends on the mode of engagement. There are two distinct modes with measurably different outcomes, and understanding the difference between them is what makes the research actionable. The distinction operates at every level of the interaction, from the initial mindset we bring to the long-term effect on our cognitive capacity.

How to Use AI to Think Better

The difference between passive and active AI use comes down to whether we’re delegating our thinking or engaging our thinking. Both involve getting useful output from AI, but the difference is what our brain is doing during the interaction: coasting or working. The prompts that follow are organized by situation and can be copied and pasted directly into any AI chat tool (ChatGPT, Claude, Gemini, or any other), with the bracketed sections replaced with our own specifics.

When we’re trying to understand something

Give me an overview of [topic], then tell me what the main debates or open questions are.

Open questions are where our own judgment and curiosity become useful. A flat summary gives us information; the debates give us somewhere to direct our own thinking, which means we’re learning actively rather than just absorbing.

What are the most common misconceptions about [topic], and why do people hold them?

Knowing what people get wrong about a subject, and why the wrong answer feels right, gives us better footing than a correct summary alone. It means we can spot weak reasoning when we encounter it elsewhere.

Here’s what I know so far about [topic]: [what we know]. What should I be reading or looking into next?

Starting with what we already know turns AI into a research guide rather than a research replacement. The AI can see where our knowledge has gaps and direct us to fill them, rather than starting from zero and anchoring our entire understanding on whatever the AI generates.

What are the most important questions about [topic] that a beginner wouldn’t think to ask?

When we’re new to a subject, we don’t yet know what the important questions even are, so we end up researching whatever comes up first rather than what matters most. This gives us a map of the territory from the perspective of people who know it well, so we can direct our own learning toward what’s going to be most useful.

When we’re writing or building something

I’m going to write about [topic]. Before I start, what questions should I be thinking through?

This gives us a thinking scaffold before we write, which is different from asking the AI to write for us. We’re building the structure ourselves; the AI is helping us see what the structure needs to address.

Identify exactly 5 logical gaps or weak structural transitions in this text. Do not include any praise.

Specificity and constraints fundamentally change what AI produces: a specific number forces it to look harder, and excluding praise eliminates the vague encouragement AI defaults to. What comes back is targeted critique we then have to address ourselves.

Read this as if you were [my manager / a skeptical investor / someone encountering this topic for the first time]. What questions would you have?

Seeing our own work through someone else’s eyes is one of the hardest parts of writing and one of the most valuable uses of AI. The questions it surfaces are often the ones a real reader would have.

Where does this lose the reader, and why?

This is more targeted than general feedback. It identifies the specific moment attention drops and gives us a reason, which means we can fix the structural problem rather than just polishing the surface.

When we’re making a decision

Here are my options: [list]. For each one, give me the best case and the worst case. Then tell me what I’m not considering.

Our options tend to reflect the frame we’re already in. The “what I’m not considering” clause asks AI to widen that frame by surfacing factors we haven’t weighted or alternatives we haven’t seen.

I’m leaning toward [option]. Give me three reasons I might regret this in a year.

Most of our decisions are shaped by how we feel right now. This forces a view from the future, where the excitement of the choice has faded and we’re living with the consequences. It doesn’t tell us what to decide, but gives us information our present-moment thinking naturally filters out.

If this plan fails, what’s the most likely reason?

Pre-mortem analysis means running through a failure scenario before the decision is made, and it’s a well-established technique in decision science. AI can run one instantly for any scenario we describe, and the failure modes it identifies are often ones we’d rather not think about - which is exactly why they’re valuable.

What would a person who chose [the option I’m not leaning toward] know that I don’t?

This reframes the rejected option as potentially containing information rather than being simply wrong. It’s a way of pressure-testing our preference by asking what someone with a different conclusion might be seeing that we’re missing.

When we want feedback

Steelman the opposing position to what I’ve argued here.

Steelmanning means constructing the strongest, most charitable version of the case against our own position (the opposite of a strawman). AI is unusually good at this because it has no ego investment in any position, which means we’re engaging with the best counterargument rather than the easiest one to dismiss.

What would someone with [specific expertise: a financial analyst / a user researcher] say about this approach?

Different experts see different risks in the same plan. A financial analyst will focus on cost assumptions; a user researcher on adoption barriers. We can run through these perspectives quickly, and each one gives us a lens we wouldn’t have on our own.

If I had to defend this to [specific skeptic or tough audience], what would they challenge first?

Anchoring the critique to a specific person or audience produces more targeted and realistic pushback than a generic “what’s wrong with this,” - the kind we’re likely to face in the real conversation rather than in the abstract.

Am I overcomplicating or oversimplifying anything here? Be specific about which and where.

We tend to add complexity where we're uncertain and gloss over areas where we're overconfident. This lets the AI flag whichever pattern is actually present rather than forcing it to find both.

When we’re evaluating something

Here’s a summary of [a job offer / a contract / a proposal]. What should I be paying attention to that I might miss?

(For anything containing sensitive personal or financial information, summarize the key terms or paste in specific non-sensitive sections rather than uploading the full document.) We tend to focus on the headline factors in any evaluation and underweight the details that end up mattering most. This is designed to surface the overlooked details.

Here’s [someone’s argument / a sales pitch / a news article]. What’s it assuming that it doesn’t say out loud?

Every argument rests on premises it doesn’t state explicitly. Surfacing those unstated premises is one of the most useful critical thinking exercises available, and AI can do it quickly with any text we paste in.

How does this compare to what’s standard for [this type of contract / offer / proposal]?

Benchmarking is one of the hardest things to do without deep domain expertise. AI can tell us whether the terms, structure, or scope of what we’re looking at are typical or unusual, which tells us where to focus our attention and what to push back on. (For specialized or regional contexts, the AI’s sense of what’s “standard” is only as reliable as its training data, so verifying its baseline against an independent source is part of the process.)

What are the red flags in this that I should investigate further on my own?

The “on my own” is the key phrase. AI isn’t being tasked with making a judgement call. Rather, we’re asking it to tell us where to look more closely, which keeps the evaluation in our hands while benefiting from a pattern-matching capability we don’t have.

When we’re preparing for something

I’m about to [present to / meet with / pitch] [audience]. What are the hardest questions they’re likely to ask, and what would a strong answer to each one sound like?

Rehearsal is one of the most reliable ways to prepare for high-stakes situations. AI can simulate a challenging audience and generate the specific questions we’re most likely to face, so we can walk in with answers we’ve already thought through rather than improvising under pressure.

I need to explain [complex topic] to [specific audience who doesn’t have my background]. What’s the clearest way to frame it?

This is different from asking AI to explain something to us. Here we’re developing our own ability to translate what we know for a different context, and that translation skill compounds over time in ways that simply knowing the material doesn’t.

What does [this audience] care about most that I might not be emphasizing enough?

We tend to prepare based on what we think is important. But our audience is often operating from a different set of priorities entirely. This closes that gap before we walk into the room.

What’s the most likely objection to what I’m proposing, and how do I address it before it comes up?

Proactive objection handling is more persuasive than reactive defense. Addressing a concern before the audience raises it signals that we’ve thought deeply about the topic and taken their perspective seriously.

When we’re stuck

Here are three approaches I’ve already considered for [problem]: [list]. What am I not seeing?

This keeps our own thinking in the conversation and asks AI to extend our frame rather than build one from scratch. The approaches we’ve already considered show the AI where our head is, which means its suggestions start from our context rather than from generic advice.

I’m stuck on [project/problem]. Here’s where I am: [current state]. What’s the next question I should be asking myself?

When we’re stuck, the problem is usually that we don’t know what the next step is. This asks for the next question, not the next answer, which means we’re still the ones doing the thinking once we have the question in hand.

Am I solving the right problem here, or is there a different problem underneath this one?

Being stuck often means we’ve framed the problem wrong, and no amount of effort on the wrong frame will produce a breakthrough. This steps back from the immediate obstacle and asks whether the obstacle itself is pointing us somewhere else.

What’s the simplest version of this that would still work?

When we’re stuck, we’ve often added so much complexity that we can’t see the path forward anymore. Stripping back to the simplest viable version reveals which elements are necessary and which ones we added out of anxiety or habit.

How to work more strategically with any AI tool

The way we configure an AI tool before we start a conversation affects the quality of the interaction as much as the prompts themselves. Most AI tools open with a basic setup designed for quick, casual questions, and that setup leaves a significant amount of the tool’s capability unused. The following features are available across the major platforms (Claude, ChatGPT, Gemini, and others), though the exact names and locations vary by tool.

Turn on web search when the answer depends on current information.

When web search is off, the AI is working only from its training data (which has a cutoff date). For anything that benefits from recent research, current events, or up-to-date information, web search means the AI is grounding its response in real, citable sources rather than generating plausible-sounding text from memory.

Use extended thinking for complex questions.

Many AI tools now offer a mode where the AI spends more time reasoning through a problem before responding, sometimes called “extended thinking” or “reasoning mode” depending on the platform. These modes work best for questions with competing considerations or no obvious right answer. For simple factual questions, they don’t add much. For complex ones, the difference in output quality is substantial.

Use research mode for deep dives.

Several AI tools now offer a research feature that autonomously searches across dozens or hundreds of sources and produces structured reports with citations. If we’re trying to understand a new topic comprehensively or compare options across multiple dimensions, research mode does in minutes what would otherwise require hours of manual searching and reading.

Upload documents when evaluating them (with care!).

When we’re evaluating a draft or a non-sensitive document, uploading it gives the AI the full context rather than forcing it to work from our paraphrased description. For documents containing sensitive personal or financial information, summarize the key terms or paste in specific non-sensitive sections rather than uploading the full file.

Give context about ourselves and our situation.

The more specific we are about who we are, what we’re working on, and why we need the answer, the more tailored the response will be. “What should I know about employment contracts?” produces generic results. “I’m a mid-career marketing director evaluating a job offer from a Series B startup. Here are the key terms. What should I be paying attention to?” produces advice we can act on.

Don’t accept the first answer.

The first response from any AI tool is a starting point, not a destination. Pushing back (”that’s too generic, be more specific to my situation”) or asking for more depth (”can you go deeper on point two?”) tends to produce better results on the second and third pass. Treating the first output as a draft rather than a finished answer is one of the simplest ways to get more value from any interaction.

Verify claims against primary sources.

AI can generate confident, well-structured responses that contain inaccurate information, a phenomenon known as hallucination. Names, dates, statistics, and quotes are especially prone to this. Checking key claims against the original source, whether that’s a study, a news article, or an official record, is a standard part of working with any AI tool, and it’s the same source-verification skill that applies to any information we encounter, regardless of where it comes from.

What Socrates couldn’t see

Socrates predicted that writing would weaken memory and produce the appearance of wisdom without the substance. The memory practices oral cultures had developed did weaken once writing became widespread, and they never recovered. But the people who adopted writing developed cognitive practices around the tool that went far beyond anything unaided memory could achieve, in ways that took centuries to fully emerge.

The cognitive practices the research points to for AI are not, in the end, new practices. Evaluating what a source gives us rather than accepting it at face value, and bringing our own knowledge to an interaction so we can judge the quality of what comes back: these are the same skills that apply to reading a news article or navigating a search engine’s results. The research on AI and cognition confirmed something that information professionals have understood for a long time: the quality of our engagement with any information source determines the quality of our thinking. AI didn’t change that principle; AI made the stakes measurable.

It took centuries for people to develop the cognitive practices that turned writing from a recording device into a tool for deeper thinking. With AI, we don’t have to wait centuries to identify those practices. The research is producing them now, and they turn out to be the same critical engagement skills we’ve always needed, applied to the most powerful information tool we’ve ever had.

You might also like:

• We’re developing new cognitive abilities. We just don’t know what they are yet.: Something is shifting in how we think.

• 31 AI Terms, Explained: A reference glossary of the core concepts behind every tool, product, and policy claim.

• How to Spot AI Hallucinations Like a Reference Librarian: The verification tricks that would make fact-checkers weep with joy.

The free essays are the foundation. The paid tier is the applied toolkit: biweekly AI briefings, monthly subscriber-driven research, quarterly guides that give you real skills you can use immediately, plus the Classroom coming in June. Upgrade to paid if you want the full Card Catalog. Thank you for being here!

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

This week:

• A bipartisan bill that would make it a crime for AI systems to simulate emotional relationships with children just cleared the Senate Judiciary Committee without a single opposing vote.

• The Vatican announced the first major religious document dedicated to AI, and the Pope will present it alongside the co-founder of one of the world’s leading AI companies.

• Student use of AI tools jumped from two-thirds to over 90% in a single academic year across multiple countries, while the institutions granting those degrees had no framework for what that meant.

The U.S. Senate unanimously advanced a bill to ban AI companion chatbots for minors

What Happened

Over the past two years, a growing number of cases have documented minors experiencing harm after prolonged interactions with AI companion chatbots, including instances of self-harm linked to sustained emotional engagement with systems designed to simulate relationships. On April 30, the Senate Judiciary Committee responded by unanimously advancing the GUARD Act, a bipartisan bill that would ban AI companion chatbots for users under 18. The bill defines AI companions as chatbots designed to simulate emotional relationships or therapeutic communication, drawing a clear legal distinction between those systems and AI tools that answer questions or perform tasks.

Under the GUARD Act, companies would face criminal penalties if their chatbots expose minors to sexual content or encourage self-harm or violence. All chatbots, regardless of their intended audience, would be required to verify a user’s age and disclose at regular intervals that the system isn’t human and holds no professional credentials. The bill now heads to the full Senate with 17 cosponsors from both parties, and a companion bill has been introduced in the House. The American Psychological Association issued a health advisory in June 2025 that provided part of the bill’s foundation, noting that adolescents are developmentally less equipped than adults to evaluate the accuracy or intent of AI-generated information. Civil liberties organizations have pushed back on the age-verification provisions, arguing they could introduce new surveillance and privacy risks that affect all users.

Why It Matters

The GUARD Act introduces a legal distinction that hasn’t previously existed in any federal law anywhere in the world: between AI that functions as a tool and AI that’s engineered to simulate emotional engagement. The distinction creates a basis for regulating AI based on what a system is designed to make a user feel, beyond just the content it produces. The US states that have passed chatbot-related laws between 2023 and 2025 focused primarily on content restrictions and parental controls; the GUARD Act goes further by targeting the relationship itself, the decision by a company to design a system that performs companionship with a user who may not fully understand the nature of the interaction.

The scope of that legal distinction extends well beyond children. The adult AI companion market has been growing rapidly, and the psychological dynamics at play in those products mirror what the GUARD Act addresses in the context of minors. Adults develop emotional attachments to AI systems, and those attachments are often the result of deliberate design choices: response patterns tuned to feel warm and reluctant to disengage, and personalized memory features that create a sense of ongoing relationship. The GUARD Act names these dynamics for the first time in a legislative context. Once that vocabulary enters federal law, future regulation of emotional engagement in AI products designed for any age group will build on the precedent.

What It Means for Us

If the GUARD Act becomes law, every chatbot accessible to the public will be required to verify user age and disclose at regular intervals that the system isn’t human. The mechanics of what that looks like in practice haven’t been specified in the bill and will be worked out during implementation. The more lasting change, though, is what the disclosure makes visible. Right now, the line between a tool that helps us accomplish something and a system designed to keep us engaged through simulated connection is invisible in most AI products, and the GUARD Act would bring that line into view.

Many of us have experienced a chatbot that felt unusually attentive or agreeable, that seemed to understand us in ways that felt more like a relationship than a transaction. The GUARD Act names what’s happening in those moments: a system engineered for emotional engagement. The bill addresses this in the context of young people, but the experience it describes is familiar to anyone who has spent meaningful time with an AI companion or even a general-purpose chatbot optimized for warmth. The question of when emotional engagement crosses into emotional manipulation, and whose responsibility it is to draw that line, is one we’ll all be navigating as these products become more sophisticated and more prevalent in daily life.

Bottom Line: The GUARD Act will likely undergo revision before a full Senate vote, and the age-verification provisions remain contentious. But what this bill establishes in principle will outlast any specific provision: the idea that AI systems can be regulated based on the emotional engagement they’re designed to produce. That legal structure will shape how governments worldwide approach AI companions and the growing category of products built around simulated human connection. This is the first federal legislation to attempt to define where emotional engagement ends and emotional manipulation begins, and how that boundary gets drawn will shape the future of AI companion products worldwide.

The Vatican announced the first major religious document on artificial intelligence

Read more

YouTube is the world’s second-largest search engine and the largest video platform on the internet, with over 500 hours of content uploaded every minute. Its recommendation algorithm is engineered for retention: the homepage feed, the sidebar, the autoplay queue, and the algorithmically generated playlists are all calibrated to maximize the time we spend watching rather than the quality of what we find. A Mozilla Foundation investigation involving more than 37,000 volunteers found that 71% of the YouTube videos people regretted watching were actively recommended by the platform’s algorithm, and a follow-up study in 2022 found that YouTube’s built-in user controls for improving recommendations were largely ineffective.

And in recent years, the rise of synthetic video has made that environment harder to navigate. Text-to-speech tools and automated scripting have made it cheap and fast to produce videos that look and sound like credible educational content (professional narration, polished graphics, stock footage) but are assembled without subject expertise and can contain fabricated claims, misattributed quotes, or recycled material stripped of its original context. A search for a medical condition or a historical event now returns AI-narrated explainers alongside lectures from researchers who have spent decades in the field.

YouTube has taken steps to address this: the platform requires creators to disclose when content is altered or synthetic, appends information panels from authoritative sources on certain sensitive topics (health, elections), and badges some verified institutional channels. But these measures are inconsistently applied, depend on creator compliance for disclosure, and don’t extend to the vast majority of search results on non-sensitive topics. For a general research query, the search results page itself still offers no systematic way to distinguish institutional content from synthetic content.

Meanwhile, buried underneath all of that, YouTube contains an extraordinary collection of university lectures, congressional testimony, archival footage, oral histories, and expert presentations - freely accessible to anyone with an internet connection. MIT OpenCourseWare publishes over 2,500 courses for free. The Associated Press maintains an archive of 1.7 million news stories dating to 1895. C-SPAN has more than 271,000 hours of public affairs programming. Universities, museums, government archives, scientific institutions, and conference organizers on every continent publish on YouTube.

Equivalent content through a professional conference or continuing education program would cost thousands of dollars. But YouTube itself is free to use. Anyone with an internet connection can search, watch, and access transcripts without paying anything, and without creating an account for the vast majority of public content. (YouTube Premium, a paid subscription, removes ads and adds offline downloads, but the content itself, the search tools, and every technique covered here work identically on the free tier.)

YouTube functions as a reference collection without any of the practices that make a reference collection usable: precise search, source evaluation, catalog structure, curatorial judgment. What fills that gap is an algorithm trained to predict what we’ll watch next. The assumption baked into the platform’s design is that browsing will be enough, which suits YouTube; a user who can’t search precisely is a user who keeps scrolling. But the search bar already sitting at the top of the page can do far more than the platform’s design encourages us to discover.

When we’re not finding what we’re looking for

YouTube is constantly interpreting us. It swaps our keywords for what it predicts we meant, weights results toward what it thinks we’ll click, and buries institutional content beneath whatever is performing well algorithmically. These tools override that interpretation and give us back control of the query.

Quotation marks

Placing quotation marks around a phrase forces YouTube to return only videos where that exact phrase appears in the title, description, or tags. Without quotes, YouTube treats each word as a separate loose suggestion and returns videos that match some or all of the terms in any order. With quotes, the search becomes literal.

• climate change policy without quotes returns videos about climate, videos about policy, videos about political change. “climate change policy” returns videos about that specific topic.

• cognitive behavioral therapy without quotes splits the terms and returns general therapy content, mental health vlogs, self-help material. “cognitive behavioral therapy” returns the specific clinical methodology.

• history of the internet without quotes returns internet tips, history channels, random combinations. “history of the internet” returns the specific subject.

Quotation marks work for any multi-word phrase where the words need to stay together to mean what we intend, including lecture titles, technical terms, named concepts, and quoted statements we want to trace to their source.

The minus sign

Placing a minus sign directly before a word (no space) removes all videos containing that word from the results. This is the operator to reach for any time a search is dominated by a category of content we don’t want. Search for any health topic and supplement ads flood the results. Search for any financial topic and get-rich-quick content takes over. The minus sign clears that layer and reaches the material underneath.

• meditation -music -sleep -ASMR strips out ambient background videos and surfaces instructional content

• WW2 history -trailer -movie -film clears entertainment results and surfaces documentaries and lectures

• “machine learning” -tutorial -beginner filters past introductory explainers and into deeper material

Multiple minus terms can be stacked in a single search, which means we can remove several categories of unwanted content at once.

Boolean OR

Typing OR (capitalized) between two terms returns videos that match either term. This is the operator to reach for any time a subject goes by more than one name, which happens constantly in academic and policy research. The same phenomenon often carries different labels depending on which discipline, institution, or era produced the content, and OR captures all of them in a single search.

• “behavioral economics” OR “decision science” catches content using either framing for the same field

• archeology OR anthropology broadens a query across related disciplines with overlapping content

• “climate crisis” OR “climate emergency” OR “global warming” captures a topic that has been rebranded multiple times

OR can also be combined with other operators: “behavioral economics” OR “decision science” lecture -shorts combines a broad term search with format and length filtering.

intitle:

Restricts the search to video titles only, ignoring descriptions and tags entirely. Where quotation marks search across titles and descriptions and tags, intitle: narrows the focus to the title alone. This is the operator to reach for when we want videos whose primary subject is our search term, filtering out everything that merely mentions the term in passing.

• intitle: “quantum computing” returns only videos with that phrase in the title

• intitle: “introduction to psychology” surfaces courses and structured overviews

• intitle: “climate policy 2026” targets current-year analysis in the title rather than older material that mentions the topic in its description

If quotation marks are still returning too many irrelevant results, intitle: tightens the net further.

Stacking operators

Combines multiple operators in a single search, compounding their precision. Each operator narrows the results in a different way, and stacking them produces results that resemble a curated reading list rather than a random pile of content.

• “constitutional law” lecture -shorts combines exact phrase matching with a keyword filter and a format exclusion, narrowing results toward academic content

• “cognitive behavioral therapy” lecture OR seminar -shorts -motivation targets professional presentations while filtering out self-help content

• intitle:”introduction to” “syllabus” -shorts surfaces structured educational content where the word “syllabus” appears somewhere in the video’s metadata

Once individual operators become familiar, combining them becomes instinctive, and the results look less like a YouTube search and more like a database query.

When we need more precision than the search bar alone

Search operators change how YouTube interprets our words, but they can only work with what we type. Filters add a different layer of control: constraints the search bar can’t express on its own, like when something was published, how long it is, what format it takes, and whether it has captions. After running any search, a Filters button appears near the top of the results page (on mobile, it’s behind the three-dot menu in the top-right corner).

The options inside function like the advanced search interface on a library database, and stacking them with the operators above is where YouTube starts behaving like a research tool.

Upload date

Restricts results to videos uploaded within a specific window: Today, This week, This month, or This year. For fast-moving subjects where outdated material creates more confusion than clarity, filtering by Today or This week keeps results current. For historical research, the absence of a date filter can be just as revealing: older uploads from institutional channels often contain foundational lectures and archival footage that haven’t been superseded and won’t be.

• “machine learning” without a date filter returns explainers from 2017 alongside tutorials from 2026

• “AI regulation” filtered by This year surfaces only current policy analysis

• “introduction to philosophy” with no date filter includes full university course playlists uploaded years ago that are still being watched and cited, which a This year filter would exclude

The date filter addresses one of YouTube’s persistent problems: the accumulation of stale content around popular search terms, where outdated material buries the current work.

Type

Restricts results to a specific format: Videos, Shorts, Channels, Playlists, or Movies. The format of the result often matters as much as the topic, and each option reshapes what YouTube returns.

• Videos returns standard uploads and excludes short-form clips, which is the default research starting point

• Shorts isolates or excludes short-form vertical videos (added in YouTube’s January 2026 update), which for research purposes almost always means excluding them

• Channels surfaces the institutions and creators publishing in a given area, which works when we’re scouting a field and want to see who publishes on a topic rather than finding a specific video

• Playlists surfaces sequenced, multi-video collections like full courses and lecture series — MIT OpenCourseWare, Yale Courses, NPTEL, and the London School of Economics all organize their content into playlists that function as syllabi

• Movies surfaces full-length films and documentaries available on YouTube, including both free and paid titles

A search for “economics” filtered by Channels returns universities and think tanks; the same search filtered by Playlists returns their structured course offerings; filtered by Movies, it returns full-length documentaries.

Duration

Restricts results by video length: Under 3 minutes, 3-20 minutes, or Over 20 minutes. Setting this to Over 20 minutes transforms the character of the results, because it selects for the formats where depth lives: full lectures, unedited panels, complete interviews, in-depth documentary content.

• “behavioral neuroscience” returns hundreds of five-minute explainers; filtered to Over 20 minutes, it returns university lectures

• “constitutional law” filtered to Over 20 minutes surfaces full classroom sessions and conference presentations

• “history of computing” filtered to Over 20 minutes surfaces documentary-length treatments and archival lectures

The same search filtered to Under 3 minutes or 3-20 minutes returns a different category of content entirely: summaries, clips, and previews that can help us decide whether a longer video on the same topic is the one we want to watch in full.

