What "Do Your Own Research" Actually Means

A step-by-step methodology for evaluating any claim and exploring any subject.

“Do your own research” used to be straightforward advice, the kind of thing a doctor might say to a patient weighing treatment options or a teacher might assign as the first step of a term paper. The phrase assumed a process: go find credible sources, read them carefully, weigh what they say, and come back with a more informed perspective than the one we started with. It also assumed that the person hearing the advice had access to a set of skills for doing this, or at least access to institutions (libraries, schools, professional advisors) that could guide the process.

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.

Photo by Gabriel Federa on Unsplash

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.

Photo by Ryunosuke Kikuno on Unsplash

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.

---