How to Use AI to Make Us Smarter
The headlines say AI is making us dumber. The research tells a more complicated story.
In Plato’s Phaedrus, written around 370 BCE, Socrates tells a story about the invention of writing. An Egyptian god named Theuth brings his newest creation to King Thamus and calls it “an elixir of memory and wisdom” that will make the Egyptians wiser. Thamus pushes back: writing won’t strengthen memory, he argues, but weaken it, because people will stop practicing recall and rely on external marks instead. Worse, writing will produce the appearance of wisdom rather than the real thing, filling Egypt with people who’ve read widely but understood little. Socrates appears to have used the story as a mouthpiece for his own argument, dressed in the authority of Egyptian myth: in the dialogue that follows, he goes on to make his own case against writing that echoes Thamus’s critique almost exactly.
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.
Want to go further?
This article has a companion lesson inside the Card Catalog Classroom: Use AI to Sharpen Our Thinking. This expands the framework and includes worked examples and exercises that help you put it to use in your own work. Paid subscribers can access the full Classroom and everything inside.
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.
This piece highlights precisely the enormous gap I encounter all the time between intellectually mature and pre-internet educated adults (along with a very few unusual, and precocious, young people) and the vast majority of students in school, K-bacc.
The distinction between active AI use by people who already know how to conduct research, how to ask questions and pursue sustained inquiry, how to evaluate what the tools generate, and especially that asking questions and evaluating responses and results are even things humans do at all, on the one hand, and passive AI use by (currently younger) people who have never learned any of these things and for whom LLMs are simply ATMs with more buttons and steps, makes all the difference.
My older colleagues who love using AI tools and gush about all the new stuff they can do, if they are not also people who teach, especially people who teach students from backgrounds that don't bring with them a culture of reading, or researching, or reasoning, find themselves *completely unable to imagine* what it would be like to use AI e.g. for a writing project without being able to rely on what they already know, on the habits of inquiry and analysis and persistent evaluation developed over a lifetime of upbringing and schooling that forced this development, and offered only isolated shortcuts, like the calculator. They cannot get their heads around what it's like to treat every assignment like a new banking or transit app, just another thing to click on until you get the thing that will get you what you want, namely, in this case, the A. They can't see that without the experience, and the appetite, for the discovery and successful evaluation process, merely adding more steps to the assignment that *mimic* this discovery and evaluation just makes for more clicks that instruct the AI tool to itself *mimic* these processes and preserve them in the output a.k.a. the completed assignment, whether essay, or journal, or project, or poem.
The analogy to the advent of browser-based search engine internet experience actually bears this out, and distinguishes both that advent as well as the new AI tools from the emergence of writing as Plato experienced it, not least because of the staggeringly different timescales: we ourselves are the beneficiaries of a couple of millenia worth of adjustment (in the west) to the losses of oral memory culture and the benefits of writing, benefits that were and are still the subject of laborious evengelicalism in education to reach everyone in the classoom. With the browser- and search engine-internet, we're talking at most 30 years, not even, really, for general exposure and dependency. In my own experience over thirty years of teaching post-secondary students from all kinds of backgrounds I have found that we have--or had--barely begun to make the adjustment the OP here analogizes to the historical response to the advent of writing. That adjustment itself has, in my view, now been completely overrun by the arrival of LLMs and their universal adoption by the vast majority of current students.
When I hear about AI use *for students* that forces, de novo, rather than *relies on*, the development of habits of inquiry and evaluation, as well as actual, non-mimetic, knowledge, I'll be all ears. Until then, I cannot see that this tech is anything but harmful, in general use, for people coming of age intellectually from the turn of this millenium on.
Card Catalog's mission is an outstanding model and guide for thinking about these issues, and for bringing the fruits of our written millenia to an AI world--but it is crucial, in my view, to distinguish in this way between users of AI, because them has to have ears to hear.
Hana, your writing is advanced course work. Solid gold. Thank you so much for this substack. As others have mentioned, I will be referring to it often. And the readers' comments are just as valuable; like sitting in a class with graduate level thinkers. It's worth every penny of a paid subscription.