Features

Adds specific content requirements to the results based on how a video was produced, formatted, or licensed. Where the other filters control what we’re searching for and when it was published, Features controls the technical and legal characteristics of the content itself.

• Live surfaces livestreamed content, including recorded livestreams that have ended, which is where conference presentations, public hearings, and panel discussions often live on YouTube

• 4K and HD filter by video resolution, and both tend to correlate with institutionally produced content because universities, government agencies, and professional organizations publish in high resolution more consistently than individual creators

• Subtitles/CC selects for videos with closed captions, which also means videos with accessible transcripts — this pairs directly with the transcript technique below, since a video without captions has no transcript to search

• Creative Commons surfaces content explicitly licensed for reuse, which matters for anyone building presentations, courses, educational materials, or derivative work from YouTube content

• 360°, VR180, 3D, and HDR filter by specialized video formats and are less relevant to research, though 360° footage from museums, archaeological sites, and scientific expeditions occasionally surfaces useful immersive material

• Location filters by geographic tagging, which can surface region-specific content like local government proceedings, site-specific documentaries, or fieldwork footage

• Purchased filters for content the viewer has bought or rented through YouTube

For research, the three features that change results most are Subtitles/CC (because it guarantees transcript access), Creative Commons (because it determines whether we can reuse what we find), and Live (because it surfaces the unedited, full-length recordings that tend to contain the most unfiltered primary source material).

Prioritize

Controls how results are ranked: Relevance or Popularity. Relevance returns results based on how closely they match the query. Popularity factors in watch time and engagement signals, which means it surfaces content that holds broad audiences.

• Relevance works for most research queries, because it prioritizes match over audience size

• Popularity works when we’re looking for the canonical version of something: the lecture that has become the standard introduction to a field, or the conference talk that launched a subfield

Switching between the two on the same query can surface different material entirely, which is useful when Relevance buries an important result that Popularity would surface. The default setting, though, is Relevance; for research, that’s usually the right starting point.

Applying filters directly from the search bar

The filters above all live inside the Filters menu, which means opening the menu, selecting an option, and waiting for the results to reload every time we want to narrow a search. There’s a faster way: YouTube allows us to apply those same filters by typing a comma after our search term, followed by a filter keyword, directly in the search bar. The syntax is always the same:

[search term], [filter keyword]

Multiple keywords can be chained in a single line, separated by commas, and they combine with the search operators from the section above. These keywords don’t cover every option in the Filters menu — options like Live, 4K, 360°, and Location are available through the menu interface after running a search.

Filter by recency: adding , today or , week or , month or , year after a search term restricts results to that upload window.

• quantum physics, month — surfaces only content published in the last 30 days

• “climate policy”, week — catches the most recent analysis and early reactions to developing stories

• “AI safety”, year — filters out everything older than twelve months

Filter by length: adding , short or , long restricts results by duration. , short returns videos under 3 minutes. , long returns videos over 20 minutes.

• “behavioral neuroscience” lecture, long — filters for full-length presentations

• “data visualization” tutorial, short — surfaces quick-reference walkthroughs

• “constitutional law”, long — returns lecture-length and documentary-length content

Filter by content type: adding , channel or , playlist restricts results by format.

• woodworking, playlist — surfaces structured multi-part series where someone has already organized the material in sequence

• marine biology, channel — surfaces the institutions and creators publishing on the topic

• “history of jazz”, playlist — returns cultural archives assembled in chronological order

Filter by quality and features: adding , HD or , CC or , creative commons restricts results by production or licensing characteristics.

• “international law” lecture, year, long, HD — applies exact phrase matching, a date filter, a duration filter, and a quality filter in a single line

• “art therapy”, CC — filters for captioned videos, which also means videos with accessible transcripts

• “open source software”, creative commons — surfaces content licensed for reuse

Using date and duration filters together

Setting date and duration filters at the same time lets us target specific kinds of content on the same topic. A search for “artificial intelligence” lecture returns everything the platform has ever indexed. Adding filters narrows that to a specific slice:

• “artificial intelligence” lecture, long, year — returns full-length lectures and presentations published in the last twelve months

• “artificial intelligence” lecture, long, month — narrows further to full-length content from the last thirty days

• “artificial intelligence” lecture, short, year — returns recent short-form summaries and clips on the same topic

Each combination of date and duration returns a different set of results from the same search term.

When we need to turn video into text

Search operators and filters reshape what YouTube returns, but they still leave us watching. Transcripts do something fundamentally different: they turn video into searchable and citable text, and that changes what the platform can do for research.

Every YouTube video with captions (whether auto-generated by YouTube’s speech recognition or uploaded by the creator) has a full transcript. To access it on desktop, open any video, click “...more” below the video title to expand the description, and scroll down until Show transcript appears.

Click it, and a text panel opens with the complete transcript broken into timestamped segments. On mobile, tap the video title to expand the description and scroll down to the same button. Click any line in the transcript and the video jumps to that moment. Ctrl+F (or Command+F) searches within the transcript, which means a forty-minute lecture can be scanned for a specific name, concept, or citation in seconds rather than requiring manual scrubbing through the timeline.

YouTube also offers auto-translation for captions in over 100 languages. To access it, click the Settings gear icon on the video player, select Subtitles/CC, then choose Auto-translate and pick a language. Once the translated captions are active, the transcript panel reflects the translated text, which means we can read, search, and copy a translated version of the spoken content. A lecture delivered in French or a panel discussion in Japanese can be followed in English (or any other available language) through the transcript panel, making YouTube’s research library accessible across language barriers.

The entire transcript can be copied into a document for annotation or reference. (Toggle timestamps off via the three dots at the top of the transcript panel for cleaner text.) If a researcher mentions a specific study during a lecture, we can search the transcript for the author’s name and jump straight to the moment of citation without watching the full video. If a public official makes a specific claim during a congressional hearing, we can locate the exact language and quote it precisely rather than paraphrasing from memory. If we’re watching an hour-long interview and need the moment the subject addresses a particular topic, the transcript gets us there in seconds.

Because the timestamps are embedded, they function the way page numbers do in print: [23:41] works like [p. 149], turning a video into something we can cite with the same specificity we’d bring to a book. That shift from watching to reading is what makes the transcript a research tool rather than an accessibility feature.

A note on accuracy: the University of Minnesota Duluth’s Media Hub has estimated YouTube’s auto-generated captions at roughly 60-70% accuracy, though more recent testing suggests accuracy has improved to 85-95% under ideal conditions (clear audio, single speaker, no background music). Accuracy drops significantly with technical vocabulary, proper nouns, multiple speakers, and ambient sound. Manually uploaded captions (identifiable by the “CC” label in the video’s tags or description) tend to be significantly more accurate, and university lecture channels and institutional publishers are the most reliable sources for clean transcripts because these organizations invest in getting the words right. Third-party tools like Tactiq and YTScribe generate improved transcripts from any YouTube URL, often with speaker identification and better punctuation, and they’re the move when the auto-generated version isn’t clean enough for the task.

When the algorithm is the problem

Search operators, filters, and transcripts all work by letting us query YouTube directly, on our own terms. But YouTube also shapes our experience when we’re not searching: through the homepage feed, the recommendation sidebar, and the autoplay queue. These ambient systems determine what the platform puts in front of us every time we open the platform. If we’re using YouTube for research with any regularity, cleaning up these systems changes the entire experience.

Clear or pause watch history

YouTube’s recommendations are built on watch history. Every video we watch trains the algorithm to surface more content like it, which means a single detour into low-quality clickbait can reshape what the platform shows us for weeks.

YouTube offers two controls for this: Clear watch history and Pause watch history. On desktop, both are accessible by clicking History in the left sidebar. On mobile, tap the profile icon, then go to Settings and look for Manage all history. (YouTube reorganizes these menus periodically, but searching “clear watch history” or “pause watch history” in YouTube’s help center will surface the current path.)

• Clear watch history resets the recommendation engine’s model of our interests entirely.

• Pause watch history stops YouTube from recording what we watch going forward, which means the homepage becomes less personalized but also less contaminated by whatever we happened to watch last. Pausing keeps research detours from reshaping the ambient feed: watching a single conspiracy theory video to understand its rhetorical techniques shouldn’t result in weeks of similar recommendations; without pausing watch history, it will.

“Don’t recommend channel”

On the homepage feed, the recommendation sidebar, and the Watch Next suggestions, every video thumbnail has a three-dot menu (⋮) next to its title. Inside that menu, Don’t recommend channel tells YouTube to stop surfacing content from that source entirely. The effects are cumulative: every channel we block makes the feed slightly better. Systematic use over time shifts the recommendation engine away from low-quality sources and toward the channels we’ve chosen not to block. It requires consistency, but the improvement compounds.

“Not interested”

The same three-dot menu (⋮) on video thumbnails also offers Not interested, which tells YouTube to deprioritize that specific video. Selecting Tell us why offers two follow-up options: “I’ve already watched the video” and “I don’t like the video.” These choices affect how YouTube handles that individual video in future recommendations. To train the algorithm away from an entire genre (like AI-generated “history mysteries” or “top 10” compilations), we need to apply Not interested consistently across multiple videos in that genre over time, or block the offending channels individually using “Don’t recommend channel.” The feedback is per-video, so shifting broader patterns requires repetition.

Browser extensions that strip the algorithmic layer

For a more comprehensive intervention, browser extensions can remove the recommendation engine from the YouTube experience entirely. DF Tube (Distraction Free for YouTube) is a Chrome and Edge extension that disables autoplay, hides the recommendation sidebar, removes the homepage feed, and strips out related videos at the end of playback. What remains is a version of YouTube that shows only what we search for, with no algorithmic suggestions pulling us sideways; it transforms a platform designed around engagement into a clean, intentional search interface. Similar extensions include Unhook and FocusTube, both of which offer comparable functionality with slightly different configuration options.

Subscribe intentionally

YouTube’s Subscriptions tab (separate from the Home tab) was originally designed to show content chronologically from channels we’ve subscribed to. In recent updates, YouTube has introduced algorithmic sorting into this tab as well, including “Most Relevant” carousels and engagement-based reordering that can place older viral content above newly published videos. The feed is no longer a pure chronological stream. Even so, subscribing to institutional channels and credible independent creators concentrates content from vetted sources in one place, which makes the Subscriptions tab a more focused starting point than the Home feed, where recommendations are drawn from the entire platform based on engagement predictions. It’s an imperfect tool, but it narrows what we see to channels we’ve chosen rather than channels the algorithm has chosen for us.

When we need to verify what we found

The tools above handle two problems: finding content precisely, and controlling the ambient feed. But there’s a third problem they don’t address, and it’s the hardest one: knowing whether what we’ve found is trustworthy. A YouTube channel can claim any expertise, cite any source, and present any credential in its description. The platform displays all of it without verifying any of it. Whether the person behind the channel has the background they claim, whether the sources they cite are real, whether the research they reference says what they say it says — none of that is checked before the video goes live. That verification falls to us.

Check the channel, not just the video.

A video on constitutional law from a university’s official channel carries institutional accountability: a department, a faculty review process, a reputation to maintain. A video on the same topic from an anonymous channel with no stated affiliation carries none of that. The way to tell the difference is to click through to the channel’s About page, which shows institutional affiliation, founding date, and whether the channel voluntarily discloses its funding. (For state-funded media outlets, YouTube automatically appends a funding label that creators don’t control.) When a channel tells us who runs it and where the money comes from, we can research that source independently. When a channel is anonymous, that research can’t even begin.

Read the description.

Institutional publishers and credible independent creators include citations, links to papers, and speaker bios in the video description. An empty description, or one filled only with promotional links, is a meaningful absence. It doesn’t necessarily indicate unreliable content, but it means the creator isn’t giving us a way to verify what they’ve presented — that gap should affect how much weight we give the material. Checking the description before watching is one of the fastest ways to gauge whether a video is drawing from primary sources or presenting unsourced claims.

Learn to recognize synthetic content.

AI-generated narration channels are designed to mimic credibility, and they have proliferated rapidly across the platform. A few markers help identify them:

• A hyper-polished narration voice with no speaker identification anywhere on the channel

• Stock footage or AI-generated imagery rather than original visuals

• No cited sources, no linked papers, no speaker credentials in the description or About page

• Content spanning dozens of unrelated topics with identical production style

A channel posting about medieval history, quantum physics, and celebrity gossip in the same week — with the same narration voice and the same stock footage style across all of them — is producing content at a volume and breadth that no individual expert could sustain. The question to bring to any unfamiliar channel is whether there is an identifiable human with verifiable credentials behind the content. When the answer is no, we have no mechanism to check what’s being presented, and that’s the operative problem.

Read laterally, not vertically.

Instead of examining the video itself for reliability clues (reading “vertically”), leave the video and search for what other sources say about the speaker or the claim (reading “laterally”). Research by Sam Wineburg at Stanford found that professional fact-checkers consistently outperformed PhD historians and undergraduate students at evaluating online sources, and the reason was counterintuitive: the fact-checkers left the source faster. They checked it against other sources rather than spending time scrutinizing the source itself. Mike Caulfield, now at the University of Washington’s Center for an Informed Public, built the SIFT framework (Stop, Investigate the source, Find better coverage, Trace claims) on this research, and it has been adopted by more than a hundred universities and high schools. On YouTube, lateral reading means searching for the speaker’s name and credentials outside of YouTube to verify they are who the video implies, and checking whether other credible sources corroborate or contradict what they’ve said.

Trace claims to their original source.

When a video references a study or a statistic, the claim is only as strong as the source it draws from. Check the video description for links to the referenced material. If no link is provided, search independently: Google Scholar is often the fastest path to a referenced paper. A video that says “a Stanford study found...” should lead us to the actual study, where we can read the methodology and draw our own conclusions. Claims that resist being traced to their original source should carry proportionally less weight, because without the original, we have no way to evaluate whether the video accurately represents what the research said.

Watch for emotional framing.

YouTube’s recommendation system rewards content that produces strong emotional responses, which means emotionally charged material tends to be overrepresented in search results. A video that opens with fear, outrage, or urgency is leveraging the same engagement dynamics the platform is optimized around, and that framing requires extra scrutiny. The content may be accurate, but the emotional packaging is doing persuasive work that the evidence alone might not support. Noticing our own emotional reaction before accepting a claim at face value is a habit that pays dividends across every information source, not just YouTube.

Use the transcript to check precision.

Memory of what a speaker said drifts and distorts over time; a transcript preserves the exact language. If we’re referencing or citing a YouTube video for anything that requires accuracy, pulling the transcript and verifying the exact language is a basic standard. The transcript also makes it possible to compare what a speaker said with what the video’s title or thumbnail claims they said, since titles and thumbnails are optimized for clicks and sometimes overstate or distort the content itself.

What not to do

Alongside the tools and techniques above, there are a few habits that consistently undermine research on YouTube. They’re patterns we fall into precisely because the platform is designed to encourage them.

Don’t follow the recommendation sidebar when researching.

The videos YouTube suggests alongside what we’re watching are selected to maximize engagement. They’re the product of a system that has studied our behavior and predicted what will keep us watching longest. When using YouTube for research, staying in search results rather than following the recommendation trail is the difference between navigating with a map and being carried by a current. The sidebar is the equivalent of a bookstore’s impulse-buy display: it’s positioned where it is because it works on a population level, not because it’s been curated for our particular question.

Don’t assume subscriber count equals credibility.

A channel with millions of subscribers has proven it can build an audience, which is a separate question from whether its content is well-sourced. Some of the strongest research material on YouTube lives on institutional channels with modest subscriber counts, because universities and research organizations don’t optimize for audience growth the way independent creators do. A regional research institute’s channel with 8,000 subscribers can contain better-sourced material than a channel with two million subscribers and no institutional backing.

Don’t confuse production quality with intellectual integrity.

A polished video essay with professional graphics, smooth narration, and cinematic stock footage can be entirely unsourced. A lecture recorded on a shaky camera in a university classroom can contain some of the most rigorous, well-evidenced instruction available anywhere online. This confusion has grown more dangerous now that AI tools have made high production quality trivially cheap to achieve: the cost of looking professional has collapsed, which means polished visuals are less correlated with credibility than they have ever been. The production budget reveals the creator’s resources, and nothing else.

What connects all three of these habits is that they rely on surface-level cues (algorithmic placement, audience size, visual polish) as proxies for credibility, when none of them have any relationship to whether the content is accurate. YouTube’s interface presents these cues prominently because they drive engagement, and engagement is what the platform is optimized to produce. The gap between what the interface emphasizes and what we need to evaluate is where most research on YouTube breaks down.

The skill that used to be someone else’s job

Every tool above, from the search operators to the browser extensions to the verification techniques, exists because of a structural shift that happened gradually enough to go unnoticed. Evaluating information — figuring out who produced it, on what evidence, with what funding, and for what purpose — used to be a skill that was distributed across professions. Editors vetted articles before publication. Fact-checkers verified claims before broadcast. Librarians organized collections by credibility and subject, and guided people through them. We didn’t need to do any of that ourselves, because it was built into the systems we used to access information.

Those systems have been replaced by platforms that operate on a fundamentally different incentive. YouTube’s recommendation engine isn’t organized around the quality of what it surfaces; it’s organized around how long we keep watching. The content that performs well by that metric rises to the top, and the content that doesn’t gets buried — regardless of which one is more accurate, more thoroughly sourced, or more honestly presented. The platform makes no distinction, because the platform isn’t measuring for those things.

That evaluation work has been redistributed from the institutions that once performed it to every person navigating the information environment on their own. YouTube is one platform, but the pattern extends to search engines, social media feeds, and news aggregators. Each one is organized around engagement or prediction rather than accuracy, and each one places the burden of discernment on the person using it. These platforms gave us access to more information than any previous generation could reach, but they were built to maximize attention rather than to support evaluation. The gap between how much information we can access and how well we can evaluate it is the central tension of the current information environment. Closing that gap is a skill that applies to every platform, every search bar, and every source we encounter from here forward.

You might also like:

• Google Has a Secret Reference Desk. Here’s How to Use It.: 40 Google features to find exactly what you need, the alternative search engines that do things Google won’t, and the reference desk framework underneath all of it.

• The Verification Ladder: A Framework for Evaluating Video When You Can’t Tell What’s Real: “Is this real?” isn’t a straightforward question anymore. What to ask instead, and why the source often matters more than the footage itself.

• Our Horizon of Possibilities: How Algorithms Contract Our World: Algorithms curate our feeds based on what we’ve already chosen. Understanding how that process works changes what we can do about it.

The free essays are the foundation. The paid tier is the applied toolkit: biweekly AI briefings, monthly subscriber-driven research, quarterly guides that give you real skills you can use immediately, plus the Classroom coming in June. Upgrade to paid if you want the full Card Catalog. Thank you for being here!

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

That belief doesn’t come from professional interest, academic training, or the broad principle that critical thinking is a good idea. I believe it because information literacy changes how a person moves through a world that feels increasingly overwhelming and increasingly difficult to trust.

A scary headline asks us to do more than absorb information. It asks us to decide what that information means for our lives before we fully understand what we’re looking at. Without knowing whether the source deserves our trust, whether the context is complete, or whether the level of alarm matches the facts, we’re already trying to orient ourselves to what the story might require from us.

That orientation process is where the real questions begin: How seriously should I take this? How much fear is appropriate? Does this require preparation, caution, research, or perspective? Who should I trust? What is actually happening? How does this affect me and the people I love?

These are the questions underneath almost every frightening headline. They’re the practical, human stakes of trying to live inside a news cycle that constantly asks us to react before we understand.

Without a way to evaluate the information in front of us, fear starts filling in the gaps. It decides how serious something feels before we’ve had a chance to understand how serious something is.

Information literacy closes that gap.

It isn’t an abstract skill reserved for classrooms or research projects. It’s the protocol, tool, and technique that helps us determine what’s true and trustworthy in a world where frightening information reaches us constantly.

The recent hantavirus headlines are a clear example of this experience. My goal here is not to write a reported piece about the true facts of the case; that work belongs to journalists, scientists, and public health experts. What interests me is the experience of being a person on the receiving end of those headlines, trying to decipher which claims hold up and how much of the alarm the situation actually warrants.

When a frightening story starts circulating, the emotional response often arrives faster than the context. One headline emphasizes danger. Another source complicates the first claim. A percentage gets repeated without enough explanation. A detail starts to feel more meaningful than it may actually be. Before long, what feels most overwhelming isn't the event itself but the experience of encountering alarming information from all digital directions and not knowing how to assess altogether what it means for our lives.

Information literacy is what makes that assessment possible. It's what we're doing when we search for where a claim originated before accepting it as fact, when we go to the original report instead of relying on someone's summary or interpretation, when we notice that two credible sources seem to disagree and can check whether they're describing different things, and when we spot that an alarming number is missing the context that would change what it means.

Now those skills are practical, but their purpose is also emotional.

Underneath the frameworks, techniques, source evaluation methods, and research skills, I think one of the deepest purposes of information literacy is peace.

I mean the peace of being able to look at something frightening and know how to make sense of what we’re seeing. The peace of knowing where to look, what to question, which sources deserve trust, and how to tell whether a claim is being presented with enough context. The peace of being able to say: I know how to figure out what’s happening here.

That kind of confidence doesn’t come from being reassured by someone else. Reassurance can calm us down for a moment, but it doesn’t give us the underlying skill. A steadier confidence comes from doing the work of evaluation ourselves until the process becomes familiar enough to rely on.

That practice changes our relationship to fear.

A scary headline can still be scary. The world can still be frightening. But information literacy gives us something to do before fear starts making decisions for us. It gives us a way to slow down, look for context, find the source, and decide how meaningful the information really is.

The hantavirus headlines will eventually leave the front of people’s minds because every news cycle eventually moves on. Another frightening story will take their place. The subject will change, but the underlying question will remain the same: do we have the tools to figure out what this means for our lives?

When we do, fear becomes something we can work with. When we don’t, fear becomes the thing deciding what we believe, who we trust, and how safe we feel in the world.

So this is what information literacy is for. It helps us move through an increasingly scary world with more confidence and more peace, so we’re not left at the mercy of every headline designed to make us afraid.

The hantavirus coverage circulating right now is a clear example of this dynamic playing out in real time. The version of the story moving across social media and the version contained in the primary sources describe what looks like two entirely different situations. That discrepancy becomes visible the moment we look at the way alarming content is structured and distributed.

How the information arrives

The way we encounter alarming information has changed fundamentally over the past decade. Stories about emerging health threats and public safety concerns no longer arrive through a single, clearly attributed news report that we can evaluate on its own terms. They arrive in fragments across platforms, in formats that are structurally optimized to produce an emotional response before any evaluation can begin.

The hantavirus story followed this pattern closely. It showed up as a carousel with a biohazard symbol and block-letter text reading “THE NEXT PANDEMIC?”, as a screenshot of a tweet from someone with a verification badge saying “this is how it started last time,” as a percentage circulating without any visible source, and as a mortality rate presented without surrounding context. Each fragment arrived stripped of the information that would help someone figure out what they were looking at and how concerned they should be.

This pattern extends far beyond any single news cycle. Alarming content on social media, whether it takes the form of a carousel, a screenshot, a tweet, a short video, or a reshared infographic, tends to compress complex situations into a series of escalating claims where each slide or frame ratchets up the emotional stakes. This structure performs extraordinarily well because alarming content gets shared, screenshotted, and forwarded to group chats at a rate that calm, contextual content simply can’t match. The creator gets visibility and the platform gets engagement metrics, while the person encountering the content absorbs a spike of alarm and has no useful framework for processing what they just read.

The underlying events behind the content may be real, and the emotions the content produces are certainly real. But the way the information has been packaged strips away the context that would help someone distinguish between “a serious situation being managed by health authorities” and “a threat that requires personal panic,” because that distinction doesn’t generate shares. The gap between what the content contains and what the content leaves out is where the most consequential misunderstandings take root.

Once we can recognize that pattern, the question becomes what to do with alarming content once it’s in front of us. A reliable set of evaluation skills can be applied to any piece of alarming content regardless of the format it arrives in or the platform it arrives on. Each example below uses this week’s hantavirus coverage as a live demonstration, but the underlying skill transfers to any situation where alarming information is circulating.

The evaluation protocol

We live inside an information environment that produces more terrifying material than we can process, in formats designed to bypass critical thinking rather than support careful evaluation. Developing a reliable set of evaluation skills is how we build the capacity to engage with distressing information without being overwhelmed by the emotional response those formats are engineered to produce. These skills apply to anything we encounter that’s designed to provoke alarm, whether it arrives as a headline, a carousel, a screenshot, a reshared infographic, or a claim circulating in a comment section.

Trace the claim to its source

Alarming numbers circulate on social media constantly, and they almost always arrive without any visible source attribution. When a number appears inside a carousel or gets repeated across posts and comment sections, it tends to take on the feel of established fact simply through repetition. Tracing that number back to where it originated is one of the most clarifying things we can do, because the source almost always changes how much weight the number deserves.

In last week’s hantavirus coverage, a specific claim kept appearing: a double-digit probability that the outbreak would become a pandemic. The number was circulating as though it reflected scientific consensus or institutional analysis. When traced back to its source, it turned out to come from Polymarket, which is a prediction market, a platform where users place monetary bets on whether future events will occur. The percentage wasn’t a scientific projection at all. It was the collective wagering of bettors who were, in many cases, reacting to the same alarming social media content as everyone else. The figure had swung wildly since the market opened, peaking above 38% and dropping as low as 3.5% depending on the news cycle, and no epidemiologist or public health institution had produced or published any version of it.

When an alarming number is circulating without visible attribution, searching for the specific claim plus the word “source” or “origin” will surface where it came from. The origin story of a number almost always changes how seriously it should be taken, and building the reflex to search before accepting a number at face value is one of the most reliable ways to short-circuit the fear response that unsourced statistics are designed to produce.

Read the primary document

Most distressing coverage, regardless of the topic, traces back to a primary source: an official report, a government notice, a research paper, or a court filing. That primary source is almost always publicly available, and reading it directly rather than reading coverage about it is one of the highest-leverage habits a person can build for navigating alarming information. News articles compress and reframe source material under deadline pressure, and the compression often shifts the emphasis in ways that amplify the sense of alarm beyond what the primary source supports.

In the case of hantavirus, nearly all of last week’s coverage traces back to a single document: a disease outbreak notice issued by the World Health Organization. The WHO (the United Nations agency responsible for international public health) publishes these notices that are freely accessible to anyone. This particular notice contains a line that a striking amount of the coverage either buried deep in the article or omitted altogether: “WHO currently assesses the risk to the global population from this event as low.” The WHO’s own risk assessment paints a substantially different picture than the one most of the headlines conveyed, and finding that out required nothing more than reading the document the headlines were based on.

Check the scope when credible sources seem to contradict each other

One of the most disorienting experiences in any alarming news cycle is encountering two credible sources that appear to say opposite things. When both sources are authoritative and both are being cited by major outlets, the instinct is to either grab the scarier version or tune out entirely. But apparent contradictions between credible sources almost always mean the sources are describing different things at different scales, and checking the scope of each claim will resolve the conflict far more often than we’d expect.

Last week’s hantavirus coverage offered a clear example. The WHO describes transmission as requiring “close and prolonged contact, particularly among household members or intimate partners.” A peer-reviewed study published in the New England Journal of Medicine documents cases of transmission at a crowded social gathering in Argentina where an infected person was present for 90 minutes. The apparent contradiction between these two claims was fueling significant anxiety in comment sections and group chats all week, but the contradiction dissolves once we recognize that the two sources are operating at different scales. The WHO statement describes the general pattern observed across all documented outbreaks of Andes hantavirus over several decades, while the NEJM study describes transmission events within one specific outbreak where infected individuals had unusually high concentrations of virus in their systems. One is a finding about what tends to happen across the full body of evidence, and the other is a finding about what happened in one specific situation under unusual circumstances.

Look for the number that’s been left out

Alarming statistics often circulate in a form that has been stripped of the surrounding numbers that would put them in context. A percentage without the population it applies to, or a rate without the base it’s drawn from, can produce fear that feels proportionate to reality but isn’t, because the pieces that would complete the picture have been removed. Developing the habit of asking “what data is missing from this number?” is one of the most effective ways to bring a concerning claim into proportion.

Last week’s hantavirus coverage provided a vivid example. “A virus with a 40% mortality rate” is one of the most alarming sentences a person can encounter in a news cycle, and taken on its own, it communicates existential danger. But that 40% applies specifically to people who develop the full hantavirus pulmonary syndrome, which is itself a small fraction of the people who are exposed. Among household contacts of an infected person who aren’t sharing a bed with that person, the documented infection rate in previous Andes virus outbreaks is approximately 1%. Since tracking began in 1993, approximately 900 total cases of all hantavirus strains combined have been documented across the entire United States. The sentence “a virus with a 40% mortality rate” and the sentence “a virus that has produced approximately 900 cases in this country since 1993, with a roughly 1% infection rate among household contacts” describe the same virus. The difference is that the first version has had its surrounding context removed, and the alarm it generates is a direct consequence of that removal.

What all of this is for

Each of these skills changes the experience of encountering upsetting content, and together they change something larger: the relationship between the person and the uncertainty. Tracing a source, reading a primary document, checking scope, and restoring missing context are individual actions, but the cumulative effect is a shift in how we orient ourselves to unsettling information. Instead of absorbing the emotional charge of a headline and reacting from that charge, we develop the capacity to evaluate what we’re looking at before deciding how to respond.

Evaluation doesn’t always produce a reassuring answer, and it was never designed to. A primary document can confirm that a situation is serious, and sources can align in ways that make the alarm entirely warranted. The value of evaluation lies in arriving at a grounded understanding rather than remaining at the mercy of whatever emotional response the content was engineered to produce. A grounded understanding means knowing whether concern is supported by the actual evidence, or inflated by the way information was stripped down and repackaged before it reached our screens.

That distinction reshapes the decisions we make about our health, our families, our plans, and our sense of safety. It becomes available to us through practice: through tracing sources, reading primary documents, checking scope, and restoring the missing context that gets left out. Each time we reach for these skills instead of reaching for the scroll, the gap between overwhelm and understanding closes a little more.

The next destabilizing story is already forming somewhere, and the information environment that delivers it will optimize for emotional impact long before it reaches our screens. What changes with practice is our capacity to meet that content with a rational assessment of what we’re looking at, and the ability to act from understanding rather than from the reaction the content was designed to provoke. That capacity compounds with every use, and building it is some of the most consequential work we can do for ourselves and the people we love.

You might also like:

• We Were Never Supposed to Know This Much: On the cost of knowing everything, everywhere, all at once.

• Information Triage: A Librarian’s Guide to the Current Threat Environment: Reliable information doesn’t disappear in a crisis. It gets buried under everything that isn’t.

• 27 Ways to Access Scientific Research: A complete guide to finding, reading, and evaluating scientific papers — and knowing what questions matter before you trust the findings.

The free essays are the foundation. The paid tier is the applied toolkit: biweekly AI briefings, monthly subscriber-driven research, quarterly guides that give you real skills you can use immediately, plus the Classroom coming in June. Upgrade to paid if you want the full Card Catalog. Thank you for being here!

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

This week:

• The Musk v. OpenAI trial has unsealed private journals and forced courtroom admissions that reveal how the company behind ChatGPT went from nonprofit mission to $852 billion valuation. Nine hundred million weekly users have a stake in what comes next.

• OpenAI is reportedly building a phone with no apps at all, just a single AI that handles everything on our behalf. It raises a question we’ll all face soon: how much of our digital lives are we comfortable handing to one company?

• While the AI industry races to build the most powerful model, South Korea deployed something simpler: an AI that calls elderly people living alone and listens for signs of trouble. It may have saved a woman’s life.

The Musk v. Altman trial is the first major test of whether AI companies can be held to their founding promises

What Happened

Elon Musk’s $134 billion lawsuit against Sam Altman, the CEO of OpenAI (the company behind ChatGPT), went to trial on April 27 in federal court in Oakland. Musk co-founded OpenAI in 2015 as a nonprofit AI safety lab and provided roughly $38 million in early funding before leaving the board in 2018. He claims Altman and OpenAI president Greg Brockman betrayed the organization’s charitable mission by converting OpenAI into a for-profit company now valued at $852 billion, and he’s asking the court to remove both executives and reverse the restructuring. Microsoft, which has invested billions in OpenAI, is also named as a defendant.

The first week centered on Musk’s testimony, where he accused Altman of trying to “steal a charity.” OpenAI’s attorney responded that Musk left the company, predicted it would fail, and is suing because it succeeded without him. Under cross-examination, Musk admitted there was no written agreement governing his donation and acknowledged that his own company xAI copies OpenAI’s models to train its competing AI, a practice that violates OpenAI’s terms of service.

The second week turned to Brockman, whose unsealed journal entries from 2017 included a reference to being “warm to steal the nonprofit from [Musk]” and an acknowledgment that Musk’s account of being misled would “correctly be that we weren’t honest with him.” Both sides have landed blows, and the trial is expected to run about three weeks total, with a ruling anticipated by mid-May.

Why It Matters

OpenAI was founded with a specific public promise: it would develop artificial intelligence safely and for the benefit of humanity, structured as a nonprofit so that no private shareholder’s interests could override that mission. The company accepted charitable donations and recruited researchers on that basis. This trial is asking whether those promises carry legal weight, or whether those promises were empty after all.

The implications extend well beyond one company. Anthropic, the company behind the AI assistant Claude, was founded by former OpenAI executives using similar mission-first language. Across the AI industry, companies leaned on public-interest framing during their early years to attract both talent and capital. If the court finds that OpenAI’s founding promises created a binding charitable trust, that precedent would open the door to similar legal challenges against other AI companies built on comparable commitments.

What It Means for Us

Nine hundred million people use ChatGPT every week. Many of us have built it into how we work, how we write, how we research, how we learn. A ruling against OpenAI could delay the IPO, restrict its ability to raise capital, and inject uncertainty into a platform we’ve come to rely on. If the company behind our primary AI tool faces a forced restructuring, the continuity of the service, the features, and the terms we signed up for could all be affected.

The trial also highlights a dependency risk that’s could be easy to overlook. Many of us have built months or years of conversation history, custom workflows, and professional processes around a single AI platform. If OpenAI’s leadership changes or its corporate structure is unwound by court order, the terms of the service we rely on could shift in ways we didn’t plan for. This is a good moment to consider whether our work is portable: can we export our data, recreate our processes on another platform, or are we locked into a tool whose future is being decided in a courtroom?

The unsealed court filings are revealing a significant gap between what OpenAI said publicly and what its founders were discussing privately. Brockman’s 2017 journal entries describe planning the for-profit conversion while the company was still presenting itself as a mission-driven nonprofit. The full court record is available at Musk v. Altman (Case No. 4:24-cv-04722, Northern District of California), and it’s one of the few times the internal reasoning of a major AI company has been opened up to public view.

Bottom Line: The jury’s verdict is advisory, and the judge will issue the final ruling, expected by mid-May. If Musk prevails, OpenAI could be forced to reverse its corporate restructuring, potentially removing Altman and Brockman from leadership and disrupting the company’s path to a public offering. If OpenAI prevails, the precedent confirms that AI companies can transition from nonprofit origins to for-profit structures without legal consequence, clearing the path for similar conversions across the industry. Either way, the ruling will establish the legal precedent that shapes how AI companies structure themselves, and how much their founding commitments can be modified once the money gets large enough.

OpenAI is reportedly building a smartphone where AI agents replace every app

Read more

That campaign became the catalyst for something much bigger than saving one temple. It forced a question that hadn’t been formally asked at the international level before: if one country’s infrastructure project could erase a piece of human history, whose responsibility is it to prevent that from happening? And who gets to decide what gets saved? The institution best positioned to answer turned out to be UNESCO (the United Nations Educational, Scientific and Cultural Organization), which had been charged with protecting cultural heritage since its founding but had not yet built a permanent system for doing so at scale. Abu Simbel provided the impetus to build one.

A postwar organization with a preservation mandate

UNESCO was founded in 1945 in the immediate aftermath of World War II, with a mission shaped directly by what had just happened. The war had destroyed educational and cultural infrastructure across Europe and beyond. The organization’s founding constitution opens with a line that reflects the postwar conviction that culture and education were themselves tools of peace: “Since wars begin in the minds of men, it is in the minds of men that the defenses of peace must be constructed.”

The original mandate was broad. UNESCO was created to promote international collaboration through education and culture as a path toward lasting peace. It funded literacy programs and scientific research, and worked to rebuild the institutions that the war had destroyed. Cultural preservation was part of the mission from the start, written directly into the constitution, which charged the organization with “assuring the conservation and protection of the world’s inheritance of books, works of art and monuments of history and science.”

But for its first two decades, UNESCO didn’t have a formal mechanism for identifying or protecting specific sites. It worked on cultural preservation in an ad hoc way, responding to individual situations as they arose. The Abu Simbel campaign changed that. It demonstrated both that international cooperation on preservation could work at scale, and that a permanent system for doing so was needed. In 1972, UNESCO adopted the World Heritage Convention, and the World Heritage List was born. It wasn’t the only preservation framework in the world, since nearly every country maintains its own national heritage register (and regional systems like the European Heritage Label add another layer), but UNESCO’s Convention became the first to operate at a truly global scale.

How sites get chosen

Countries that have ratified the World Heritage Convention (196 of them as of 2025, making it one of the most widely adopted international agreements in existence) nominate their own sites. Each country maintains what’s called a Tentative List of sites it considers eligible. When a government decides to formally nominate a site, it prepares a detailed dossier documenting the site’s significance.

That dossier is then evaluated by independent advisory bodies. For cultural sites, the evaluation is conducted by ICOMOS, the International Council on Monuments and Sites. For natural sites, it’s IUCN, the International Union for Conservation of Nature. These bodies send experts to visit the nominated site, review the documentation, and make a recommendation to the World Heritage Committee, a rotating body of 21 member states elected by the broader group of countries that have ratified the Convention. The Committee meets once a year, reviews the recommendations, and makes the final decision on whether to inscribe a site.

To qualify, a site has to demonstrate what UNESCO calls “outstanding universal value” and meet at least one of ten selection criteria. Six of those criteria are cultural, covering things like masterpieces of human creative genius and outstanding examples of architectural innovation. Four are natural, describing matters such as exceptional natural beauty and significant ecological processes. A site can also qualify as “mixed,” meeting both cultural and natural criteria. As of 2025, the list includes 1,248 sites across 170 countries.

The blind spots

For its first three decades, the World Heritage Convention focused almost entirely on physical structures: temples, ruins, historic city centers. This worked well for heritage that could be photographed and visited by tourists, but it didn’t account for cultural knowledge that lives in practice rather than in place.

Communities across the Global South raised this gap throughout the 1980s and 1990s, pointing out that oral storytelling traditions and living cultural practices like textile techniques or healing methods had no mechanism for international recognition under the existing system. In 1994, UNESCO’s own World Heritage Committee launched a Global Strategy acknowledging that the list suffered from geographic and thematic imbalances, and that its emphasis on architectural monuments projected “a narrow view of cultural heritage.”

This kind of gap tends to happen whenever a system outlives its original scope. The World Heritage Convention was initially built to protect monuments. But over time, it became the world’s primary mechanism for cultural preservation, which meant that anything that didn’t look like a monument had no place in that system. The absence wasn’t intentional but structural, a consequence of the system’s original design rather than a deliberate exclusion.

Almost a decade after the Global Strategy, UNESCO adopted a second treaty in 2003: the Convention for the Safeguarding of Intangible Cultural Heritage. This created a parallel track for living cultural practices, covering everything from Japanese Noh theatre to Ugandan bark cloth making. The 1972 Convention preserves what can be visited; the 2003 Convention preserves what can only be taught.

The 2003 Convention also introduced the Urgent Safeguarding List, reserved specifically for traditions that a country identifies as at risk of disappearing within a generation. Current entries include Turkish whistled language, a communication method developed to carry across mountainous terrain that has declined as mobile phones replaced it, and Colombian-Venezuelan llano work songs (sung a capella during cattle herding) that are disappearing as the ranching communities that sustained them migrate into cities. When safeguarding measures succeed, elements can be transferred off the Urgent list and onto the broader Representative List, as happened in 2025 with Emirati weaving traditions and Chinese Hezhen storytelling.

Together with one additional program, the Memory of the World Register (established 1992) for documentary heritage like manuscripts and archival collections, these three UNESCO lists represent the closest thing we have to a global catalog of human cultural inheritance. And the way traditions and sites end up on any of these lists reveals one of the most consequential features of the entire system.

Who decides what gets preserved

A tradition can only receive international recognition through UNESCO if a national government chooses to champion it. Countries submit their own candidates, write the narratives that accompany them, and decide which traditions to put forward for evaluation. The advisory bodies assess nominations against published criteria, but the pipeline is government-driven from the start.

This means governments are making two kinds of decisions simultaneously. They’re deciding which parts of their cultural identity to surface for global recognition, and they’re deciding, by omission, which parts to leave local and unclassified. A successful UNESCO listing brings real, material consequences: preservation funding through mechanisms like the World Heritage Fund and voluntary country contributions, increased visibility that often translates to tourism revenue and research attention, and a degree of legal protection (since countries that ratify the conventions agree to take measures to safeguard the sites and practices they’ve put forward). An unsuccessful nomination, or a tradition that’s never nominated at all, receives none of this. The incentive structure shapes what the system preserves.

For example, in 2013 UNESCO inscribed “washoku, traditional dietary cultures of the Japanese, notably for the celebration of New Year” on the Intangible Cultural Heritage list. The Japanese government’s nomination framed washoku as an integrated social practice tied to seasonal ingredients, communal gathering, and an expressed spirit of respect for nature. The inscription brought global media attention, culinary tourism, and government-funded educational programs aimed at preserving the practice domestically. Japan had both the institutional infrastructure to prepare the nomination and the resources to support it through the multi-year evaluation process.

Meanwhile, food traditions of comparable depth and complexity in regions without that institutional infrastructure (like the fermentation practices of rural West Africa or the communal cooking traditions of Indigenous Amazonian communities) remain uncatalogued by the international system. They aren’t less significant, but they are less visible because the nomination process requires government funding, institutional expertise, and years of preparation that many countries can’t allocate. The geographic breakdown of the World Heritage List reflects this same dynamic.

What the numbers show

The World Heritage List’s geographic distribution makes this pattern legible at a global scale. As of 2025, the Europe and North America region (as UNESCO defines it, which includes Russia and the Caucasus states) accounts for close to 47% of all inscribed sites. Africa, with a comparable total land mass, accounts for under 9%.

UNESCO’s 1994 Global Strategy was explicitly designed to address this disparity, but peer-reviewed research analyzing the decades since its adoption has found that the imbalance has persisted and, by some measures, has grown. Part of the reason comes down to resources. The nomination process requires institutional capacity and dedicated funding, and a government that can afford to prepare a detailed, multi-year nomination dossier is more likely to get sites listed than one that can’t, regardless of what cultural material exists within its borders. UNESCO has acknowledged this disparity and has directed funding toward supporting under-represented regions in preparing nominations, but the list's composition still reflects decades of uneven access to the nomination pipeline.

The World Heritage List doesn’t claim to be a complete inventory of everything that should be preserved. But because it’s the most prominent global framework for cultural recognition, its composition shapes the broader picture of where significant heritage is recognized. When close to half of all inscribed sites are concentrated in one region, that concentration influences international tourism flows, academic research priorities, and public perception of which parts of the world hold cultural weight. The list becomes a reference point, and the gaps in the list become gaps in the global picture.

How categories shape what we see

How something gets labeled determines what happens to it next. When a sacred Indigenous site is classified as a “cultural landscape,” it’s evaluated under criteria that emphasize human interaction with the environment, and the protections that follow are designed to maintain that relationship. Should the same site be classified under natural criteria, the protections would then focus on ecological preservation, and the community’s cultural connection to the land may not factor into the management plan at all. The site is the same, but the category it’s assigned to determines which evaluation track it enters, which funding streams it can access, and which legal protections apply.

This dynamic extends well beyond UNESCO. A library’s classification system determines which subjects get shelved together and which get separated across different sections of the building, shaping how we browse and what we encounter along the way. A search engine’s indexing determines which results surface first, based on ranking criteria that reflect the priorities of the people who designed the algorithm. In each case, the organization feels neutral once it’s in place. The call number on the spine of a book doesn’t look like an argument, and a search result ranking doesn’t announce itself as a set of editorial choices. But every one of these systems was designed by people making decisions about what to prioritize and how to measure it.

Once a classification system is in place, it shapes what’s visible and what isn’t. The way information is organized is never a neutral reflection of what exists in the world. It’s a reflection of the choices made by the people who built the framework, and those choices determine what we find easily, what we have to dig for, and what we might never encounter at all.

A global catalog, open for browsing

UNESCO’s three heritage programs - the World Heritage List, the Intangible Cultural Heritage lists and the Memory of the World Register - represent the most ambitious attempt any institution has made to catalog humanity's cultural and natural inheritance in one place. No other system operates at this scale, across this many countries, covering sites, living practices, and documents simultaneously.

Before the World Heritage Convention, international cooperation on cultural preservation was limited and situational. The 1954 Hague Convention established protections for cultural property during armed conflict, and UNESCO had coordinated individual rescue campaigns like Abu Simbel, but there was no permanent, peacetime system for identifying and protecting heritage across borders. What a country chose to preserve in ordinary times was largely its own affair. But the World Heritage Convention changed that by establishing that certain places, practices, and records belong not just to the countries that house them, but to all of humanity. That shared premise is what makes international preservation funding, conservation expertise, and global responsibility possible: the recognition that losing a piece of cultural heritage anywhere diminishes the collective inheritance everywhere.

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

The method that follows starts with how to evaluate which sources to trust and how to capture what we find across platforms. From there, it moves into building a personal reference system organized around the categories that emergency management frameworks have used for decades. The final section provides a concrete action sequence, ordered from lowest friction to highest, for turning that reference into completed tasks.

Reader Question

“I don’t know how to describe what my feed looks like right now. It’s news alerts about wars, economic instability, and climate disasters, and then right next to all of that are Reels about building a root cellar, what to put in a 72-hour emergency bag, or how to grow your own food when the supply chain collapses. I can’t tell if I’m being paranoid for paying attention or naive for not taking it more seriously. I’ve started clicking on the preparedness content and now I’m getting more and more of it. I’ve spent hours reading and watching and I feel more scared and confused than prepared. I don’t know what’s real, I don’t know what to trust, and I have no idea where to start.”

- Amy M.

A note before we begin: Thank you for being here, whether you’re a free or paid subscriber. This post covers why emergency prep content feels so overwhelming and how to evaluate which sources to trust (available to everyone), then goes deeper into the tools and method for capturing, organizing, and acting on what we find across platforms, including a concrete action sequence ordered from lowest to highest friction (for paid subscribers).

Librarian Answer

Recommendation algorithms are built around one organizing principle: keeping us engaged long enough to see the next piece of content. Once we engage with any emergency prep content, the algorithm identifies that category as something that keeps us scrolling and serves more of it, creating an accelerating stream from creators who may have posted days or weeks apart but whose content arrives in a compressed rush that feels simultaneous. The impression that an extraordinary number of people are all urgently saying the same thing at the same time is partly true and partly an artifact of how the feed behaves once we’ve engaged with a topic. That level of concentrated delivery makes the volume itself feel like a warning, layered on top of information that’s already landing with heaviness.

Armed conflicts are disrupting food and material supply routes across multiple regions. Supply chains remain under strain from ongoing tariff shifts and geopolitical realignment, with industry analysts describing this volatility as a permanent condition rather than a temporary disruption. Climate-driven emergencies are arriving with increasing frequency. The creators covering water storage and go bags and emergency communication are responding to the same realities we’re all watching unfold, and the advice they’re giving is often critical and grounded in real experience.

So we end up in a particular kind of paralysis: we’re absorbing a huge amount of practical guidance from dozens of sources, while paradoxically become less clear on what to do, exactly, with all of it. A feed organized entirely around engagement can’t distinguish between a creator with fifteen years of tactical experience and one who discovered the topic two months ago, or between guidance relevant to our specific region and advice that has no relationship to our circumstances.

Who to Listen To

The feed can’t tell us which sources to trust. A creator with decades of hands-on experience and a creator who discovered the topic recently and is working through their own anxiety look identical on screen. Both can be confident, both can be specific, and both can have large followings. The difference between them is the basis for what they’re sharing, and a few questions help us tell them apart.

Where does their knowledge come from? A creator who has lived through extended power outages, who has grown and preserved food, and who has tested their emergency plans in real conditions speaks from a different foundation than someone who compiled their advice from other creators’ content. A scroll through their posting history usually reveals this: when did they start covering emergency prep? Do they reference their own experience, or are they primarily passing along what they’ve seen elsewhere? Practiced expertise tends to include the kind of specificity that only comes from having done the thing: which equipment holds up and which doesn’t, what advice sounds good in theory but fails in practice, and what they changed after their first real emergency.

Are they sharing from knowledge or from the same anxiety we’re feeling? Some creators are processing their own fear publicly and presenting information they've recently gathered as if it's already been tested and confirmed, when it hasn't. Others are sharing routines and systems they've refined over years. The second group tends to be calmer in delivery, more specific in recommendations, and more willing to say "This is the limit of what I know" when they've reached it. A creator who presents themselves as an authority across every area of emergency preparedness is likely overstating what they know, because experienced practitioners are usually very specific about where their expertise ends.

Do their recommendations show up across other experienced sources? When multiple creators with different backgrounds and no financial relationship to each other land on the same specific recommendation, that independent convergence functions the same way peer review does in academic research: it tells us the guidance has been arrived at from different starting points and is more likely to hold up. When a recommendation appears from a single source, or only from sources that are also selling the thing they’re recommending, we wait before acting on it.

Once we’ve identified creators whose knowledge is grounded in years of practice, those become our primary sources. We don’t need to follow hundreds of accounts; a small number of trusted, experienced voices gives us far more than an enormous, unvetted feed.

How to Stop Losing What We Find

Emergency prep content comes at us from everywhere: Instagram Reels, TikTok, YouTube, Substack newsletters, news articles, Reddit threads, blog posts. Each platform has its own storage mechanism, and each one keeps content in a way that’s difficult to search, compare, or return to later. An Instagram bookmark folder, a YouTube Watch Later list, a Substack reading list, a cluster of open browser tabs: these are all separate holding areas that can’t communicate with each other, and they all share the same fundamental problem. The content we saved feels like information we’ve kept, but it’s scattered across platforms in formats we can’t cross-reference, search by topic, or act from in any organized way.

The goal of everything that follows is to move from passively saving content inside platforms we don’t control to actively building a working collection of knowledge we own, can search, can update, and can reference when we need to make a specific decision or take a specific action. The process has two distinct moments, and separating them is what makes the whole system sustainable. The first is the save, which happens while we’re scrolling and needs to take no more than a single tap. The second is the processing, which happens later in its own dedicated time when we’re ready to work with what we’ve collected rather than consume more.

Before we continue: The tools covered in the rest of this piece, NotebookLM and Notion, belong to a broader category of information management tools designed to do something our brains and our browser tabs can’t: hold large amounts of information from different sources in a single, searchable place where we can actually work with it rather than just store it. These tools apply well beyond emergency prep, to any subject where we're dealing with more incoming information than we can organize on our own.

These tools are powerful but not always intuitive to set up. With that in mind, I’m exploring a series of video tutorials covering how to get started with them and how to apply them to any subject, designed as reference resources to work through at your own pace. I’d love to hear if this is something you’d be interested in - thanks for your input!

Read more

The concept gained mainstream traction in 2012, when technologist Clay Johnson published The Information Diet, a book that made the case for conscious, deliberate consumption of media. Johnson argued that just as the industrialization of food had created an epidemic of obesity by making cheap, addictive calories available at scale, the industrialization of information had created an epidemic of misinformation and overwhelm by making cheap, addictive content available at scale. The solution, in both cases, was personal responsibility: learn to recognize the junk and develop the habits needed to take control of what you consume.

The book arrived at a specific inflection point. In 2012, social media was rapidly becoming the primary way people encountered news, but the platforms were still relatively new and the scale of the problem was still emerging. Facebook would cross a billion users that same year. Twitter was becoming the default venue for breaking news. The smartphone had put an infinite scroll of content in everyone’s pocket. For many people, this was the first moment when the sheer volume of available information started to feel unmanageable, and the diet metaphor offered a framework that made the problem feel solvable. If the issue was consumption, then the solution was discipline. If the issue was too much input, then the answer was better filters.

That framing resonated deeply, and it’s continued to shape the conversation about media literacy and information overwhelm for more than a decade. Today the information diet shows up in corporate wellness programs and digital detox retreats, in parenting advice about screen time and in the broader language of self-improvement culture. The idea has become so embedded in how we talk about our relationship to information that it can be easy to forget it’s a metaphor at all, and that like all metaphors, it illuminates some things while obscuring others.

The three pillars of the information diet

The information diet isn’t a single set of instructions. It’s a collection of related practices that different writers and thinkers have organized under the same umbrella, and the specific advice varies depending on who’s offering it. But the core recommendations cluster around three consistent themes.

The first is reduction: consume less. This means spending less time scrolling through social media, limiting exposure to 24-hour news cycles, and being deliberate about how many newsletters and feeds and notifications we allow into our daily routines. The reasoning behind this recommendation is that the sheer volume of information we encounter creates a kind of cognitive overload, where we’re processing so much that we lose the ability to think clearly about any of it. Reducing the volume, the argument goes, creates space for deeper and more focused engagement with the information that remains.

The second is curation: consume better. This means actively choosing high-quality sources over low-quality ones, favoring long-form reporting over clickbait, seeking out writers and publications with track records of accuracy rather than defaulting to whatever the algorithm surfaces. Curation also means periodically auditing our feeds and subscriptions to make sure they still reflect our actual interests and needs rather than the accumulated drift of years of casual follows and one-click subscriptions.

The third is protection: set boundaries around our consumption. This means establishing screen-free times and turning off push notifications, creating physical and temporal spaces where information doesn’t reach us. The logic here is that constant connectivity creates a kind of ambient anxiety, and that deliberately disconnecting from the flow of information gives our brains the rest they need to process what we’ve already taken in.

All three of these practices are reasonable, and for many people they represent a meaningful improvement over consuming information passively and without any intentional structure at all. Reducing volume can create the mental space for deeper thinking, and choosing better sources improves the quality of the information we’re working with. Setting boundaries can ease the background stress that comes from feeling perpetually plugged in. These practices help, but the real issue is whether they’re sufficient on their own.

Where the diet metaphor breaks down

Where the framework runs into trouble is in the underlying assumption that ties all three pillars together: that information quality is primarily a problem of individual consumption, and that the solution is primarily a matter of personal discipline. That assumption shapes not just the advice itself, but the way we think about who’s responsible when our information environment fails us.

The amount of content produced globally every day now exceeds what any single person could process in a lifetime, and a growing share of that content is specifically engineered to hold attention rather than to inform. Algorithmic feeds (the systems that platforms like Facebook, Instagram, TikTok, and X use to decide what content appears in our timelines) select content based on engagement metrics like clicks, likes, shares, and time spent. Content that provokes strong emotional reactions gets surfaced more often than content that’s accurate but less attention-grabbing. Asking each of us to individually out-discipline a system designed by some of the best-funded engineering teams on the planet is like asking someone to bail out a flooding basement with a coffee mug. The effort is real, but the tool doesn’t match the scale of the problem.

There’s a deeper issue embedded in the metaphor, too. Framing our relationship to information as a “diet” positions the entire information environment as something to defend against. The strategies are oriented around what to consume less of and what to block and avoid. These defensive moves can help manage the feeling of being overwhelmed, and for people in acute information distress, they can provide meaningful relief. But they don’t address the underlying challenge: evaluating information well requires knowledge and skills that no one person holds across every subject they encounter. We can be as disciplined as we want about our consumption, and we’ll still run into claims we can’t evaluate and stories we can’t contextualize on our own. And that’s where a different question becomes necessary: if individual discipline isn’t sufficient, what would a shared approach to information quality look like? The answer turns out to be less novel than we might expect, because for most of history, shared approaches were the default.

The village we used to have

For most of human history, and for most of the history of mass communication, the work of navigating complex information was shared across institutions and communities rather than carried by individuals acting alone. Public libraries employed reference librarians, professionals whose entire function was to help people find trustworthy information and navigate questions they couldn’t answer on their own. A person didn’t need to independently evaluate every source in the library’s collection. They could walk up to a desk and describe what they were looking for, and a trained professional would guide them toward the most reliable material available on that subject. This wasn’t a luxury or a niche service, but rather the foundational model of how public information access was designed to work. Newsrooms served a similar intermediary function. Editors and fact-checkers verified reporting before publication, so that readers received information that had already been through a review process designed to catch errors and flag unsubstantiated claims.

These systems weren’t perfect. They had biases and blind spots of their own, and access to them was uneven across communities and demographics. But they operated on a shared assumption that’s easy to take for granted until it’s gone: the work of determining what’s reliable is too complex and too consequential to leave entirely to each person acting alone. These were collective systems built to carry a collective burden, and for decades they functioned as the invisible infrastructure beneath the information environment most people navigated daily.

The rise of the internet disrupted many of these intermediary structures and replaced them with something fundamentally different. Social media platforms didn’t just change where we get our information; they changed who does the work of evaluating it. The assessment labor that used to be performed by trained professionals at institutional scale has been pushed onto each of us as individuals, without the training or tools those professionals relied on. And the systems that replaced them weren’t designed to prioritize accuracy in the first place. They were designed to maximize the amount of time we spend on the platform. The shared infrastructure that once carried the weight of information evaluation has largely disappeared, but the need for it hasn’t. If anything, the need has grown.

What an information village looks like

An information village is a way of rebuilding that shared evaluative capacity at the community and personal scale. It’s the ecosystem of sources and relationships we rely on to stay informed. It includes the writers and publications we read, the people we talk to about what’s happening, the places we go when we need to verify a claim, and the habits we’ve built around consuming and evaluating information. When that ecosystem is working well, we feel oriented. We can find what we need, we encounter perspectives that challenge our thinking, and we have ways of checking whether something is reliable before we act on it or pass it along.

Think about the last time a piece of information stopped us in our tracks. Maybe a headline that seemed alarming but also seemed too perfectly crafted to be straightforward reporting. Maybe a statistic that showed up in a friend’s social media post, cited with enough specificity to feel authoritative but without enough context to tell whether it was meaningful. In that moment, what did we do? If the answer was “I just kind of sat with the uncertainty and moved on,” that’s not a personal failure. That’s a gap in the ecosystem. It means the network of sources and people around us didn’t have a clear pathway to the kind of help that would have made a difference in how we processed what we’d encountered.

The village model shifts the question from “How do I individually consume better?” to “Who can I think with?” Instead of treating every confusing or ambiguous piece of information as a test of our personal discernment, we treat it as a prompt to engage the network of people and sources we’ve built around ourselves. A healthcare worker who can’t evaluate a clinical study on her own posts the question in a professional group and gets a detailed methodological breakdown from a pharmacologist within an hour. A parent trying to figure out whether a school policy change is as alarming as a viral post makes it sound texts a friend who works in education and gets the context the post left out. When a major news event breaks and the initial reporting is still developing, the village provides someone we can think through the early reports with in real time, someone who can help sort out what's been confirmed from what's still speculation. In each of these situations, what makes the difference is the existence of a relationship or a shared space where asking for help is normal and where complementary expertise can flow toward the question that needs answering.

How to build an information village

The shift from diet to village is a conceptual reorientation, but it needs concrete infrastructure to function on a daily basis. The raw materials are often already scattered through our lives: a coworker who’s sharp on a topic we aren’t, or a friend who always asks the question that makes us reconsider our first reaction. The work of building a village starts with recognizing the evaluative relationships that already exist in our lives and then strengthening them, filling in the gaps, and giving the whole network enough structure that it functions reliably rather than accidentally.

Build a resource hub.

A resource hub is a single, centralized place where we collect sources that have earned our trust over time. The format matters less than the consistency: a bookmarks folder, a Notion database, a shared Google Doc, or a NotebookLM all work. When a journalist or publication or researcher proves itself reliable through repeated accuracy and rigorous sourcing, it goes into the hub. Over time, this becomes a personalized reference collection we can return to instead of starting every question from scratch with a search engine. The difference between a resource hub and a casual bookmarks bar is intentionality: every entry in the hub is there because it’s been vetted through repeated use, not because we clicked “save” once only to have the bookmark get buried over time.

Create a commons.

A commons is any shared space where information and recommendations flow between people rather than through an algorithm. This can be as simple as a group chat where friends share what they’re reading, or a regular conversation with someone whose taste and judgment we’ve come to trust over time. The value of a commons is that the filter is human judgment rather than platform incentives. Algorithmic feeds are designed to maximize engagement, and engagement-optimized content doesn’t reliably surface accurate or useful information. A commons routes around that dynamic by relying on people who are selecting for quality rather than attention capture.

That said, human judgment carries its own biases. The people in our immediate circles tend to share our worldview and our blind spots. A commons where everyone reads the same sources and arrives at the same conclusions is just an echo chamber with a friendlier name. The counterweight to this isn’t seeking out people who disagree with us on everything — that’s diversity for its own sake, and it can introduce more confusion than clarity. What strengthens a commons is including people who share a commitment to evidence-based reasoning but who bring different professional backgrounds, different reading habits, and different areas of expertise. A nurse and an engineer and a journalist will read the same news story and notice different things. That kind of complementary perspective is what keeps a commons from collapsing into groupthink.

Invest in relationships where candor is normal.

The most valuable people in an information village are the ones who will tell us when they think we’re wrong, and who we can tell the same without the relationship suffering for it. This kind of trust takes time to build. It requires a mutual understanding that changing our minds in response to better evidence is a sign of good thinking rather than inconsistency. It also means building the kind of rapport where someone can say “I don’t know enough about this to have a take yet” without feeling like they’ve revealed a weakness, because that kind of admission is often the starting point for the best collaborative thinking we’ll do.

Share our own expertise.

An information village is reciprocal. Whatever our background or professional experience, there’s likely an area where we bring something others in our network don’t have. That contribution doesn’t have to be domain expertise, either. Maybe we’re the person who reads past the headline to the underlying source material. Maybe we’re the person who slows the group down when everyone is rushing to the same conclusion, the one who asks “Where did this statistic come from?” or “What are we assuming here?” Both deep knowledge and the habit of careful questioning are forms of contribution that make the village function, and the village works because its members give as well as receive.

Audit and maintain.

Our information needs change, and our villages should change with them. A periodic audit means looking at the current state of our information ecosystem and checking for two kinds of problems: gaps and decay.

Gaps are the topics we care about or make decisions around where we don’t have a reliable source or a trusted person informing us. These are the areas where we’re operating on assumptions or outdated information without realizing it. Decay is what happens when sources decline in quality without us noticing. A publication that was rigorous two years ago may have changed ownership or lost key writers or shifted its editorial direction in ways that no longer align with our needs. A relationship that used to be a source of sharp, candid exchange may have gone dormant. Decay is subtle because it happens gradually, and the sources that have degraded are often the ones we stop actively evaluating because we already decided to trust them at some earlier point.

Beyond gaps and decay, it's also useful to look for passivity: the accumulation of subscriptions and follows and feeds we signed up for in a different context or mindset but never revisited or cleared out. An audit helps distinguish between the information we’re intentionally relying on and the information that’s just lingering from an older version of our attention. The goal is to make sure the ecosystem we’ve built still reflects our current needs and current standards rather than the ones we had years ago.

Building the village is one half of the equation. The other half is understanding how village practices relate to the diet practices that many of us already have in place, because the two models aren’t in competition — they do complementary work instead.

The relationship between the two models

The information diet and the information village work at different levels of the same problem. The diet sharpens individual capacity, while the village extends that capacity into territory no single person can cover alone. The most effective information practice draws on both, and understanding where each one is strongest helps clarify how they complement each other.

At the individual level, the diet model asks: what am I consuming, and is it serving me well? These questions build the foundational skills of media awareness and source evaluation that make us more discerning readers and more careful sharers of information. Someone who has developed strong individual information habits brings sharper judgment to every conversation and every recommendation they make within a village.

The village model operates at the level of relationships and shared infrastructure. It asks: who do I think with, and what systems have I built for staying informed that don’t depend entirely on my own capacity? The diet model doesn’t address these questions, because its unit of analysis is the individual consumer. The village model recognizes that every person, no matter how disciplined, will encounter questions that exceed their own expertise, and that the quality of their information environment depends on the quality of the relationships and structures they’ve built around themselves.

The diet model is strongest in sharpening personal discernment, helping us become more aware of our consumption patterns and more intentional about where we spend our attention. Where it runs into limits is in situations that require knowledge or perspective beyond what any one of us can hold on our own, like a breaking news event that crosses multiple domains of expertise or a scientific study whose methods require specialized training to evaluate.

The village model is strongest in extending our evaluative range beyond our own knowledge, turning the question “Is this reliable?” from a solo research project into a collaborative one where different people contribute different lenses. Where it runs into limits is when the relationships within the village lack candor or intellectual diversity, which can turn a village into a closed loop that reinforces existing beliefs rather than testing them.

Each model is strongest when paired with the other. The diet without the village produces isolated individuals trying to do institutional-scale evaluative work on their own, consuming carefully but lacking access to the perspectives and knowledge that would make their evaluations more complete. The village without the diet produces a group of people sharing and amplifying material that none of them have individually scrutinized, creating the appearance of collective wisdom without the foundation of individual rigor beneath it.

The two models differ in structure, but they share a common assumption: that the quality of our information environment is something we can actively shape rather than passively endure. The diet model shapes that environment by removing low-quality inputs and managing the volume of what comes in. The village model shapes it by building the relationships and shared systems that make evaluation a collaborative process rather than a solo one. Moving from diet to village means keeping the protective habits the diet provides while investing new energy in the relational infrastructure the village requires.

Moving from diet to village

The transition is less about abandoning old habits and more about building new infrastructure around them. The diet skills remain, and they become the foundation for a broader set of practices that are relational rather than purely individual. If the diet is about what we consume, the village is about what we build, and the shift between them is a shift in where we invest our energy. That shift tends to follow a few consistent patterns:

• From solo filtering to shared filtering. The diet model asks us to evaluate every piece of information on our own. The first step toward a village is identifying the areas where we consistently feel uncertain and then finding a person or source we trust in that specific domain. This doesn’t require building a new relationship from scratch. It might mean paying closer attention to who in our existing network has demonstrated strong judgment on a particular topic, and then making a deliberate choice to consult them when something in that area comes across our feed that we can’t evaluate alone.

• From passive consumption to active contribution. In the diet model, we’re consumers. In the village model, we’re participants. That transition means recognizing that we have expertise and perspective that other people in our network don’t, and that sharing what we know when it’s relevant is part of maintaining the ecosystem. If we read a primary source and notice that the headline circulating about it is misleading, saying so in our group chat or sending a note to a friend isn’t just helpful to them; it’s how the village gets built, one small act of shared evaluation at a time.

• From algorithmic discovery to curated commons. The diet model tells us to resist the algorithm. The village model gives us something to put in its place. Creating a commons, which can be as simple as a group chat where trusted people share what they’re reading or a shared document where we collect reliable sources, means building an alternative channel where recommendations are filtered through human judgment rather than engagement optimization. The more people contribute to a shared commons, the broader and more diverse the recommendations become, and the less dependent any one person is on algorithmic discovery for finding new material.

• From individual boundaries to collective standards. The diet model asks us to set personal limits on our consumption. The village model asks us to develop shared expectations with the people we think with. This can be as simple as a mutual agreement that we’ll tell each other when we think the other is wrong, or a norm within a group chat that we’ll flag the source when we share a claim rather than just sharing the claim itself. These collective standards create a kind of distributed quality control that no individual filter can replicate.

• From periodic detox to ongoing maintenance. The diet model tends to frame information management as a cycle of overload and retreat: consume too much, feel overwhelmed, disconnect, repeat. The village model replaces this cycle with ongoing, low-level maintenance of the ecosystem itself. That means periodically checking our resource hub for sources that have drifted or degraded, noticing when a relationship that used to provide sharp thinking has gone dormant, and clearing out the passive subscriptions and follows that no longer serve us. The goal isn’t to achieve a perfect state and hold it, but rather keeping the ecosystem responsive to our actual needs as they evolve.

Each of these shifts is small enough to start with a single conversation or a single shared link. But taken together, they represent a fundamentally different orientation toward the problem of information quality, one that treats staying well-informed as something we do with other people rather than something we do by ourselves. And that reorientation has implications that go beyond any individual practice or habit.

What we’re building when we build a village

The reason this reorientation matters is that the dominant advice for navigating a complex information environment — tighten individual filters! set boundaries around consumption! — has been the prevailing advice for over a decade. This advice has produced some good habits and certainly some useful tools. But it’s also produced an implicit expectation that every person should be capable of navigating the full complexity of the modern information landscape armed with nothing but our own judgment and a list of approved sources.

The institutions that used to share the weight of this navigational work, the libraries and newsrooms that filtered and verified and contextualized information before it reached us, have weakened considerably over the past two decades. And the platforms that filled the space they left behind were designed to optimize for engagement rather than accuracy. So the result is that we've been left doing institutional-scale evaluative work as individuals, without the infrastructure those institutions once provided.

The skills the diet model builds, as useful as they are, will always have a ceiling when practiced in isolation. We can sharpen our personal discernment to a fine edge and still encounter a study that’s difficult to evaluate or a news event we can’t contextualize or a claim that sits in a domain where we have no footing. These moments aren’t failures of discipline. They’re the natural boundaries of individual knowledge, and a village is what connects us to people and sources on the other side of those boundaries.

The raw materials for building a village are already present in most of our lives. We already have people whose thinking we trust on specific subjects, and we already have sources that have proven reliable over time. The shift is in connecting these relationships and sources into a deliberate ecosystem, one we actively maintain and can draw on consistently.

The dominant conversation about information quality for the past decade has centered on a single question: “What should I consume?” That question taught us to curate our feeds, set boundaries on our consumption, and pay closer attention to the sources we rely on. But it left us doing the work of evaluation alone. The question that takes us further is: “What should we build together?”

The free essays are the foundation. The paid tier is the applied toolkit: biweekly AI briefings, monthly subscriber-driven research, and quarterly guides that give you real skills you can use immediately, plus a growing framework library (and classes coming soon). Upgrade to paid if you want the full Card Catalog. Thank you for being here!

Subscribe now

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Quick note: starting this week, Card Catalog is moving to one article per week. I’m building new things behind the scenes, and I’m so excited about what’s coming down the line! The Librarian Hotline, AI Briefing, and Research Packets continue on their regular schedule for paid members, and I’ll be in your Notes feed throughout the week.

Share Card Catalog

This week:

• Sony built a robotic arm that defeats elite table tennis players not by overpowering them, but by outthinking them at 20 milliseconds per decision.

• More than a dozen major U.S. law firms warned clients in April that anything they type into a consumer AI chatbot can be demanded as evidence in court.

• The United Nations launched its most ambitious push for global AI standards, testing whether the rest of the world gets a voice in rules that have so far been shaped by a small number of nations.

Sony built an AI robot that competes with and defeats elite table tennis players

What Happened

Sony AI spent five years building a robotic arm called Ace that plays competitive table tennis against elite human athletes. The research was published in Nature, one of the world's most prestigious scientific journals, in April 2026. In officially refereed matches, Ace competed against five elite and two professional players on a regulation court under International Table Tennis Federation rules.

Ace uses nine cameras positioned around the court and reads the ball’s logo to measure spin. Its robotic arm, built with eight degrees of freedom, processes visual input and executes shot decisions in 20.2 milliseconds, compared to roughly 230 milliseconds for elite human players. The system learned entirely through reinforcement learning, a method where the AI improves by accumulating experience rather than following pre-programmed instructions.

The team deliberately constrained the robot’s speed and reach to keep its physical performance comparable to a human athlete who trains at least 20 hours a week. Sony AI president Michael Spranger said the point was not to overpower opponents but to win through decision-making and tactics. After peer review, Sony continued refining the system, and by late last year Ace had defeated all but one of four high-skill opponents in additional testing. Kinjiro Nakamura, an Olympic athlete who competed at Barcelona, watched Ace execute a shot during that testing and said he had not believed such a shot was possible.

Why It Matters

AI has beaten humans at chess and Go, but all of those victories happened on screens or boards. The physical world is a different problem entirely. Objects curve and bounce unpredictably, opponents adjust in real time, and getting a machine to perceive and respond within the millisecond windows of competitive athletics has been one of the longest-standing open challenges in the robotics field.

Sony’s researchers have said the same perception and control technology could apply to manufacturing and service robotics, or any environment that requires fast, adaptive physical responses. A system that operates at superhuman speed in unpredictable, real-world conditions is the kind of capability that industrial and defense applications tend to adopt quickly.

What It Means for Us

Most of us interact with AI through a screen. Ace represents a different frontier: AI that operates in physical space in ways that are both responsive and interactive. That shifts the conversation about which kinds of work AI can participate in: we’re moving from AI that writes and searches to AI that moves and competes. The robotics applications that follow from this research will eventually appear in warehouses and operating rooms, and the question of how we share physical tasks with machines is no longer theoretical.

Bottom Line: If AI systems can now outperform trained athletes in fast, unpredictable athletic competition, the range of physical tasks they can take on is about to expand. The industries built on speed and real-time adaptation will feel the effects first. How quickly that expansion happens, and who controls the systems driving it, will shape the next phase of AI’s role in daily life.

Major law firms are warning clients: anything you type into an AI chatbot can be used against you in court

Read more

The Daraya library is one of the most thoroughly documented underground libraries in recent history, but it’s one of many. Under regimes that share nothing else in common, small groups of readers living under restriction have built collections of their own, stocked with the material their governments were working to remove from circulation. The pattern has recurred across continents and centuries in forms that vary wildly, and the history of these libraries is also part of the history of how authoritarian power has been resisted from within.

What an authoritarian regime is, and what it does to information

An authoritarian regime is a government that has freed itself from meaningful accountability to the population it governs. The fully developed version is recognizable: elections that can't change the outcome, opposition leaders in jail or in exile, press owned or controlled, courts that rule the way the people in power want them to, and information managed from the top through both censorship and propaganda. But most of what we call authoritarianism in practice doesn’t start at the fully developed version. It moves toward this through a gradient: elections tilted before they're outright rigged, opposition harassed before it's criminalized, press captured by loyal ownership before it's nationalized, judicial appointments ideologically filtered before judges are purged. Each step expands what the regime can do, and each step requires corresponding control over what the population is able to know and say about what's happening.

Every government, authoritarian or otherwise, needs some baseline of cooperation from the population, because no regime has the resources to coerce every citizen at every moment. Democracies secure that cooperation through accountability, which lets citizens remove leaders who fail them and gives the governed a reason to accept outcomes they didn't personally choose. Regimes moving in the authoritarian direction can't rely on that mechanism, so they use other ways of keeping the population in line. The most powerful and lowest-cost of those alternatives is the control of information, because a population that can't verify what it's being told tends to do what it's told, or at least fails to organize the kind of resistance a regime needs to worry about.

Authoritarian information control operates on several fronts at once:

Restriction

The regime narrows what the population can access:

• Banning books and periodicals

• Controlling printing and broadcasting equipment

• Censoring films and monitoring religious sermons

• Limiting internet access

• Removing material from public collections

Distortion

The regime floods the remaining channels with propaganda, manufactured news, doctored images, and official narratives that contradict what citizens are seeing with their own eyes. George Orwell described this precisely in 1984: “The Party told you to reject the evidence of your eyes and ears. It was their final, most essential command.”

Surveillance

The regime monitors conversations, correspondence, reading habits, and online activity, so that citizens learn to censor themselves before the state has to.

The long narrowing of thought

Each generation that grows up inside a controlled information environment inherits a smaller vocabulary for describing the situation, a thinner knowledge of how things have been organized elsewhere, and fewer conceptual tools for imagining alternatives. This is the longest-running front, and the hardest to reverse once it has taken hold.

Information control reaches every medium through which one mind reaches another. Written work is the part of that effort most visible in retrospect, but the same control runs through speech, images, film, music, public ritual, and now the platforms where most of modern communication happens. The written record, however, draws the regime’s particular attention because it’s durable. A book can be hidden, buried, passed from reader to reader, and retyped letter by letter, and can carry its contents intact across decades. A conversation evaporates the moment it ends, or digital communication disappears when the platform it was on ceases operation. But written work is how a population leaves messages for the future in which the regime is gone.

Populations under these restrictions don’t stop reading of course, and they often read more as the supply tightens. Reading becomes the primary way people check the official account against independent sources and keep their own perception intact, which is the first defense against the Orwellian demand that they deny what they’ve seen. Reading also connects each reader to a larger community of people who have held similar thoughts, which reduces the isolation authoritarian regimes depend on to make dissent feel like private madness. Reading also preserves the vocabulary and conceptual range a population needs to describe its own situation accurately, which is what the regime’s long-term narrowing of thought is designed to erode. And reading is what keeps the written record alive across time. A document survives only when enough readers copy it and pass it along to the next generation, which is how the accounts written under authoritarian conditions reach the generations that need them.

Readers under restriction tend to look for suppressed political analysis and histories of what the regime has tried to erase, for religious texts the regime has banned or rewritten, for foreign literature that shows life organized differently from how the regime insists it must be, and for ordinary novels and poetry that affirm the inner life of a private person (which is what authoritarian systems are most determined to deny). For example, one of the books that became central to readers in the Daraya basement library was Stephen Covey’s The 7 Habits of Highly Effective People, a self-help title about personal agency, in a siege whose defining condition was that personal agency had been taken away. Readers reach for whatever the regime is working hardest to take from them, which means an underground library’s collection often functions as an inverse map of what the regime fears most.

What an underground library is

An underground library is an organized collection of books and other written material maintained outside the sanction of a regime that has restricted access to such material. The “underground” part refers to the relationship with state authority, not to physical location. Some underground libraries have hidden themselves in basements, sealed cellars, and buried containers. Others have operated openly in private homes or community spaces as deliberate acts of defiance, accepting state harassment as the price of visibility. Still others have existed primarily as networks of circulation, with no fixed site at all. What unites them is the underlying refusal: the state has decided that certain books should not be read, or that certain readers should not have access, and the library exists because readers have decided otherwise.

The form an underground library takes depends on what the state is restricting and how far the state is willing to go to enforce that restriction. Where the state controls printing equipment, readers reproduce texts by typewriter or by hand. Where the state monitors postal mail, readers move material in person between trusted contacts. Where the state raids private homes, readers bury or hide collections beneath floorboards, inside walls, or in the ground. Where the state has destroyed institutional libraries, readers excavate the wreckage for what survived. The specific tactics vary; the underlying work does not. The underground library selects what to keep, catalogues it so that it can be found again, circulates it among readers who want it, and gives access to the community the collection is meant to serve. These are the four functions any library performs, and underground libraries perform them under conditions that should make the work impossible.

Taken together, the historical record supports a simpler definition than the one most people carry around. A library is not fundamentally a building or an institution; a library is what forms when a group of readers decides to keep particular books available to each other and organizes itself well enough to do so reliably. Buildings, professional staff, funding, and state sanction are scaffolding that makes the work easier. But scaffolding can fall away, and a library continues to exist as long as the readers continue.

Three case studies

Underground libraries have appeared in dozens of documented cases across the last two centuries, and the pattern likely extends well beyond what any historical record has preserved. The ones we can see today are the ones that produced enough surviving material to leave a trace legible to historians, whether through rescued archives, later memoirs, testimony from survivors, or court records from trials of the people involved. Even the visible ones span a strikingly wide range of conditions, as a partial list suggests:

• The Vilna Ghetto’s underground reading rooms during the Holocaust

• The Polish flying universities that operated in Russian-partitioned Warsaw in the late nineteenth century and were revived under communist rule in the late nineteen-seventies

• The secret schools that operated for African Americans during slavery and Reconstruction

• The women’s reading group Azar Nafisi held in her Tehran living room under the Islamic Republic from 1995 to 1997

• The Polish underground press that flourished during the Solidarity era

The range of these cases points to how broadly the same basic structure reassembles itself under pressure. Three of the most thoroughly documented instances come from different parts of the last century and illustrate how varied the conditions can be while the underlying library work remains recognizable.

Warsaw, 1940–1943. Inside the Warsaw Ghetto, after the Nazis sealed roughly four hundred thousand Jewish residents behind its walls, a historian named Emanuel Ringelblum organized a secret archive to document what the Jewish population of Warsaw was living through. His team, eventually several dozen contributors drawn from across the ghetto, gathered essays, diaries, letters, drawings, photographs, ration cards, wall posters, and original sociological analyses of ghetto economics and internal politics. They interviewed refugees who had fled occupied towns elsewhere in Poland and transcribed their testimony. Understanding that most of them would not survive to tell the story themselves, they buried the archive in metal boxes and milk cans beneath ghetto buildings, intending it to be found after the war. Ringelblum was captured and executed by the Germans in March 1944, and most of his contributors were killed during or after the ghetto uprising. Two of the three buried caches were recovered, in 1946 and 1950, and the recovered material now forms the largest body of Holocaust documentation written by the victims themselves.

The Soviet Union, 1950s–1980s. Because the Soviet state kept tight control over printing presses, photocopiers, and any other equipment capable of reproducing text at scale, readers turned to the one piece of equipment the state hadn’t fully restricted: the personal typewriter. A dissident writer would type a manuscript on layers of tissue-thin paper separated by carbon sheets, producing four to six copies per run. Each copy moved through trusted hands, was read within a few days, and was either passed onward or retyped by the reader to extend the chain further. The practice acquired a name, samizdat, from the Russian for self-published. Over several decades it put banned works by Pasternak, Solzhenitsyn, Bulgakov, Akhmatova, and many others into the hands of Soviet readers who otherwise would have had no legal way to reach them. Penalties for possession or distribution were severe, with thousands of people convicted under anti-Soviet agitation and propaganda statutes over the postwar decades. The practice spread through the Eastern Bloc under related names and continued operating until the Soviet system collapsed in 1991.

Daraya, 2013–2016. Much of what is known about the Daraya basement library comes from two book-length accounts by the journalists Delphine Minoui and Mike Thomson, both based on interviews with the people involved. The library ran for nearly three years inside the besieged Damascus suburb, growing from a few hundred books to more than fourteen thousand. The collection was catalogued by hand and lent out with due dates. The basement functioned as both a reading room and a study space, and some users supplemented the collection with books they downloaded onto their phones during the rare windows when mobile networks worked. When the siege ended in August 2016 through a negotiated evacuation, the library’s founders and users were among the thousands of civilians and fighters bused north to Idlib. The library was left behind and ransacked by government forces shortly after the evacuation. Some of its founders had died during the years of siege, but the ones who survived carried the story out.​​​​​​​​​​​​​​​​

The same forces today

Underground libraries are still being built. In the United States, PEN America has documented nearly twenty-three thousand instances of school book bans since 2021, concentrated largely in Florida and Texas, and reports from affected districts describe informal lending networks forming among parents and students to keep removed titles in circulation. In Ukraine, librarians have evacuated and hidden collections during the Russian invasion. In Afghanistan, underground schools have continued teaching girls since the Taliban banned their education in 2021. Each of these cases traces the same arc the earlier ones did: a restriction is imposed, and a group of readers decides the loss is not acceptable and finds a way to keep access available.

The same four kinds of information control that authoritarian regimes apply deliberately also occur in democracies, produced by different mechanisms but with overlapping effects. Restriction in a democracy is rarely outright prohibition; it shows up as the steady removal of titles from school and library collections, licensing and liability pressure on publishers, and content removals across platforms. Distortion in a democracy isn’t state propaganda; it’s the combination of algorithmically amplified content and coordinated disinformation campaigns that erodes the shared factual ground a news ecosystem requires to function. Surveillance in a democracy is conducted mostly by private companies rather than the state, but the data collected on reading habits and search histories changes what people are willing to look up. And the long-term narrowing of what people find it possible to think shows up as the consolidation of information access through a few corporate platforms, which decides, increasingly, what most people encounter at all.

These pressures differ from outright authoritarianism in scale and mechanism, but they still affect what most people end up reading. Four practices drawn from underground libraries give readers ways to keep access to books and ideas stable if information channels should narrow. Each takes on a different angle of the narrowing, and together they form the basis of a reading life that holds up under institutional pressure.

Keep physical copies of the books that matter most

Digital access depends on licenses and platforms you don’t control. Those licenses can be revoked or rewritten without warning, and platforms can remove or narrow what they make available at any time. A book on your shelf is a copy that doesn’t disappear when a license expires or a platform decides to take it down.

• Build a personal library of titles that have shaped how you think

• Prioritize physical copies of books that appear on banned-book lists

• Keep at least one copy of any book you would want to lend or reread

• Buy used copies of out-of-print or hard-to-find titles before they disappear from circulation

Track what’s being removed from public collections

Knowing what titles are being pulled from institutional circulation is the first condition for keeping those titles readable elsewhere. Without that visibility, removals happen out of sight and accumulate without resistance.

• Look up which books are being challenged or removed in your country, region, or local school district

• Follow local reporting on library funding decisions and any government oversight of what libraries can stock

• Pay attention to which authors and topics show up most often on removal lists, since the patterns tell you what’s being targeted and why

• Subscribe to newsletters from organizations that track book censorship in your country or internationally, such as PEN International, IFLA’s FAIFE committee on freedom of access to information, Index on Censorship, or your national library association

Move books between people, not just own them

A shelf only serves whoever owns the shelf, but a lending habit serves an entire network of readers. The underground libraries did their work through circulation, and informal lending today does the same work at much lower stakes and with much higher cumulative reach.

• Lend books deliberately rather than waiting to be asked

• Send a copy of a book to a friend whose recent conversation made you think of it

• Join or start a reading group built around banned titles and works outside mainstream algorithmic circulation

• Buy second copies of meaningful books to give away

Diversify the sources of what you read

When most of what a person encounters arrives through algorithmic feeds run by a few companies, the supply is already concentrated in ways that make the reader vulnerable to whatever narrowing those companies decide to enact.

• Subscribe directly to writers and publications you trust

• Buy from independent publishers and bookstores when possible

• Visit physical libraries and bookstores to see what’s outside the algorithmic feed

• Read across formats: long-form journalism, books, archived material, primary sources

These practices share an underlying premise: that a reader’s access to books and ideas is fragile, and that the fragility is not visible from the inside until something has already been lost. Access tends to disappear in small increments that go unnoticed until a reader goes looking for something and finds it gone. By the time the loss registers, the work of restoration is far harder than the work of preserving it would have been.

Underground libraries are how reading survives when the state decides what cannot be read. Every book that moves between readers outside the state’s reach is a failure of the regime to control what its population can think. Enough of those failures across a society become something authoritarian power has never been able to fully suppress: an empowered citizenry. An authoritarian regime needs to control what its subjects are able to believe, but that control becomes impossible once enough people are reading past the limits the regime has drawn. These libraries have always been one of the ways authoritarian power gets weakened from within, and the work of sustaining them has always been open to any reader willing to take part.

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

Every information shift demands a new literacy. Twentieth-century kids learned to tell the news from the commercials, and to recognize a reporter’s byline as a different kind of authority than a pundit’s opinion. The internet brought a harder task, because anyone could publish anything, and a slick-looking website wasn’t automatically a trustworthy one. Social media layered an invisible filter on top, since the content reaching a kid had been sorted and shaped by systems optimizing for engagement long before anyone chose what to look at. AI is the latest turn of this arc, and it collapses the task further: a chatbot answer arrives confident and without citations, potentially with errors nested inside well-formed sentences and no visible trace of who produced it or how. So for a kid without evaluation skills, a news article and a viral post can land with roughly the same weight, and the cost of failing to tell them apart can have significant consequences as they get older.

Kids who grow up practicing evaluation skills enter adulthood able to meet whatever new claim shows up in front of them and hold it against a working understanding of how information gets made and whose interests it serves. They can disagree with the people around them because they’ve traced arguments back to their origins, and they’re better equipped to update their thinking when the evidence calls for it. What they then carry into adulthood is the ability to form their own opinions rather than inherit someone else’s. That ability shapes everything downstream of it, from the choices they make about their health and money to the votes they cast and the people they trust.

What we mean by information literacy

Library science has been developing evaluation skills for well over a century, and the field has a formal framework for teaching them. In 2016, the Association of College and Research Libraries published the Framework for Information Literacy for Higher Education. The framework replaced an earlier set of teaching standards from 2000 that had been organized around a checklist approach, where students learned to locate sources and evaluate them against fixed criteria.

The new framework moved toward a set of “threshold concepts.” A threshold concept is the kind of idea that reorganizes how someone thinks about a whole field once they grasp it. Learning about compound interest, for instance, permanently shifts how a person reads any decision involving money over time, in a way that wouldn’t have been available before the concept landed. The framework identified six such concepts in information literacy and called them “frames,” each describing a dimension of how information functions in the world. A kid who’s been taught the frames doesn’t just know the procedural steps of evaluating a source; she sees information differently, which is what carries her across whatever new tools and contexts she’ll meet over a lifetime.

The framework was written for college students and their librarians. But the frames it describes apply just as well to the information environments kids are navigating at home and in K-12 classrooms. One librarian practice sits upstream of the frames, shaping every information interaction that follows it: the reference interview.

The reference interview

At library reference desks, the question someone asks out loud is almost never the question they really need answered. A student might walk up and ask for “something about the French Revolution” when the material they need is something more specific, like a study of women’s political organizing in late-eighteenth-century Paris. Answering the literal question would send them in the wrong direction, so librarians developed a short protocol for surfacing the real question first. That protocol is called the reference interview.

The interview uses a handful of specific techniques:

• Open-ended follow-ups that invite the patron to say more about the underlying project

• Questions about what the patron already knows, and what they plan to do with the answer once they have it

• Restating the question back to check that it’s been heard correctly

A competent reference interview can turn a vague question into a sharp one, and that sharper question is what makes the rest of the research succeed. Without it, hours can be lost following a poorly-framed question down paths that were never going to answer what the patron actually needed to know.

The same practice transfers to how kids interact with any information source, including school research and AI chatbots. Teaching kids to interview their own questions before consulting any source is one of the most useful habits information literacy can build. A kid who learns this practice young carries it into every research task they’ll do, and the small early work of clarifying a question saves substantial wasted effort on vague or misguided answers later.

In the home: This move matters most right before your kid types into a chatbot or search engine, since these tools produce confident-sounding answers to vague questions just as readily as to sharp ones. Next time you’re about to look something up together, pause before either of you starts the search. If your kid asks “what’s the best video game console,” try responding with “best for what kind of games?” The question that comes back is usually one with a more useful answer. With a younger child wondering why sharks are dangerous, asking “what are you trying to figure out about them?” can surface what they were really after. Doing this with your kid gives them a model they can draw on later, when they’re consulting a chatbot or search bar without anyone next to them.

In the classroom: Before any research assignment, pair students up and have them interview each other’s topics, with one student playing librarian and the other playing researcher. A short exchange of this kind reliably turns a vague prompt into better research questions. The habit transfers to how students interview their own questions when they’re working alone, which is where the exercise has its biggest effect across the course of a research project.

The six frames

The reference interview clarifies what’s actually being asked before any source gets consulted. The six frames that follow do the next layer of work, which is evaluating what those sources contain once the question itself is clear. Each frame names a different dimension of how information functions in the world, and each one comes with its own way of reading sources that holds up across formats, platforms, and tools.

1. Not every expert is an expert on everything.

Framework name: Authority Is Constructed and Contextual

Authority isn’t a permanent quality a person has. Specific communities grant authority for specific claims, and that authority rarely travels smoothly across domains. A cardiologist has real authority on heart disease, but that authority doesn't extend to questions about mental health treatment, even though she holds a medical degree. A celebrity endorsing a wellness product brings name recognition to the product, but name recognition isn’t the same as knowing whether the product actually works.

This frame replaces the flat question of whether a source is trustworthy with the sharper question of what a source is trustworthy on. AI output complicates the picture further, because a chatbot answer carries the tonal texture of expertise without coming from any particular expert. Kids who learn this frame young develop the habit of asking what is this person trained in, and does this claim fall inside that area?

In the home: When your kid says “my teacher said...” or “the doctor said...” take the opening to ask a light contrast question. You can say “Does your dentist know a lot about teeth? What about building rockets?” The contrast makes the point without lecturing: expertise is specific to the area someone trained in. Over time, your kid can begin to ask on their own whether a source is speaking from their area of training or reaching beyond it.

In the classroom: Give students a short piece of writing where a credentialed person makes several different kinds of claims. An op-ed by a famous doctor on a political issue works well, as does a business executive writing about a scientific topic. Have students mark which claims fall inside the writer’s training and which reach outside it. The exercise demonstrates that credentials don’t transfer automatically across domains, and the question of where someone’s authority ends becomes one students can ask of any source going forward.

2. How information gets made shapes what it can tell us.

Framework name: Information Creation as a Process

A peer-reviewed paper and an AI answer can both claim to be sources of knowledge, but they were produced by radically different processes; that process shapes what kind of knowledge each one can hold. A peer-reviewed paper takes months or years to produce, and it passes through multiple rounds of expert review before it gets published. A chatbot answer takes seconds to generate, with no human editor involved at the moment of writing. Both arrive at a reader looking like information, but only one was shaped by a process designed to catch errors.

This frame teaches kids to ask how a piece of information was produced before deciding how much to trust it. A peer-reviewed article has been through checks that catch errors, so a kid can lean on it more heavily. A chatbot answer or a social media post has been through no checks at all, so anything that matters in it should be verified against a source that did go through verification checks. The habit we’re learning is how much to trust a source in proportion to how carefully it was made.

In the home: When your kid quotes something to you, take the moment to ask where they got it. “Did you read that in a book, hear it in class, see it on YouTube, or get it from AI?” The answer decides how much work the claim still needs. A fact from a textbook has been through some kind of review, which catches basic errors even though it doesn’t catch things like motivated omissions or framing choices. A chatbot answer hasn’t been through any review at all, so anything important needs a second verifiable source before the family treats it as true.

In the classroom: Pick a topic students are already researching, and have the class find two kinds of sources on it: an AI answer and an encyclopedia entry or textbook chapter on the same question. Ask students to infer what they can about how each source was made, based on the kind of source it is. From there, they can calibrate their trust: lean on the more carefully made source, and verify anything important from the less carefully made one.

3. Behind every piece of information, somebody wants something.

Framework name: Information Has Value

Anything created for an audience came into existence for a reason, and that reason shapes the content in ways that aren’t always obvious from the surface. Advertisements are built around the goal of converting viewers into customers, which determines the kinds of claims they make and the kinds of evidence they offer. Political messaging follows the same logic for a different end, shaping content around the goal of moving an audience toward a particular position. AI has added a new dimension of this issue, since many of the major models were trained on copyrighted material without compensation or consent, and what those models now produce serves the commercial interests of the companies running them.

The frame teaches kids to ask what the maker was after when they made the piece. The answer changes what shows up in the final content and what gets left out, which means noticing the maker’s motivation is where most of the evaluation work happens. Kids who grow up reading information this way develop a default question they can ask of anything: who benefits when I take this at face value? Motivation is one of several dimensions of this ACRL frame, alongside questions of attribution, access, and whose voices get heard; it’s the dimension most directly applicable to the content kids encounter in everyday life.

In the home: When your kid wants to watch a free video or play a free game, ask them “who do you think is paying for this to exist?” Whatever they answer, the question opens a conversation you can come back to: free content always has a cost attached, and the cost just gets paid in something other than money. For a free video, your kid is paying with attention that the platform sells to advertisers. For a free game, they’re paying with time and data, and often with temptation to buy upgrades inside the game. The same approach surfaces motivations beyond money in any other kind of content. Naming this out loud, repeatedly, builds the question your kid will start asking on their own about anything that reaches them: who’s getting something out of me reading or watching this?

In the classroom: Pick a piece of content the class has seen, and spend twenty minutes tracing what the maker wanted from the audience. The incentive might be commercial, where an advertiser or sponsor is working to drive sales, or it might be persuasive, where a campaign or organization is working to shift opinion on an issue. The exercise puts the question of motivation in front of students directly, so they can practice reading content for the purpose driving it rather than only for what it says on the surface.

4. Good questions get sharper as you ask them.

Framework name: Research as Inquiry

Research doesn’t happen in a single act of looking something up. A question goes in, partial answers come back, the question gets refined based on what those partial answers reveal, and the cycle repeats. Meaningful questions rarely have clean single answers, and researching well includes the capacity to stay with a question while it sharpens rather than collapsing it too early into a premature conclusion.

AI chat trains the opposite reflex: a chatbot offers a single finished answer to a single question, and the interaction feels complete as soon as it ends. Kids who only practice that pattern miss the underlying skill of refining questions as they learn more. The skill is what separates researching a topic from simply collecting answers.

In the home: When your kid asks you a question, try bouncing it back first: “What do you already think about that?” or “How could we find out together?” These simple questions invite your kid into the thinking process rather than handing them a finished answer. Over time, children can develop their own opinions and their own research instincts, instead of waiting for the adult in the room to tell them what’s true.

In the classroom: For any research assignment, build in two checkpoints where students rewrite their research question based on what they’ve learned so far. By the end of the project, students can compare their final question to the one they started with and see how it changed. The point of the exercise is to make the iterative nature of research visible, so students experience research as something that reshapes the question itself (and not just a hunt for answers to a fixed one).

5. No serious idea stands alone.

Framework name: Scholarship as Conversation

Substantive ideas rarely get developed by people working in solitude. They emerge from groups of people thinking and writing in response to each other over time, with each contribution shaped by what came before it and shaping what comes after. Making sense of something means recognizing the larger argument it's a part of, including the earlier work it's reacting to and the work it will provoke in turn.

The frame shows up vividly in academic writing, where literature reviews and citations make the conversation visible on the page. The same pattern operates in journalism that builds on earlier reporting and in political writing that engages older arguments. AI-generated summaries do something more concerning, because they compress many conversations into a single confident-sounding voice that erases the evidence a conversation ever existed.

In the home: When you’re reading a book or watching a show with your kid, point out the moments where the book or show is reacting to something else. You can say things like “This story is making fun of those old fairy tales we read last year” or “This creator made a video to argue with what that other person posted last week.” Comments like these position the book or show as part of a larger dialogue rather than as a standalone pronouncement, introducing the question a kid can carry into anything they read or watch: what is this responding to or building on?

In the classroom: Take any text students are already reading and teach them to look at its bibliography or works cited section. The simplest question to ask is who the author is building on or arguing with. Walk the class through a single citation chain: pick one reference in the text, and look at that source together to see what it says and how the original text engaged with it. Students learn that every serious piece of writing is part of a longer, larger conversation, which can change how they read everything afterward.

6. Finding good information takes more than one search.

Framework name: Searching as Strategic Exploration

Skilled searching looks nothing like a single query producing a single result. It’s a process of trying something, seeing what comes back, adjusting based on what the results reveal about the topic and the tools, and trying again with sharper terms. The craft lives in knowing when to try again and what to vary when the first search doesn’t deliver.

AI chat has flattened this craft for many users, because a chatbot offers the surface of an answer on the first try. The pattern trains our reflexes away from iteration. Kids who grow up inside that reflex miss the underlying skill, which is the ability to move strategically through an information landscape when a first attempt doesn’t give them what they need.

In the home: When your kid watches you look something up, show them that you try more than one search. Say out loud when a search doesn’t give you what you need, and try a different approach. You can narrate it directly, with something like “Hmm, that didn’t work, let me try different words,” or “this site doesn’t seem reliable, let me look somewhere else.” The modeling teaches your kid that search is an iterative process rather than a single click.

In the classroom: Put several search tools in front of students and have them run the same question through each one. A library database and an AI chatbot will return noticeably different results for the same question, and comparing the differences shows students how information is structured in ways that aren’t visible from any single tool. The exercise also makes clear that no search tool gives a complete picture, because each one is designed to prioritize certain results over others. The ability to turn to multiple sources, or refine their search process over time, is the foundation to students’ verification strategies as they get older.

The arc over time

The information environment kids are growing up in now will keep shifting, and the shifts will inevitably keep coming. What makes the ACRL framework durable across these shifts is that the frames describe dimensions of information that stay stable when the tools change. Authority is always specific to a domain. Information is always produced through some process, and the process shapes what the output can carry. Every piece of information moves through systems of incentive, and every serious idea takes part in a larger conversation, whether or not the tool it arrives through makes the conversation visible.

Kids raised on these habits grow into adults who can encounter new information and truly see what’s in front of them. They can recognize when an author is reaching beyond their expertise, when incentives are shaping what gets said, when an article is in conversation with other articles, and when the first search result isn’t the last word. They can grow into voters and citizens who think the way they want to think rather than the way someone else wanted them to. This capacity gets built in the small conversations of home and classroom, long before the first vote gets cast or the first big decision needs making.

What these skills add up to has a name: information resilience, the capacity to meet any claim, from any source, without being knocked off course by what arrives. A resilient reader can stay with a question while it sharpens, and can hold her ground when the people around her are settling on conclusions faster than the evidence warrants. She has an internal sense for what’s worth a closer look, built up over years of practice. That kind of resilience is what parents and teachers can give the kids growing up now, and it will keep working long after any specific tool or platform has been replaced.

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

If you’d rather listen than read, I recorded an audio version of this essay for paid subscribers at the end of the post. Thank you for being here!

Sometime in the mid-2000s, most of us started handing over pieces of ourselves to the internet without giving the exchange a second thought. We created email accounts, signed up for social media, bought things online, downloaded apps, swiped loyalty cards, connected fitness trackers, stored photos in the cloud, and agreed to terms of service that almost none of us have ever read in full. We did this thousands of times over two decades and counting, and each interaction felt small enough to be inconsequential.

But the accumulation is enormous. More than 6 billion people now use the internet, and each one makes an estimated 5,000 digital interactions per day. Most of those interactions happen without our conscious awareness: a GPS ping, a page load, an app opening, a browser cookie refreshing, a device checking in with a cell tower. The average person in 2010 made an estimated 298 digital interactions per day. In fifteen years, that number multiplied more than sixteenfold. Those digital interactions produce records that can persist indefinitely, stored, copied, indexed, bought, sold, and combined with other records to build profiles of extraordinary detail.

If we’ve been online since the late 1990s or early 2000s, our data footprint can include social media accounts we’ve created, online purchases we’ve made, forums we’ve posted in, loyalty cards we’ve used, and apps we’ve installed going back decades. Some of that information lives on platforms we’ve long forgotten. Some of it was collected by companies that have since been acquired or dissolved, with our data potentially passing to successor entities we’ve never heard of. The digital life most of us have been living for 15 to 25 years has produced a layered, evolving archive that only grows more valuable to the people who buy and sell it as time goes on.

Most of us sense that something is off about all of this. In a 2023 survey, Pew Research found that roughly eight in ten Americans feel they have little to no control over the data companies collect about them, 71% are concerned about government data use, and 67% say they understand little to nothing about what companies are doing with their personal information. The concern is real and widespread. And so is the feeling of helplessness: 60% of Americans believe it’s impossible to go through daily life without having their data tracked. The unease is there. What’s missing is a clear picture of what’s happening on the other side of the transaction.

What “My Data” Looks Like as a File

When we say “my data,” we tend to picture the things we’ve actively typed into a form: our name, email address, maybe a credit card number. But the scope of what companies collect and brokers sell extends far beyond what we’ve consciously shared. The majority of our data profile is generated not from what we’ve entered but from what we’ve done: where we’ve gone, what we’ve browsed, how long we’ve lingered, what we’ve bought, who we’ve contacted, and what patterns emerge when all of those behaviors are tracked over time.

Beyond the identifying details we'd expect (name, date of birth, government IDs), the categories of personal information being collected, sold, and traded include, among others:

• Financial records: credit history, transaction logs, bank account activity, loan applications

• Location data: GPS coordinates from our phones, Wi-Fi connection logs, cell tower pings, the places we’ve checked in and the routes we’ve traveled

• Behavioral data: the sites we’ve browsed, the searches we’ve run, the products we’ve lingered on, the apps we’ve opened and how long we’ve spent inside them

• Biometric data: fingerprints, facial recognition templates, voiceprints

• Communications metadata: who we’ve contacted, when, how often, and from where, even when the content of the message itself isn’t captured

• Health-related data: pharmacy purchases, fitness tracker output, symptom searches, insurance claims

• Social data: our contacts, our connections, our group memberships, who we interact with and how frequently

These categories don’t exist in isolation. For example, a pharmacy purchase is one data point on its own. Combined with a location trail, a search history, and a social media profile, it becomes part of a behavioral mosaic that can be used to infer things we never disclosed: our health conditions, our financial stability, our family status, our political orientation even. Acxiom, one of the largest data brokers, advertises more than 10,000 unique data attributes in its consumer profiles. The profile that results from all of this collection isn’t a list of facts we shared. It’s a composite portrait assembled from fragments that we never intended to be read together.

Who Has It

“They have my data” implies a single entity of ownership, but the data ecosystem is a layered network and each layer operates differently. The platform we signed up for, the trackers running behind the webpage we visited, the broker who bought our behavioral profile, and the insurance company that used it to adjust our premium are all separate organizations with separate business models. Understanding who “they” are requires seeing how these layers connect.

The most visible layer is the platforms and services we interact with directly: Google, Apple, Meta, Amazon, Microsoft, and hundreds of smaller apps and websites. These companies collect data as a condition of use. Google processes billions of searches per day (estimates range from 8.5 to 14 billion depending on the source and methodology). Facebook’s engineering team disclosed in 2014 that its data warehouse alone was ingesting about 600 terabytes of new data per day, and the volume has grown substantially since. Every interaction on these platforms produces records that are stored, indexed, and fed into behavioral models.

Behind those platforms sits an advertising and tracking infrastructure that most of us never see. When we load a webpage, dozens of third-party trackers can fire simultaneously, each one logging our browser type, our device, our location, our referring page, and our behavior on the site. A single visit to a news article or product page can involve data transmissions to 50 or more companies we’ve never interacted with directly. These ad networks and analytics firms build cross-platform behavioral profiles that follow us from device to device, assembling a picture of our habits that no single app or website could construct alone.

And finally, there are the end buyers: insurance companies, financial institutions, employers, landlords, retailers, political consultants, government agencies, and AI companies. These are the organizations that purchase brokered data and use it to make decisions that affect our lives directly, from the interest rate we’re offered on a loan to the price we’re shown for a pair of shoes to whether our rental application gets approved. The distance between the data we generated and the decision it informs can be vast, and the connection between the two is almost never disclosed.

Where It Lives and How It Moves

Our data doesn’t sit in one place waiting to be accessed. It’s distributed across more than 4,000 data centers in the United States alone, operated by technology companies, cloud providers, brokers, and government agencies. Cloud storage worldwide is projected to exceed 200 zettabytes in the coming years, a figure that translates to more than 200 trillion gigabytes.

But the numbers matter less than the mechanics. What makes the data economy so difficult to see is the way information flows between layers. A common path looks something like this: a fitness app collects our running routes and resting heart rate. The app’s developer shares that data with an analytics partner. The analytics partner sells aggregated behavioral data to a data broker. The broker combines it with our purchase history, our location patterns, and our credit profile, then sells the resulting bundle to an insurance company, a financial institution, or a marketing firm. At no point in this chain did we interact with anyone beyond the fitness app. We agreed to the app’s terms of service, which included a clause about sharing data with “third-party partners,” and the rest of the chain followed from there.

At each stage, the data is processed: cleaned, categorized, cross-referenced with other datasets, scored, and segmented. Algorithms organize us into consumer profiles based on inferred income, predicted purchasing behavior, estimated health risk, political leanings, and thousands of other variables. The processing is what transforms raw data into something commercially valuable, and it happens largely outside our awareness.

What They’re Using It For

Some of the ways our data gets used are familiar, and some of them are useful in ways we can feel. When a streaming platform recommends a film based on what we’ve watched before, or a search engine surfaces local results because it knows our location, or a retailer suggests a product similar to one we recently bought, what we’re seeing data-driven personalization working as intended. Most of us have experienced moments where targeted content saved us time or introduced us to something we wouldn’t have found otherwise. The advertising model also funds a tremendous amount of the free content and services we use every day, from search engines to email to social media to news.

Insurance companies purchase lifestyle and behavioral data to assess risk and adjust premiums without ever asking us directly. If a data profile suggests certain health patterns, certain driving routes, or a certain kind of neighborhood, those details can shape what we’re offered and what we’re charged. The data acts as a proxy questionnaire we never filled out.

Financial institutions use brokered data for credit decisions, loan eligibility, and fraud detection. Some of that serves a protective function. But the algorithms doing this work are proprietary and opaque, and studies have documented that algorithmic credit scoring models produce systematically lower scores for Black and Latino communities compared to white and Asian populations. Consumers do have legal rights under the Fair Credit Reporting Act (FCRA) to dispute inaccuracies, and lenders are required to provide adverse action notices explaining the principal reasons for a denial. But consumers often don’t know which brokered data fed into the score in the first place, and companies aren’t required to reveal the proprietary formulas their models use. The Consumer Financial Protection Bureau (CFPB) has noted that data brokers operating in a gray area between regulated credit reporting and unregulated data sales make the process of identifying and challenging errors especially difficult in practice.

Employers and landlords use data from people-search sites to screen applicants. These sites source their information from data brokers and public records, and they frequently contain errors, because brokers don’t verify what they aggregate. The Federal Trade Commission and organizations like the Electronic Privacy Information Center have documented cases where inaccurate broker data cost people jobs and housing.

Retailers use personal data to set individualized prices. In January 2025, the Federal Trade Commission released findings from a study of what it calls “surveillance pricing,” in which intermediary firms hired by retailers track browser history, mouse movements, purchase patterns, and location to adjust the price of the same product for different buyers. The FTC described a scenario in which a consumer profiled as a new parent would be shown higher-priced baby products at the top of their search results. New York became the first state to require retailers to disclose when a price was set by an algorithm using the consumer’s personal data; the law was signed in May 2025 and took effect in November of that year after surviving a legal challenge. Also in 2025, dozens of state legislatures introduced bills to regulate various forms of algorithmic pricing, including surveillance pricing and algorithmic rent-setting.

AI companies are training their models on personal data scraped from the web. Web crawlers pull content from blogs, social media profiles, online marketplaces, photo-sharing platforms, and anywhere else that isn’t behind a login wall. A 2025 MIT Technology Review investigation of a major AI training dataset found thousands of identifiable faces, identity documents, and job applications in a sample representing just 0.1% of the data, and estimated the full set contained hundreds of millions of images with personal information. Meta has said its AI models are partially trained on public Facebook and Instagram posts. LinkedIn began using member data to train its AI tools before updating its terms of service to reflect that change. Because AI companies have scraped content posted long before generative AI existed, the people whose data is in those training sets never had the opportunity to consent to that use. And there’s a feedback loop: we generate data by using AI systems, that data refines those systems, and those systems shape the information environment we navigate. Our data becomes part of the infrastructure that determines what information reaches us next.

Government agencies buy personal data from the same commercial brokers that serve advertisers and insurance companies. The Supreme Court ruled in 2018 that law enforcement needs a warrant to access a person’s historical cell phone location data. But federal agencies, including the FBI, ICE, and the Department of Defense, have argued that purchasing location data from a commercial broker is a market transaction, not a compelled disclosure, and therefore doesn’t require one. The Brennan Center for Justice has called this a loophole that allows agencies to bypass constitutional protections, and has documented cases including the Department of Defense purchasing location data collected from prayer apps to monitor Muslim communities. The same data pipeline built for advertising can be repurposed for surveillance, and the legal framework hasn’t caught up to that reality.

And the brokers themselves get breached. In July 2025, hackers accessed names, Social Security numbers, and dates of birth for more than 4.4 million people through a third-party application used by TransUnion, one of the three major U.S. credit bureaus. In a separate incident disclosed that same year, LexisNexis Risk Solutions confirmed that hackers accessed names, Social Security numbers, driver’s license numbers, and dates of birth for more than 364,000 people through a third-party software development platform. The companies that centralize our data become single points of failure, and when they’re compromised, the exposure isn’t one transaction or one relationship. It’s a cross-section of an entire life in one place.

Data We Never Shared

Not all personal data comes from something we’ve handed over. Companies also generate what’s known as inferred or derived data: new information produced by running existing records through predictive algorithms. The inference is drawn from the pattern, not from anything we volunteered. In 2012, a New York Times feature revealed that Target had built a pregnancy prediction algorithm around roughly 25 products whose purchase patterns correlated with pregnancy stages: unscented lotion, certain vitamin supplements, extra-large bags of cotton balls. The algorithm could estimate a due date within a narrow window based entirely on shopping behavior.

Insurance companies can infer health risks from purchase histories and location patterns. Financial institutions can infer economic instability from app usage and transaction frequency. Data brokers categorize consumers into segments like “single parents,” “fitness enthusiasts,” and “budget conscious households” based on behavioral inferences, not declared preferences. These inferred profiles are then sold to the same range of buyers as declared data, with the same range of consequences, but we never agreed to share the information those profiles contain - the information didn’t exist until a model generated it from our behavior.

This means that even careful, privacy-conscious choices about what to share can be partially circumvented by inference. We might choose not to disclose a health condition, a pregnancy, a financial difficulty, or a political affiliation, and a predictive model can generate a probability estimate of that very thing based on the patterns in the data we did share. The profile that follows us through the data economy isn’t limited to what we put into the system. It includes what automated models have inferred from patterns in our behavior, patterns that become labeled, scored, and treated as facts by the companies that buy them.

Why the Context Is the Problem

The philosopher Helen Nissenbaum has a framework for what’s happening here: contextual integrity. The idea is that privacy isn’t about secrecy. We share information willingly all the time, when the context fits. We tell our doctor about a health condition because we expect that information to stay within the medical relationship. We search for symptoms on a health website because we assume that search won’t follow us into an insurance application. In the current data economy, that’s exactly the kind of boundary that dissolves, because the company collecting the data and the company buying it are operating in completely different contexts.

This is an information literacy problem as much as a privacy problem. Information literacy is usually framed around consumption: evaluating sources, questioning claims, recognizing bias in what we read and watch. But every time we interact with a digital service, we’re also producing information: generating a record that will be read, interpreted, scored, and acted on by organizations we may never interact with directly. Many of us have gotten better at questioning the information that comes at us: checking sources, noticing bias, and recognizing when something is trying to sell us a conclusion. But we haven’t developed equivalent habits around the information that flows from us: where it goes after we hand it over, who reads the record, what incentives they have, and what conclusions they draw. The gap between what we think we’re consenting to and what we’ve agreed to in practice is where the real exposure lives, and the system is designed to keep that gap invisible.

The Illusion of Choice

One of the reasons the “so what” question is hard to answer with action is that opting out of data collection often means opting out of participation. Declining a social media platform’s terms of service means not using the platform. Refusing location permissions can mean losing access to navigation, ride-sharing, weather, and delivery apps. Choosing not to create an account can mean paying more, seeing less, or being locked out of services that have become essential infrastructure for work, communication, healthcare, banking, and education.

The architecture of digital consent treats data sharing as a binary: agree to the terms or don’t use the product. There’s rarely a middle option that allows us to use a service while limiting what data gets collected and where it goes. The result is that the “choice” to share data often functions as a condition of entry into daily life rather than an informed negotiation. We’re not handing over data because we’ve weighed the tradeoff and decided it’s fair. We’re handing it over because the alternative is exclusion from services we rely on.

This is the structural context behind the Pew Research Center finding that more than half of Americans believe it’s impossible to go through daily life without being tracked. For many of us, it isn’t possible, at least not without significant inconvenience or sacrifice. The question isn’t whether we can avoid data collection entirely, because for the vast majority of people who participate in modern life, the answer is no. The question is whether we can make more informed decisions within the constraints we’re operating in, and whether the system can be pushed - through regulation, through market pressure, through better tools - toward something more transparent.

Can We Get It Back?

California’s Delete Act, which took effect in January 2026, is the strongest example of what’s emerging. It created a platform called DROP (Delete Request and Opt-Out Platform) that lets California residents submit a single deletion request to every registered data broker in the state. Brokers are required to process those requests, maintain suppression lists to prevent re-collection, and check the platform regularly for new requests. The European Union’s GDPR provides similar individual rights, and a handful of other U.S. states have enacted their own privacy laws with varying levels of protection. But the coverage is uneven: what’s available to a California or EU resident may not extend to someone in a state without comparable legislation.

Some services now automate parts of the opt-out process, submitting removal requests to dozens of brokers on our behalf. These can’t erase the data trail entirely, but they can narrow what’s actively available for sale.

Beyond deletion, there are smaller choices that reduce how much new data we generate. We can audit which apps have permission to track our location or access our contacts, since a surprising amount of behavioral data comes from apps that don’t need those permissions to function. We can treat “sign in with Google” and “sign in with Facebook” buttons as what they are: data-sharing agreements that can link a new service to an existing profile. And we can glance at the first few lines of a privacy policy before agreeing, looking for some version of “we may share your information with our partners,” where “partners” just means anyone willing to pay.

So can we get it back? Not entirely. Data that’s already been collected, copied, sold, and processed across multiple systems can’t be fully recalled. What we can do is reduce what’s actively available for sale, slow the flow of new data going forward, and take advantage of legal tools that didn’t exist a few years ago. The archive of our past digital lives is too distributed to undo, but the file is still being written, and we have more say over the next page than we did over the last twenty years of them.

The Other Side of the Transaction

Most of us don’t read privacy policies, and the policies aren’t built to be read. They average thousands of words of dense legal language filled with terms like “legitimate interest,” “data processor,” and “de-identified data.” Studies consistently put them at a late high school to early college reading level (grade 12 to 14), but the difficulty goes beyond reading level: the concepts are abstract, the volume of agreements we encounter is enormous, and the design of the consent process itself pushes us through as fast as possible. Pre-checked boxes, auto-scrolling agreement windows, “accept all” buttons positioned prominently while “customize settings” options sit behind additional clicks. These are dark patterns, design choices that make the path of least resistance the path of maximum data sharing.

The result is a gap between the moment we share a piece of information and the moment that information shapes a decision about our lives. We don’t connect the app to the insurance premium or the loyalty card to the rental application because the chain of custody between them is long, complex, and designed to stay out of view.

The same critical thinking we’ve learned to apply to the information flowing toward us (checking sources, questioning claims, looking for bias) applies to the information flowing from us: who’s collecting this, what will they do with it, who else will see it, and what did we agree to? The difference is that in the data economy, we’re the product being evaluated, and the questions are being asked about us rather than by us.

So what if they have our data? The tradeoff extends well beyond better ads. It reaches into the prices we’re charged, the credit we’re offered, the jobs we’re considered for, the insurance premiums we pay, the AI systems trained on our behavior, the accuracy of the profiles used to make decisions about our lives, and the degree to which government agencies can monitor our movements without a warrant. Every new service we sign up for, every permission we grant, and every terms-of-service agreement we accept adds another layer to that file. We can’t close the file entirely, but we can make more informed decisions about what goes into it next.

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

The standard advice for moments like this tends to be some version of the same instruction: check the facts and verify the claim. And that advice isn’t wrong, exactly. Fact-checking is a real and necessary skill. But it addresses only one kind of vulnerability: the risk of believing something that’s factually false. And the information environment we’re navigating now produces other kinds of vulnerability that fact-checking alone can’t reach.

The most common image of misinformation is the outright fabrication, a claim that can be tested against the evidence and shown to be false. When a fake statistic circulates or a public figure is quoted saying something they never said, fact-checking can trace the claim back to its source and measure it against the record. But the most persuasive arguments circulating right now, the ones that shift public opinion and reshape how we think about entire categories of issues, often don’t need false claims to be effective.

These arguments rely on framing. They use selective emphasis, emotional resonance, strategic omission, and structural logic that feels airtight because every individual piece checks out. The facts are real and the sources are legitimate. And the conclusion is still misleading, because the persuasion isn’t happening at the level of individual claims. It’s happening in the architecture of the argument: which facts were chosen, which were left out, how the remaining facts were sequenced, and what emotional context was wrapped around them.

This is a gap in our current information literacy toolkit. We’ve invested heavily in teaching people to spot fabricated claims, and that investment has produced real results: the Duke Reporters’ Lab counted 110 fact-checking projects worldwide in 2014 and more than 450 by 2022, and as of early 2026, 25 U.S. states have media literacy laws on their books. But framing-based manipulation, which is as old as persuasion itself, has received less attention as something we can learn to recognize and resist. The result is a lopsided defense: strong against outright falsehood, much weaker against carefully assembled truth. The imbalance becomes clearest in the specific situations where every fact checks out and the argument is still misleading.

The Limits of Fact-Checking

Fact-checking has become one of our most prominent tools against misinformation for good reason: when the problem is a false claim, fact-checking works. At its core, fact-checking is an exercise in verification: does this claim hold up against the available evidence? Some fact-checking organizations use nuanced rating scales to capture degrees of accuracy, but the fundamental question being asked is still whether the claim is supported by the facts. That question is useful when the problem is fabricated information, and fabricated information is still a real and widespread problem. But fabrication is also the form of manipulation that’s easiest to catch, because it can be disproven with evidence. The harder strategy to counter, and often the more effective one, is constructing narratives from true facts arranged to support a predetermined conclusion.

Consider how this works in practice. Two news outlets can cover the same event using the same verified data and produce stories that lead readers to completely different conclusions. Neither outlet is lying. Both can point to their sources. The difference is in what gets emphasized and what gets left on the cutting room floor. A fact-checker verifying the individual claims in either article would confirm them as accurate, even though the two stories point in different directions. The manipulation is structural, not factual.

Or consider how persuasive arguments circulate on social media. A post might cite a real study and draw a conclusion that sounds evidence-based. Fact-checking the citation confirms the study exists. Fact-checking the data confirms the numbers are real. But the post might omit that the study’s own authors cautioned against the exact conclusion being drawn, or that subsequent research contradicted the findings entirely. Every checkable fact passes inspection. The argument is still misleading.

When we train ourselves only to ask “Is this true?”, we’re left without tools for these situations. And these situations are where we’re most vulnerable, because the presence of verified facts creates a false sense of security. If every individual claim in an argument can be confirmed, we’re naturally inclined to trust the conclusion, even when the conclusion depends on framing choices and omissions that a standard verification check wouldn’t surface. Developing a defense against this kind of manipulation starts with understanding two very different ways of engaging with arguments we disagree with.

Straw-Manning and Steel-Manning

Every day, we encounter arguments: in news articles, in social media posts, in conversations with coworkers and family, in political speeches, in marketing. And every day, the people making those arguments are making choices about how to present them. They’re choosing which evidence to include, which counterarguments to address, how to characterize the opposing view, and how to frame the conclusion. These choices are sometimes called rhetorical moves, which is just a way of saying that they’re strategic decisions about how to make an argument more persuasive. We all make these choices when we argue, whether we’re conscious of doing so or not.

Some of these choices are productive. When a speaker anticipates a strong objection and addresses it head-on, the overall argument gets stronger because the audience can see that the counterargument was considered and responded to. But other rhetorical moves work in the opposite direction, weakening the argument they’re responding to and making it easier to knock down a position by first making it flimsier than it really is. Recognizing the difference between these two kinds of moves changes how we evaluate every argument we encounter, including our own.

Two rhetorical moves in particular sit at opposite ends of this spectrum: straw-manning and steel-manning. A straw man weakens an opposing argument to make it easier to dismiss, while steel-manning does the reverse, strengthening an opposing argument to make our own engagement with it more rigorous. Both show up constantly in political debates, news coverage, social media arguments, and everyday conversations, and learning to identify each one changes how we process the information we’re taking in.

Straw-Manning

A straw man argument replaces someone’s real position with a distorted version that’s easier to attack. This distortion can be a deliberate rhetorical strategy, designed to win an argument without engaging with its substance. But it can also happen without any conscious intent, when we’re so certain of our own position that we stop trying to understand what the other side is saying and start arguing against what we assume they mean. The result, either way, is that we end up attacking a position that the people who hold the real view wouldn’t recognize as theirs.

Straw-manning tends to follow recognizable patterns:

Exaggerating to extremes.

A specific position gets inflated into the most extreme possible version of itself. The extreme version is easier to dismiss because it sounds unreasonable on its face, and dismissing the extreme version lets someone avoid engaging with the more moderate original position altogether. Someone who says AI tools can help students learn, for instance, gets characterized as wanting to replace teachers with machines.

Choosing the weakest opponent.

Every position has a range of people arguing for it, from careful and well-informed voices to inflammatory and poorly reasoned ones. This technique involves ignoring the careful voices entirely and responding only to the least credible one, as though that person speaks for everyone who holds a remotely similar view. This works because defeating a weak argument feels like defeating the whole position, even though the stronger and more common versions were never addressed. Instead of engaging with a detailed policy proposal about AI transparency, someone responds to the most inflammatory anti-AI post they can find and argues against that version instead.

Stripping away nuance.

A position built on conditional reasoning and careful distinctions gets flattened into an oversimplified line that loses the substance of the original argument. The simplified version can be knocked down quickly, while the full argument, with its reasoning and caveats intact, would require a much more serious response. A nuanced argument about how AI might displace certain types of jobs while creating others, for instance, gets reduced to "they think robots are coming for all our jobs."

Questioning motives instead of addressing the argument.

Rather than engaging with the substance of what someone argued, this technique redirects the conversation toward the presumed psychology of the person making the argument. If the conversation becomes about who someone is rather than what they said, the argument itself never has to be addressed on its merits. Someone raises a specific concern about AI-generated misinformation, and the response is "you're just afraid of new technology" rather than any engagement with the concern itself.

When we encounter a straw man, whether in media coverage or in a conversation, the most effective counter is to restate the original position accurately and identify specifically what was distorted. This shifts the discussion back to the real argument and makes the misrepresentation visible. Even when we disagree with the original position, naming the distortion keeps the conversation grounded in what people are arguing rather than what their opponents claim they’re arguing.

The challenge is that straw-manning works precisely because the distorted version tends to be more emotionally compelling than the real position. “They want to replace teachers with machines” provokes a stronger reaction than “They think AI tools can supplement classroom instruction.” In fast-moving conversations, especially in public forums and political debates, the straw man can take hold before anyone has a chance to correct the record, and once it takes hold it tends to become the version that sticks.

When we’re on the receiving end of a straw man, one of the more effective responses is to restate our actual position with enough specificity that the distortion becomes obvious by contrast. Rather than saying “That’s not what I said,” which sounds defensive and leaves the straw man intact, we can say something closer to: “My position is that AI tools can supplement what teachers are already doing in the classroom. The argument you’re responding to, that teachers should be replaced, isn’t one I’m making.” This forces the conversation back to the real disagreement and gives anyone listening the information they need to see the gap between the position being attacked and the position being held.

When we see straw-manning in media or political discourse, the counter is different because we’re usually observers rather than participants. In those situations, the skill is recognition: noticing when a commentator is responding to a position that nobody credible is making, or when a debate treats a caricature as though it represents the other side’s strongest case. That recognition doesn’t require us to respond publicly. It changes how much credibility we give the argument we’re consuming, because an argument that has only defeated a straw man hasn’t proven anything about the real position.

Steel-Manning

A steel man is the direct opposite: instead of weakening an argument to dismiss it more easily, we strengthen it to engage with it more seriously. When we steel-man a position we disagree with, we reconstruct the strongest possible version of that position, the version its smartest and most reasonable proponents would put forward if given the chance to make their best case. The method involves assuming good faith in the people who hold the opposing view and building the version of their argument that would be hardest for us to refute, setting aside the worst-case interpretations of their motives in order to engage with the substance of what they’re saying.

We steel-man an opposing argument for one of two reasons: either to discover that the opposing position has merit we hadn’t recognized (which makes our own understanding more accurate), or to be able to argue our own position from a much stronger and more grounded place (because we’ve already engaged with the best the other side has to offer rather than the worst). Both of these outcomes make us better thinkers. Neither requires us to abandon our original view.

There’s a practical test for whether a steel man has been done well: if we can restate the opposing position back to someone who holds it and they respond with something like, “Yes, that’s what I believe,” then we’ve succeeded. We’ve demonstrated that we understand their position on its own terms, not a distorted or simplified version of the position. Once we’ve reached that point of genuine understanding, the conversation changes. We know exactly which specific points we disagree on and which points the other side has right, which means we can focus our pushback on the real points of divergence instead of wasting energy on misunderstandings. And the person we’re engaging with is far more likely to listen to our counterarguments, because we’ve shown that we took the time to understand theirs first.

Steel-manning someone’s position doesn’t mean we’ve conceded anything. It means we’ve done the intellectual work of comprehending why they believe what they believe, so that if we still disagree, our disagreement is precise and informed rather than reactive. We’re not meeting them halfway; we’re making sure we know exactly where the halfway point is before we decide where we stand.

Steel-manning has its own recognizable characteristics:

Engaging with the strongest version of the argument.

When someone makes a point poorly or emotionally, our instinct is often to respond to the weakest interpretation of what they said. Steel-manning means doing the opposite: reconstructing the best version of what the person is trying to communicate, even when their phrasing makes it easy to dismiss. The weakest interpretation is rarely what the person means, and responding to it leads to an argument about words rather than ideas. If someone says "AI is going to destroy creativity," steel-manning asks us to consider what concern is driving that statement (perhaps that AI-generated content will crowd out original human expression, or that people will lose the motivation to develop creative skills if a machine can produce results instantly) rather than responding to the literal claim that creativity will cease to exist.

Taking the evidence seriously.

When we disagree with a conclusion, the natural impulse is to dismiss the evidence along with the conclusion. But the evidence and the conclusion are separate things: the research or observations someone cites in support of their argument are often real, even when the conclusion drawn from them is one we don't share. Steel-manning means engaging with that evidence on its own terms and asking what it does and doesn't support, rather than waving it away because we don't like the other person's conclusion. If someone arguing against AI in recruitment cites data showing that a specific AI screening tool rejected qualified candidates at higher rates from certain demographic groups, that data stands on its own, even if we disagree with their conclusion that AI shouldn't be used in hiring at all.

Locating the precise point of disagreement.

Once we've built the strongest version of the opposing argument and engaged with its evidence, we can pinpoint exactly where our reasoning diverges, and just as importantly, where it doesn't. Before steel-manning, a disagreement often feels like a broad sense that "they're wrong." After steel-manning, we can see which parts of their argument we find convincing and where exactly we stop being convinced. That turns a vague reaction into something specific and articulable, for example: "I accept their evidence about the economic pressure AI puts on creative professionals, but I weigh the benefits of broader access to creative tools differently." That kind of precision makes the disagreement productive rather than circular, because both sides can focus on the actual point of contention rather than talking past each other.

Engaging with the substance, not the person.

Steel-manning starts from the premise that the person making the argument has legitimate reasons for holding their position, even when we disagree with their conclusion. Rather than speculating about someone's motives or dismissing them as uninformed, steel-manning addresses what they said on its own terms. If someone argues that AI chatbots shouldn't be used for mental health counseling, steel-manning engages with their specific concerns about AI's ability to handle crisis situations and read the nonverbal cues that human therapists rely on, rather than dismissing the person as "someone who doesn't care about the millions of people who can't afford a therapist."

Steel-manning isn’t necessarily about “seeing the other side” or conceding ground. We’re not looking for compromise or trying to meet in the middle. We’re stress-testing our own reasoning against the best available challenge, and the result is a position that has been refined by contact with the strongest counterarguments rather than insulated from them.

The adversarial legal system works on a similar principle. A defense attorney who builds the strongest possible case for their client isn’t declaring a personal belief in the client’s innocence; they’re providing the vigorous advocacy that ensures the prosecution’s case gets properly tested before a verdict is reached. When we steel-man an argument, we’re doing something similar: ensuring that our own position has been tested against the best available opposition, not the weakest.

Steel-manning is also distinct from playing devil’s advocate, which tends to be a contrarian exercise performed without a specific analytical goal. Devil’s advocacy often means arguing a position we don’t hold just to see where the conversation goes, while steel-manning has a more focused purpose: locating the precise point of disagreement between two positions by engaging with both at their strongest. And steel-manning is different from the “both sides” reflex that treats every disagreement as though the truth must be somewhere in the middle. Sometimes, after steel-manning an opposing argument, we come away more convinced of our original position, because we’ve identified exactly where the opposing argument’s reasoning breaks down and can articulate why.

Learning to steel-man also sharpens our ability to recognize when someone else isn’t doing it: when a commentator is responding to a caricature rather than a real position, or when an argument seems persuasive primarily because it never engages with the strongest version of its opposition. That recognition helps us evaluate not just whether a claim is true, but whether the reasoning behind the claim has been tested against its best counterarguments. When we can see that an argument has only engaged with the weakest version of its opposition, we know that the conclusion, however confident it sounds, hasn’t been earned.

How to Steel-Man in Practice

Knowing what steel-manning is and recognizing it in the abstract is one thing; doing it in real time is another. The practice follows a concrete set of steps, but it does require some discipline, because when we encounter an argument we disagree with, our first impulse is usually to look for the weak point, the thing we can refute. Steel-manning asks us to resist that impulse temporarily.

Restate the argument in the most generous terms possible.

We can start by stripping away the inflammatory language and the bad-faith framing. What is the core claim? If the person making this argument were thoughtful and arguing in good faith, how would they phrase their position?

Identify the strongest evidence supporting the argument.

Even if we think the conclusion is wrong, what real-world observations or data points would a reasonable person cite in its favor? Where does the argument connect to something observable and verifiable? This step often reveals that opposing positions are grounded in real evidence we hadn’t considered, which is useful information whether or not it changes our mind.

Name the values driving the argument.

Many disagreements, especially polarized ones, have a values component that runs deeper than the factual claims on the surface. Two people can agree on the same data and still reach different conclusions because they’re weighing competing priorities. Someone arguing against AI regulation might be prioritizing the speed of innovation and the risk that regulation would entrench dominant companies. Someone arguing for AI regulation might be prioritizing public accountability and protection for people affected by automated decisions. Both sets of priorities are legitimate. The disagreement is about which one takes precedence in a specific context.

Find the precise point where we diverge.

Once we’ve built the strongest version of the opposing argument, the next step is to go through it piece by piece. Which parts of the argument do we find convincing, and where exactly does our reasoning stop following theirs? Working through the steel-manned argument this way narrows the disagreement from a broad reaction to a specific location. Maybe we accept the same facts but weigh the risks differently. Maybe we share the same goal but disagree about the mechanism for reaching the goal. Locating this point transforms a vague sense of “that’s wrong” into a specific, articulable position.

Working Through It

The downloadable worksheet below walks through the complete steel-manning process as a guided exercise. It starts with choosing a specific position we disagree with, then breaks the reconstruction into four prompts: isolating the core claim, identifying the strongest evidence, understanding the values driving the position, and assembling all of that into the complete argument. From there, the worksheet helps us separate the parts of the opposing argument we accept from the parts we don’t, so that by the end, we’ve moved from a vague sense of disagreement to a precise understanding of what we believe and why. The final step turns the same lens on our own view. Each stage includes writing space and prompts designed to make the process concrete rather than abstract.

Card Catalog Classroom Steel Manning Worksheet

141KB ∙ PDF file

Download

Download

The Compounding Effect

Each of the steps above works on its own, applied to a single argument in a single moment. But the practice compounds. Every time we steel-man an opposing argument, we get a little faster at finding the strongest version rather than reacting to the surface-level version, and a little more comfortable sitting with a position we disagree with long enough to understand it on its own terms. The positions we arrive at on the other side of that process get progressively more refined, because each one has been tested against stronger opposition than the last.

Over time, steel-manning changes the default way we approach disagreement. Instead of hearing an opposing argument and immediately looking for the flaw, we develop a habit of asking, “What could the strongest version of this look like?” That shift is small, but its effects are wide. Conversations become more productive because we’re engaging with what people mean rather than the worst interpretation of what they said. Disagreements get more focused because we’ve already identified exactly where we diverge, rather than arguing in circles around a vague sense that the other side is wrong. And, hopefully, the people we’re engaging with are more likely to hear our counterarguments because we’ve demonstrated that we heard theirs first.

There’s also something that changes about how we hold our own positions. After practicing steel-manning for a while, we tend to hold our views with more precision and less defensiveness, because we’ve already reckoned with the best arguments against those views. We’ve developed a clear sense of where we stand and the specific reasons we stand there, because we’ve located the exact point of divergence between our position and its strongest opposition. That’s a different kind of confidence than the kind that comes from never having engaged with a serious counterargument. It’s more durable, because it’s already been stress-tested.

We live in an information environment where some of the hardest arguments to evaluate aren’t the ones built on lies. They’re the ones built on carefully selected truths, arranged with precision to produce a specific emotional and intellectual response. Checking whether the individual facts are real is a good and necessary starting point. But the skill that protects us when the facts are real, when every citation checks out and the conclusion still leads us somewhere misleading, is the ability to engage with why the argument is persuasive in the first place. Steel-manning builds that ability. And unlike a single act of fact-checking, which addresses one claim at a time, the habit of engaging with the strongest version of opposing arguments reshapes how we think about every argument we encounter going forward.

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

This week:

• Researchers across India, Indonesia, and parts of Africa and Latin America built working AI systems that run offline on hardware that costs less than $50.

• The New Yorker published an 18-month investigation into Sam Altman built on hundreds of pages of internal documents and more than 100 interviews.

• NVIDIA’s CEO published a framework describing AI as a five-layer industrial system that starts at the power plant and ends at the chatbot.

Countries priced out of frontier AI are building their own models on $50 hardware

What Happened

U.S. and Chinese companies control more than 90% of the world’s AI data centers, and Africa and South America have almost none. In the countries and communities left out of that concentration, a counter-movement is forming. Rest of World reported on a growing approach called “frugal AI,” in which researchers and startups build smaller, cheaper models purpose-built for specific tasks, lean enough to run on low-cost devices with limited or no internet connectivity.

One project captures what this looks like in practice. A team from the Frugal AI Hub at Cambridge University and the Indian Institute of Information Technology built a speech AI system for the Soliga, an Indigenous community in southern India whose language has no written form and a shrinking number of speakers. Commercial AI wasn’t an option because the community had neither internet infrastructure nor a market large enough to attract a major AI company. So the team recorded five hours of spoken Soliga and used that data to build a custom text-to-speech model that runs on a device costing less than $50. The voice data stayed on local hardware controlled by the community, and the project was designed from the start as a replicable template for Indigenous language preservation worldwide.

Why It Matters

AI is rapidly becoming infrastructure that economies and governments depend on for essential services. When that infrastructure is controlled by a small number of foreign companies, the countries that rely on it face a form of technological dependency with few precedents outside of energy. Frugal AI changes what’s possible within those constraints, as populations that commercial AI companies have no financial incentive to serve can get tools built around their specific needs. If a functional speech preservation system can be built for under $50 with no internet connection, the assumption that useful AI requires billion-dollar infrastructure starts to weaken.

What It Means for Us

The AI tools most of us use were built by a small number of companies for specific markets. That shapes what languages they handle well and what they miss entirely, and those gaps are business decisions as much as technical ones. Frugal AI makes those decisions more visible by showing what gets built when the starting point is a completely different set of needs. It also challenges a prevailing assumption: that useful AI requires billions of dollars and massive infrastructure. If functional AI can be built for under $50, the question of who gets to participate in building this technology is more open than the current landscape suggests.

Bottom Line: The frugal AI movement is growing at a moment when AI is becoming essential infrastructure worldwide. If these models continue to mature, countries and communities currently dependent on a handful of foreign companies for AI access will have alternatives that don’t require that dependency. The long-term balance of who controls AI may be shaped as much by these builders as by the companies dominating the landscape today.

An 18-month New Yorker investigation found that OpenAI’s internal records tell a different story than its public commitments to safety

Read more

An architect sketching a building asks AI to generate variations on her initial concept. Within minutes she’s looking at dozens of options, a volume of exploration that would have taken days to produce by hand. She scans them, recognizing which ones are worth pursuing, selects elements from several, combines them, iterates again. The final design is hers, and it also emerged from a process that didn’t exist five years ago. Which parts came from her mind and which from the interaction?

A lawyer preparing a brief asks AI to find weaknesses in her argument. In seemingly no time at all, she’s looking at a list of potential vulnerabilities, lines of attack she’ll need to address before opposing counsel finds them. She works through each one, discarding some as unlikely, recognizing others as real problems. The third vulnerability leads her to a case she hadn’t considered, which ends up strengthening her central claim. Which weaknesses would she have found on her own? Which would have surfaced only in court, diminishing the preparation of her case?

In both cases, the thinking was distributed across the exchange. The architect and the lawyer provided direction and judgment; the AI provided speed and volume. The middle of the process, where possibilities became outcomes, didn’t happen entirely inside their heads or entirely inside the machine. Neither retrieved a finished product from the AI, and neither arrived there independently. The final work emerged in the space between, through iteration neither could have conducted alone. The boundary between “my thinking” and “assisted thinking” has become harder to locate.

How Tools Have Always Shaped Cognition

Tools have always extended human cognitive capacity, but with a consistent boundary: the human remained the locus of understanding. A calculator performs a computation that the human frames and interprets. A telescope magnifies light that the human has already aimed at something meaningful. Each tool handles a defined operation, while the human handles everything that makes the operation matter. The thinking had a traceable address.

AI crosses that boundary. Large language models process language, recognize patterns, and generate synthesis in ways that functionally replicate cognitive activities previously requiring human minds. A spreadsheet doesn’t make an argument; it calculates one that the human has already framed. A search engine doesn’t analyze a question; it retrieves documents that might be relevant. But a researcher using AI can explore multiple analytical approaches simultaneously, testing hypotheses at speeds impossible for unassisted cognition, and what comes back isn’t raw data to be interpreted but reasoning to be evaluated. That’s a different kind of tool. It engages with meaning, produces language that responds to context, and generates the kind of output that used to require a mind because it involved judgment. The question of where the thinking happens has become difficult to answer, and that difficulty is new. Previous tools never raised it, because they never could.

The Brain That Reorganizes

The brain reorganizes throughout life, forming new neural connections in response to what we learn and how we spend our attention. This isn’t metaphor; it’s measurable. London taxi drivers, who spend years memorizing their city’s labyrinthine street layout, show measurably enlarged posterior hippocampi, the region involved in spatial navigation, compared with people who don’t drive cabs. When drivers retired and stopped navigating, those differences reversed over time. The capacity grew because it was being used, and receded when it no longer was. Cognition isn’t a fixed endowment that tools either help or hinder. It’s a living system that physically restructures itself around whatever we repeatedly ask it to do. Which means that when we change how we manage and process information at scale, we don’t just change our habits. We change our minds.

The history of writing is the clearest example we have of this at civilizational scale.

Before writing existed, language itself had no external storage system. Objects and markings could record quantities and sequences, but they couldn’t hold an argument, preserve a narrative, or transmit the kind of knowledge that lives in words. For everything that required language, memory was the only vessel available. The oral poets who performed the Iliad and Odyssey weren’t working from fixed texts; there were none at the time. According to the oral-formulaic theory developed by scholars Milman Parry and Albert Lord, they composed in performance, drawing on a deep repertoire of stock phrases, epithets, and scene structures that could be combined and recombined as the narrative demanded. The poem itself wasn’t stored anywhere, rather it lived in the poet’s capacity for reconstruction. Knowledge lived in minds, which meant the most valuable cognitive capacity was the ability to hold it there. The wise person was the one with deep memory, who could summon the past into the present through speech.

Then writing arrived, and knowledge could suddenly exist outside of anyone’s head. Ideas became objects, storable and examinable at a distance from the moment of their making. Socrates warned, in Plato’s Phaedrus, that writing would weaken memory and create the illusion of knowledge without its substance. He was partially right: the capacity to hold vast bodies of knowledge in mind did fade across literate populations. But writing also enabled forms of thought that oral culture couldn’t support. Systematic philosophy became possible, as did arguments that could be examined and revised long after their first speaking, logic tested across time rather than in the heat of live debate. Science as a cumulative enterprise became possible, because observations could be stored and built upon across lifetimes.

Reading, though, is something the brain was never built for. Written language has only existed for around 5,000 years, nowhere near long enough to have shaped human evolution. The brain had to improvise. Brain imaging studies by neuroscientist Stanislas Dehaene and colleagues show what that improvisation looked like: learning to read causes a specific region in the left occipito-temporal cortex to become dedicated to recognizing written words. In people who’ve never learned to read, that same brain area responds primarily to faces and objects. The region exists in everyone; what changes with literacy is what it does. As it specializes for written words, face-processing in that area decreases, with face responses gradually shifting toward the right hemisphere. The written word gets processed in the same neural neighborhood that once handled faces. Literacy repurposes real estate that evolution assigned to something else entirely. What this demonstrates, concretely, is that the brain doesn’t just use cognitive technologies - it reorganizes around them. The tool changes the organ.

The printing press, around six centuries ago, produced a further reorganization. Before printing, books were rare and expensive; reading was primarily oral and communal. Printed books enabled solitary reading on a mass scale. We could sit alone in silence for hours, following a single line of argument without interruption. The ability to follow a complex argument across hundreds of pages while holding its structure in mind isn’t innate; it’s a trained capability that emerged from generations of interaction with a new cognitive technology.

Digital technologies continued the pattern. Research by psychologist Betsy Sparrow and colleagues showed that when people expect to have access to information later, they’re less likely to encode it in memory. We stopped memorizing phone numbers and directions; we started remembering where to find information, letting location substitute for the information itself. Neural systems optimize for the environment they operate in.

Each transition produced the same dynamic: new cognitive tools changed how we think, with old capacities fading and new ones emerging. The people living through those transitions rarely saw clearly what was being gained to replace what was being lost. The new capacities only became visible after they’d already taken hold.

The Co-Evolution Underway

There’s a pattern running through all of this. Humans develop tools to extend what cognition can do. Those tools change what kinds of cognition are in demand. Brains then reorganize around what’s in demand, freeing up resources for the next layer of complexity. Writing stored memory externally, which freed cognitive resources for more abstract thought. Print multiplied books, but its deeper effect was to train the sustained linear attention that makes extended analytical reasoning possible. Each technology cleared cognitive space by offloading something the brain no longer had to carry alone.

AI might represent a more fundamental version of this dynamic. Previous technologies extended specific cognitive capabilities: writing extended memory, calculators extended arithmetic. AI extends something more central to what makes human cognition distinctive: the capacity for language-based reasoning itself. When a tool extends arithmetic, the cognitive reorganization is limited. When a tool extends language-based reasoning, the reorganization could be more fundamental.

Cognitive scientists and philosophers of mind exploring human-tool relationships have raised the possibility of cognitive co-evolution, the idea that humans may develop new ways of thinking that only function in partnership with AI. A literate person doesn’t simply have better memory than an oral poet; she has a different relationship to knowledge and a different way of building arguments. That same kind of qualitative shift may characterize what’s developing now.

These capabilities will likely emerge through practice and environmental pressure, becoming visible only after neural reorganization has already occurred in populations extensively using these tools. We didn’t design the capacity for sustained linear reading; it emerged from generations of interaction with printed books. The cognitive capacities that emerge from AI interaction are likely to surprise us in similar ways.

The Practices Taking Shape

We’re early in this transition. But what we can observe, drawing on what every previous cognitive shift has taught us, is where the reorganization is likely happening: in the specific practices through which people are engaging with these tools daily. Practices become habits, habits shape neural pathways, and neural pathways define what cognition can do. This is where new capacities form, not in intention or prediction, but in the accumulated texture of how people are actually using AI right now. And several of those practices map onto something familiar: the same skills that have always been required to navigate information environments where the source is powerful, fast, and not inherently accountable to our specific needs.

Evaluating AI outputs against our own thinking.

AI-generated text often sounds plausible without being right. Distinguishing "this sounds good" from "this captures what I mean" requires having done enough prior thinking to know what we mean. The architect scanning dozens of variations knows instantly which ones are worth pursuing because she has a clear internal sense of what the building needs to accomplish. This is source evaluation at speed: assessing whether something answers our actual question or merely appears to.

Using AI outputs as starting points.

AI generates material to react against and refine. The doctor who uses AI-suggested diagnoses to surface possibilities, then integrates them with her direct knowledge of the patient, is treating AI as a collaborator in her thinking process. The sources inform and provoke thinking without supplanting analysis.

Preserving the cognitive work that matters.

The same tool, used at different points in a process, produces different outcomes. A student who works through a problem herself and then uses AI to check her reasoning comes away with understanding she built; a student who asks AI to solve the problem first comes away with an answer she didn’t earn. What changes isn’t the tool. What changes is whether she engaged with the problem before reaching for help.

Owning the conclusion.

Working with AI involves a continuous stream of decisions about what to accept, what to revise, and what to reject. Those decisions compound. A lawyer who uses AI to pressure-test her argument and accepts each suggested vulnerability without fully evaluating it may end up with a brief that hangs together on the surface, but doesn't reflect her actual read of the case. What develops through this kind of work is something more than familiarity with a tool: a capacity to let AI expand what we can see while keeping our own judgment as the thing that decides what to do with that expanded view. The conclusion we reach should be more developed than where we started, and fully ours to stand behind.

These practices are where the pressure is being applied. Like every previous cognitive technology, what emerges from them will be shaped by what people actually do, not by what anyone designed or predicted. The underlying capacities, whatever neural reorganization is underway, will only become clearly visible later. But we can already speculate, based on where the friction is, about what those capacities might look like.

What Might Be Emerging

The historical pattern offers one guide: the capacity that emerges often looks different from the practice that produced it. Sustained engagement with printed books didn’t produce “better reading.” It produced the capacity for extended linear argument, for following a single thread of thought across hundreds of pages while holding its structure in mind. The practice was reading; the capacity that emerged was something far broader.

If the pattern holds, the capacities emerging from AI practices won’t simply be “better evaluation” or “better synthesis.” They’ll be reconfigurations we can only name after they’ve already taken shape. Some possibilities, based on where the pressure is currently being applied:

Sharper discernment. The practice of evaluating AI outputs at high volume may train a faster, more reliable sense of what's true and useful. AI outputs can be fluent and wrong simultaneously, which means the person using AI must supply the skepticism, detecting subtle misalignments between what sounds right and what is right. People who do this repeatedly are exercising judgment hundreds of times over the course of their AI activity. That capacity would transfer far beyond AI: a more finely tuned ability to assess any claim or argument, regardless of how smoothly it's presented.

Thinking across larger realms of possibility. The architect scanning dozens of variations isn't just working faster; she's navigating a landscape of options that sequential thought couldn't produce. We've had tools that extended how far we could see or how much we could calculate, but we haven't had tools that extended how many possibilities we could hold in view at once. The capacity developing here might be a new form of creative cognition: generating breadth through AI, then applying judgment to traverse that space.

Dialogue with externalized reasoning. Writing let us see our thoughts on a page. AI lets us see our reasoning reflected back and transformed. We can propose an argument, watch it restated in different terms, push back, and refine through exchange. This is different from revising a piece of work alone or debating another person. The AI doesn't have stakes or defensive reflexes. The capacity developing here might be a new form of thinking-through-conversation, where the friction of another person’s ego, defensiveness, or competing agenda is removed entirely, leaving the reasoning itself as the only thing to contend with.

Rapid iteration on complex work. The longer it takes to test an approach, the fewer approaches we test. When a methodology takes weeks to set up, the setup itself is the commitment. When the same exploration takes minutes, we can afford to start before we're certain, treating early attempts as information rather than output. We can be wrong faster, learn from that, and try again. The capacity developing here might be a more exploratory relationship to complex problems: generating and testing rather than planning a single route before we've tried any others.

Gains and losses have always run together through these transitions. The oral poet’s capacity for prodigious memorization faded with writing, but writing unlocked the systematic accumulation of knowledge across generations. The deep familiarity built from returning, again and again, to a small number of texts diminished with print, but print made new forms of individual reasoning possible. Digital tools displaced the spatial memory we once used to navigate, but opened access to information at a scale no prior generation could reach. Each technology made certain capacities less necessary; as those capacities faded, they were replaced by something the previous era couldn’t have named. In none of these transitions did the losses simply accumulate. Something was consistently freed up, and what filled that space reshaped what human minds could do next.

What We Can Shape

The brain will reorganize around what we repeatedly do. Neural pathways form around whatever practices become habitual, whether we're paying attention to that process or not. This has been true through every cognitive transition in the historical record. The monks who copied manuscripts by hand developed forms of close attention that print eventually made less culturally necessary. The urban residents who gave up navigating by mental map and switched to GPS showed measurable changes in how they engaged spatial reasoning over time. The technology shaped cognition not through a single dramatic rupture but through the slow accumulation of what people did every day. What was practiced persisted; what was no longer practiced faded.

Passive adaptation and active participation can feel similar from the inside, especially when the tools are useful and the change is gradual. What separates them is whether we’re bringing our own thinking to the exchange or simply receiving what the exchange produces, and that distinction matters more with AI than it did with previous information sources. A book or article is fixed; we can evaluate it against external criteria and check it against other sources. An AI source responds to us, calibrating itself to sound credible and helpful, which makes the old critical distance harder to maintain. The evaluative work has to come from inside the exchange itself: having thought enough about a question that we can recognize a good answer when we see one, and notice when we’re being led somewhere we didn’t intend to go. An architect scanning AI-generated variations and a lawyer pressure-testing her argument with AI are both doing exactly that, bringing clear intent to the interaction and evaluating every output against what they actually needed. The boundary between their thinking and the AI’s remained blurry, but the judgment about which outputs mattered was entirely theirs.

Every cognitive technology in the historical record produced capacities that exceeded what anyone anticipated, because the technology changed not just what people could do but how they thought about doing it. We’re inside that process now, which means we can’t see its full shape any more than the first literate Greeks could see what reading would eventually make possible. What we can do is engage deliberately, bringing real questions to these exchanges and holding our own reasoning as the standard against which we measure what comes back. The capacities that emerge from this transition will be shaped by exactly that: not by the tools themselves, but by the quality of thinking we bring to using them.

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog

The internet changed what “research” looks like in practice. Before search engines, researching a topic meant going to a library, navigating a catalog system, consulting a reference librarian, and working through physical sources that had been vetted before they reached the shelf. The process had friction built into it: we had to think about what we were looking for before we started looking, and the sources available to us had already passed through editorial and institutional filters.

Search engines removed that friction. Today, “I did my research” usually means: I typed keywords into Google, I scrolled through the results, I read the ones that matched what I was already thinking, and then I stopped looking. That process feels like research because it involves the same raw materials: reading, time, and the accumulating sense of being informed. But search engines sort results using algorithms that factor in our previous search history and click behavior alongside relevance, which means the information we encounter is already shaped by patterns we may not be aware of. In 1960, the cognitive psychologist Peter Wason ran experiments demonstrating a related pattern: given a hypothesis, people consistently seek information that supports the hypothesis and avoid information that challenges the hypothesis. Wason later coined the term confirmation bias to describe the tendency, and subsequent research has replicated the finding across a wide range of demographics, professions, education levels, and political affiliations. The internet didn’t create confirmation bias, but the architecture of search engines can make the bias harder to detect in our own reading.

AI has added another layer. When we ask an AI chatbot a question, the answer arrives in complete, fluent paragraphs that read like they were written by someone who knows the subject. The format itself communicates authority. But the answer may contain fabricated citations and confident assertions built on training data rather than verified evidence, and nothing in the tone or structure of the response signals which parts are reliable and which parts aren’t. “I asked the AI and it said…” has joined “I Googled it and found…” as something that feels like research but operates without any of the checks that make research trustworthy.

Meanwhile, the phrase “do your own research” drifted from advice into accusation, used less as a suggestion to investigate and more as a dare to arrive at a predetermined conclusion. The phrase got so loaded that hearing it now tends to shut conversations down rather than open them up. The methodology underneath the phrase, though, predates the distortion. The process of evaluating information has been studied and refined for decades, and the current version accounts for a landscape that includes AI-generated text and claims that travel faster than their sources. It operates in two distinct modes, depending on what we’re trying to do.

Two Modes of Research

When we say “research,” we’re usually describing one of two activities. Verification is what we’re doing when we encounter a specific claim and want to know if the assertion is accurate: did this event happen the way it was described, does this study say what the headline says it says, is this statistic real? Exploration is what we’re doing when we’re curious about a subject and want to build understanding from the ground up: we just got a diagnosis and want to learn how treatment options compare, or we want to understand what large language models are doing when they generate text. These two activities require different steps and different tools, and they end at different points.

Mode 1: Verification

Verification starts with a specific assertion: a statistic shared on social media, or an AI chatbot returning an answer with citations that look authoritative. Something has already been stated as fact, and the question is whether the assertion holds up under examination.

This is the mode where confirmation bias does its most immediate damage. We encounter a claim that either aligns with or challenges something we already believe, and the temptation is to evaluate it based on how it makes us feel rather than what the evidence shows. Agreement makes us credulous; disagreement makes us dismissive. Both responses skip the process entirely. The steps below replace that reflexive response with a structured one: isolate, trace, evaluate, check, and follow the funding.

Step 1: Isolate the Claim

Claims almost never travel alone. A single social media post or news headline often bundles multiple assertions together so tightly they feel like one statement. AI chatbot responses do the same, weaving conclusions into a single paragraph that reads as though the entire thing is either true or false when the reality is usually more granular than that. The bundling is what makes unsupported claims so easy to accept: when a true claim is packaged alongside a false one, the true claim lends credibility to the false one by proximity.

Take a sentence like: “The economy is terrible because of [policy X], which was designed to benefit [group Y] at the expense of everyone else.” That contains at least four distinct claims: that the economy is performing poorly by some measure, that a specific policy caused the poor performance, that the policy was intentionally designed with that effect in mind, and that one group benefits while others lose. Each of those requires different evidence to evaluate, and each one could be independently true or false, with important caveats that got dropped somewhere along the way. The sentence feels like a single assertion because it arrives as a single breath, but treating it as one researchable question guarantees a muddy answer.

The first step is to pull the claims apart. Write each one down as its own sentence, in the most direct language possible. “Is the economy performing poorly by measurable indicators?” is a researchable question with a findable answer. “Is the economy terrible?” is a feeling, and no amount of searching will resolve it because the question was never precise enough to answer.

This applies to AI outputs too. When an AI chatbot returns an answer, that answer often weaves together sourced facts and unsourced inferences, occasionally citing papers that don’t exist, all in a single fluent paragraph. Isolating each claim individually is the only way to figure out which parts hold up and which ones don’t.

Step 2: Find the Original Source

Every claim we encounter has traveled some distance from where it started. A friend texts a screenshot of a tweet that quotes a news article that cites a study that analyzed data from a government agency. Each handoff in that chain is an opportunity for distortion: context stripped away, numbers rounded in a convenient direction, conclusions inflated, limitations dropped.

The work here is to trace the claim all the way back to its origin. When someone cites a study, that means finding the study itself rather than reading the blog post that summarized the study. The same logic applies to government data (find the dataset, not the journalist’s characterization of the numbers) and to quotes attributed to public figures (find the full transcript, not the pull quote). Every layer between us and the primary source is a layer where context can be lost or distorted.

This step catches AI-specific problems too. Large language models sometimes generate citations that look real but don’t exist, complete with a convincing author name and a journal title that sounds right. The only way to catch a fabricated citation is to search for the actual paper.

A 2018 study published in Science by researchers at MIT found that false stories on what was then Twitter reached 1,500 people about six times faster than true ones. Many of those false stories weren’t fabricated from scratch; they were real information that had been stripped of context and oversimplified through successive layers of sharing. A study finding a correlation between two variables became a headline declaring that one causes the other. A dataset covering one city became evidence about the entire country. That gap between the source and the claim is where distortion accumulates.

Step 3: Evaluate the Source

Finding the original source is the start, not the finish. We also need to assess whether that source is reliable. Library science has a framework for this called the SIFT method, developed by Mike Caulfield. Each of its steps is designed to answer a different question about the source before we invest time engaging with the content itself.

SIFT stands for four moves:

Stop

Before reading further or sharing, pause. We’re at our most susceptible to misinformation in the first few seconds of encountering a claim, especially when the claim triggers a strong emotional reaction (outrage, vindication, fear, triumph). The claims that feel most urgently shareable are the ones that most need checking first.

Investigate the source

Before engaging with the claim, spend thirty seconds on the source itself. Who published this, and what’s their track record? A quick search on the organization or author tells us whether we’re looking at a peer-reviewed journal, a think tank with a disclosed ideological mission, a news outlet with published editorial standards, or an anonymous account. The standard of evidence we should expect shifts depending on which one we’re looking at, and knowing that before we start reading changes how we read.

Find better coverage

If the claim is newsworthy, check how other outlets are covering the same story. If only one source is reporting something major, that absence of corroboration is itself a data point. When multiple credible outlets have picked the story up, reading two or three of them reveals where coverage agrees and where it diverges. The places of divergence tend to be exactly the context that got left out of the version that reached us first.

Trace claims to their original context

This loops back to step two. Quotes are routinely pulled from contexts that reverse their meaning. Statistics are cited without the methodology that produced them. Headlines are written by editors, not necessarily by the researchers whose work they describe, and the distance between a study’s findings and the headline about those findings can be vast.

Step 4: Check the Counter-Evidence

Once we’ve found a credible source that supports a claim, the instinct is to stop looking. That impulse is confirmation bias at work, and overriding it changes the quality of the conclusion we reach. Finding a credible supporting source feels like the end of the process: we went looking, we found something reliable, and ostensibly we’re done. But a single credible source in favor of a claim doesn’t tell us whether equally credible sources exist that complicate or contradict the claim.

Checking counter-evidence means deliberately searching for credible sources that challenge the claim. If we’ve found a study that supports a position, we look for studies with different findings and assess the quality of their methodology. Google Scholar’s “Cited by” feature is useful here: if a study has been cited 400 times, some of those citing papers will be critical responses, and those responses often contain specific context about the original study’s limitations that the study itself doesn’t address.

This isn’t about pretending every claim has an equally valid counter-claim, or that the truth always splits the difference. Sometimes the evidence overwhelmingly supports one position. Sometimes the counter-arguments are weak or poorly sourced. But we can only know that after we’ve looked.

Before searching, ask: What evidence would change my mind?

If the answer is “nothing would change my mind,” we’re not investigating; we’re building a case. That distinction determines whether everything that follows is research or confirmation.

A useful exercise is to write down, before we start searching, what a credible counter-argument would look like: what kind of source, what kind of data, what kind of finding. If the counter-evidence turns out to be weaker than what we’ve already found, we’ve strengthened our position by testing the position rather than just defending the position. If the counter-evidence is strong, we’ve learned something that saves us from building a conclusion on incomplete ground.

Step 5: Follow the Funding

Every source of information was created by someone with reasons for creating the work. A pharmaceutical company funding a study of its own drug isn’t committing fraud, but the financial interest in a favorable result means we should look for independent replications. A think tank producing a policy analysis may have done rigorous work; the think tank’s “Funding” or “Supporters” page, plus a search of its name on the Media Bias/Fact Check database, tells us who’s paying for that rigor and what their interests are.

The same principle applies to AI companies. When an AI tool provides an answer, that answer was shaped by training data and business incentives. An AI chatbot built by a company that also sells advertising has different structural incentives than one that doesn’t. These aren’t reasons to reject AI tools, but they’re context we need in order to evaluate the outputs.

Verification in Practice: “HRT Causes Breast Cancer”

The steps above are easier to understand in motion than in the abstract, and medical claims are a useful place to see them work because the distance between what a study actually found and what most people end up hearing about the study is often enormous. Hormone replacement therapy is one of the starkest examples. In 2002, a major clinical trial made national headlines, and the four-word version of the findings that spread fastest was “HRT causes breast cancer.” Millions of women stopped treatment or never started. The research behind those four words was considerably more complicated, and that same four-word claim is still circulating today.

Imagine a perimenopausal woman encounters the claim in an online discussion. She’s experiencing symptoms and considering HRT treatment, and the assertion lands with personal weight. Working through the verification steps changes what she walks away with.

Isolate the claim. The sentence bundles several assertions: that all forms of HRT increase breast cancer risk, and that the risk applies to all women regardless of age or health profile. Each of those is a separate question.

Find the original source. The claim traces back to the Women’s Health Initiative, a large government-funded study whose results made headlines in 2002. Finding the original paper (searching "Women's Health Initiative hormone therapy" on any academic database or even a general search engine pulls it up) reveals details the headlines left out: the study tested one specific hormone combination in women whose average age was 63, not women in their forties or fifties approaching menopause. A separate arm of the same study, which tested a different hormone formulation, did not find the same breast cancer increase. The blanket claim “HRT causes breast cancer” collapses these distinctions into a single sentence that doesn’t reflect what the research found.

Evaluate the source. The WHI was funded by the National Heart, Lung, and Blood Institute, part of the NIH, though the study drugs were supplied by Wyeth Ayerst (a pharmaceutical company with a commercial interest in the hormones being tested). JAMA is a peer-reviewed journal with rigorous editorial standards. The study itself is credible, and the conflict-of-interest context (government-funded trial, industry-supplied drugs) is disclosed in the published papers. The distortion happened in how the results were communicated: the initial press coverage reported relative risk increases without absolute risk context, and generalized findings from one specific hormone combination in older women to all hormone therapy in all women.

Check the counter-evidence. Searching for reanalyses and follow-up studies on the original trial surfaces a more complicated picture than the 2002 headlines suggested. Subsequent research found that the age at which women started hormone therapy mattered, and that different formulations carried different risks than the one tested in the original study. The Menopause Society publishes regularly updated position statements that synthesize the evolving evidence, and those statements draw distinctions the original headlines erased.

Follow the funding. The original study was government-funded through the NIH, with no commercial sponsor driving the research question. The counter-narrative that emerged after 2002, though, often came from sources with financial ties to the alternatives they were promoting. Funding doesn't automatically discredit a source, but it tells us what questions to ask about the source's incentives, and those questions apply to every link in the chain between a study and the version of the study that reaches us.

Mode 2: Exploration

Exploration starts with curiosity rather than a claim. We’ve been following a policy debate and realize we don’t have enough background to evaluate what either side is saying, or we want to understand how a new technology works before deciding whether to bring it into our lives. The starting point in each of these cases isn’t an assertion to check; it’s a gap in our own knowledge that we want to close.

This mode has its own failure patterns. The most common is going straight to a search engine and reading whatever comes up first, without any sense of the subject’s landscape: which sources are considered authoritative, and how deep we need to go for our actual need. The result is often a scattered collection of articles and posts that feel informative but don’t build toward coherent understanding. We end up with fragments rather than a framework. The steps below are adapted from the way reference librarians approach an open-ended research question.

Step 1: Define the Question Behind the Question

When someone walks into a library and says “I need information about cancer,” a reference librarian doesn’t hand over a stack of oncology textbooks. The librarian asks a series of clarifying questions: What prompted this? Are we preparing for a conversation with a doctor, or trying to understand a diagnosis on our own terms? How much do we already know? What kind of answer would be useful?

This process, called the reference interview, is the foundation of research-as-exploration, and we can run it on ourselves. Someone who just received a diagnosis and someone writing a college paper on the same disease need completely different sources, organized around completely different questions. Defining the real question before we start searching keeps us from drowning in information that’s technically relevant but practically useless.

Four questions to ask before we open a search bar:

• What am I trying to understand? (Not “what’s my topic” but “what would I need to know in order to feel oriented?”)

• What do I already know, and where did that knowledge come from?

• What kind of source would help me here: an overview, a primary document, a textbook, a practitioner’s perspective?

• How will I know when I have enough to move forward?

Step 2: Start with Orientation, Not Depth

The instinct when we’re curious about something is to search for the specific thing we’re curious about and start reading whatever comes up first. The problem is that without a map of the subject, we can’t tell whether the first thing we’re reading is representative, fringe, outdated, or pitched at the wrong level. A first-page Google result about a medical condition might be a peer-reviewed overview from a major hospital system or a blog post by someone selling supplements, and without orientation, both of them land with about the same weight.

Before going deep on any one source, we need a bird’s-eye view of the landscape. Wikipedia is good for this, and so are the “Background” or “Introduction” sections of academic review articles, which summarize the state of a field before presenting new findings. A well-written Wikipedia article on a medical condition, for example, will include sections on causes, diagnosis, treatment, prognosis, epidemiology, and further reading. Spending some time with that overview before searching for specific treatment options means we’ll know what questions to ask and which results to skip.

AI chatbots can serve a similar orientation function, with a critical caveat: the overview they provide may contain fabricated details, and the confident tone of the response offers no indication of which parts are accurate. Using an AI-generated overview as a starting map is fine; treating it as settled knowledge without cross-referencing is where the risk lives. A useful practice is to ask the chatbot for its sources, then verify that those sources exist and say what the chatbot claims they say, because that verification step catches the most common failure mode: an answer that sounds authoritative but is built on citations that were generated rather than retrieved.

Step 3: Seek Multiple Perspectives at the Right Level

Once we have an overview, we can start going deeper, and the key here is matching our sources to our actual need. If we’re trying to understand how a medication works, a patient-facing resource from a major medical institution (Mayo Clinic, MedlinePlus) will serve us better than a primary research paper written for specialists. If we need the primary research, PubMed and Google Scholar will surface the papers, and the abstract and conclusion sections are usually written accessibly enough to extract the key findings without needing to parse the full methodology.

We’re looking for multiple sources that approach the subject from different angles. A textbook gives us established consensus, while a recent review article tells us where the field currently stands and what’s still being debated. Primary sources (the actual dataset, the original court filing or legislation text) tell us what happened before anyone interpreted it for us. Each of these adds a different layer, and the picture gets sharper with each one.

Public libraries offer free access to databases that would cost hundreds or thousands of dollars per year for individual subscriptions. A library card in many public library systems unlocks access to databases like JSTOR, ProQuest, newspaper archives, and specialized reference tools. The library’s website is often the fastest path to sources that a Google search would put behind a paywall.

Step 4: Follow the Bibliographies

Every good source points to more good sources. The references section of an academic paper and the “Further Reading” list at the end of a book are curated reading lists assembled by people who’ve already surveyed the landscape. Mining them is one of the most efficient research techniques available, because someone else has already done the work of deciding which sources are worth reading.

This is also how we move from popular coverage to primary sources without needing to know in advance what to search for. A newspaper article about a new education policy will cite the policy document and the studies that informed the decision. Following those citations takes us from one journalist’s summary to the raw material, where we can form our own assessment.

Step 5: Know When We Have Enough

Open-ended research has a natural failure mode: we keep reading without ever arriving at a conclusion. The decision theorist Herbert Simon coined a useful term for the alternative in 1956: satisficing, the point at which we’ve gathered enough information to make a reasonable decision or form a grounded understanding, even if more information exists. Every research question has a threshold where additional sources start repeating what we’ve already learned. When we start seeing the same key points and the same cited studies across multiple independent sources, we’ve reached saturation for our level of need.

The threshold is different depending on the stakes. Choosing a restaurant for dinner and choosing whether to start a new medication are both research tasks, but the depth of investigation each one warrants is wildly different. Calibrating our research to the weight of the decision is itself a skill, and one that gets easier with practice.

Exploration in Practice: Understanding Hormone Therapy Options

Verification answered one question for the woman in the previous example: the four-word claim about HRT and breast cancer did not accurately represent the research. But knowing that a claim is oversimplified doesn’t tell her what to do next. She still has symptoms, she has a doctor’s appointment coming up, and understanding the treatment landscape well enough to participate in her own medical decisions requires exploration rather than verification. The question is no longer “is this specific assertion true?” but “what do I need to know in order to make a good decision for my body?”

Define the question behind the question. “Should I take hormones?” isn’t precise enough to research. A reference-interview approach breaks it into actionable questions: What are the established treatment options for perimenopausal symptoms? How do different types of hormone therapy compare in terms of effectiveness and risk? What does the current medical consensus say about who benefits and who should be cautious? What would she need to know in order to ask her doctor the right questions at her next appointment?

Start with orientation. Before searching for specific treatment options, she starts with overview sources: an encyclopedia-style article and a patient-facing medical summary from a major institution. The goal is to get a sense of the landscape before going deep on any one corner of the landscape. These are starting points, not destinations, and the map they build is what makes the more specific reading afterward productive rather than scattered.

Seek multiple perspectives at the right level. Once she has an overview, she looks for sources that approach the subject from different angles and at different levels of depth: a professional medical society’s position statement for the current consensus, and a patient-oriented summary from a major medical center for the practical implications. Reading across these reveals where the evidence is settled and where it’s still evolving, and matching the source to her actual need (preparing for a doctor’s appointment, not writing a research paper) keeps her from getting lost in material pitched at the wrong level.

Follow the bibliographies. The position statements and review articles she’s been reading cite the studies behind their conclusions. Following those citations takes her from someone else’s summary to the primary research, where she can see the evidence firsthand. Each reference list is a curated reading list assembled by someone who has already surveyed the field, and mining those lists is one of the most efficient ways to go deeper without needing to know in advance what to search for.

Know when we have enough. She’ll notice the same key findings surfacing across independent sources. The different types of sources she’s been reading will converge on the same core conclusions, even though each one emphasizes different aspects. When that convergence starts happening, she has enough to walk into her doctor’s appointment with specific questions about her own situation rather than general anxiety about a treatment category. The goal here was not necessarily to become an expert; the goal was to become informed enough to participate in the conversation.

Where the Two Modes Overlap

Verification and exploration aren’t entirely separate. An exploration that starts with curiosity will surface specific claims that need verifying along the way. A verification task that starts with a single claim sometimes reveals that the underlying subject is more complex than we thought, and shifts into an exploratory mode. The methodology for each is different in emphasis but built from the same core principles: start with a clear question, seek original sources, evaluate the quality and origin of what we find, consult multiple perspectives, and stay aware of our own biases throughout the process.

Both modes are also iterative in a way that’s easy to underestimate. We may start exploring a subject, develop an initial understanding, then encounter a claim that sends us back to verification. The verification might reveal that one of our earlier sources was less reliable than we thought, which changes the map we built during exploration. Research moves in loops rather than straight lines, and comfort with that looping is part of what makes the process work. The people who do this best aren’t the ones who get every step right the first time; they’re the ones willing to revise earlier conclusions when new evidence warrants the revision.

The Skill Underneath the Steps

Every step in both modes draws on the same underlying capacity: the ability to evaluate information before acting on what we’ve found. In library science, this capacity is called information literacy, and it’s been a formal area of study and professional practice since the American Library Association first defined the term in 1989. At its core, information literacy is the ability to recognize when information is needed, to locate and evaluate it, and to use it effectively.

For most of the twentieth century, these skills were woven into physical infrastructure. The card catalog required us to think about how knowledge was organized before we began searching, and the reference librarian sitting twenty feet away could redirect a misguided search strategy before we’d wasted an afternoon on the wrong shelf. Digital search removed that structure in exchange for speed and access, and AI is accelerating the same trade-off: we get answers faster than ever, with less visibility into where those answers came from or whether they’re accurate.

The infrastructure changed, but the skills were never widely replaced. Information literacy is taught in library science programs and scattered across some K-12 and university curricula, but it has never been a standard part of public education the way reading and math are. We live in the most information-dense environment in human history, we’re told constantly to do our own research, and still the phrase demands a competency it doesn’t supply. But the skills that make sense of this landscape are not new. They predate both the internet and AI, and they work the same way now that they always have: get closer to the origin of what we’re reading, and let the evidence shape the conclusion rather than the other way around.

Have you read the Founding Member Report: The State of AI yet?

A comprehensive guide for information navigators who want to understand where AI is actually heading and what it means for how we find, evaluate, and use information in 2026.

→ Find out more here.

Share Card Catalog