Your mind is not a neutral judge of evidence, and it cannot feel itself leaning. People look for what agrees with them, read it more kindly than what does not, and come to trust claims they have simply heard often. None of it feels like bias. It feels like being right. AI chat can make each habit easier to indulge. This page sets out what the research shows, where famous findings were overstated, and six checks you can run on yourself.
- Confirmation bias is mostly unwitting. You seek and weigh evidence in favour of what you already think, without deciding to.
- Wanting an answer bends the search. People believe what they want to, but only as far as they can build a case for it.
- Repetition makes claims feel true. In lab studies, even claims people already knew were false.
- AI chat can deepen the groove. In one study, people asked more one-sided questions of an AI search tool than of ordinary search.
- Some famous findings were overstated. That corrections usually backfire is one; a researcher who reported it says so.
- A routine beats good intentions. Whether knowing about these biases helps is unproven. Set questions are something you can actually do.
Confirmation bias: building the case without noticing
The standard review of the research defines confirmation bias as “the seeking or interpreting of evidence in ways that are partial to existing beliefs, expectations, or a hypothesis in hand”.Raymond S. Nickerson, Confirmation Bias: A Ubiquitous Phenomenon in Many Guises, Review of General Psychology, 1998, read at source 18 Sep 2026: “the seeking or interpreting of evidence in ways that are partial to existing beliefs, expectations, or a hypothesis in hand”.
The key idea is that you do not choose to do it. A lawyer builds a one-sided case on purpose. Confirmation bias is the version nobody notices: “It refers usually to unwitting selectivity in the acquisition and use of evidence.”Nickerson, Review of General Psychology, 1998, read at source 18 Sep 2026: “It refers usually to unwitting selectivity in the acquisition and use of evidence.”
It takes familiar forms: considering only one explanation, giving supporting evidence more weight, testing an idea only with cases that would fit it. In the review’s words, people “tend to seek information that they consider supportive of favored hypotheses or existing beliefs and to interpret information in ways that are partial to those hypotheses or beliefs.”Nickerson, Review of General Psychology, 1998, read at source 18 Sep 2026: “People tend to seek information that they consider supportive of favored hypotheses or existing beliefs and to interpret information in ways that are partial to those hypotheses or beliefs.”
The bias does not announce itself. Feeling that you weighed the evidence fairly does not show that you did.
Motivated reasoning: wanting a particular answer
Confirmation bias can happen even when you have no stake in the answer. Motivated reasoning is the version driven by wanting one. A 1990 review proposed that the wish to reach a conclusion changes which memories, rules and arguments you reach for: it “enhances use of those that are considered most likely to yield the desired conclusion.”Ziva Kunda, The Case for Motivated Reasoning, Psychological Bulletin, 1990, read at source 18 Sep 2026: “the motivation to arrive at particular conclusions enhances use of those that are considered most likely to yield the desired conclusion”.
The same paper sets a limit worth knowing: “People will come to believe what they want to believe only to the extent that reason permits.” You need a case that would look reasonable to an outsider, so you build one from a search you do not notice is selective.Kunda, Psychological Bulletin, 1990, read at source 18 Sep 2026: “People will come to believe what they want to believe only to the extent that reason permits.”
The paper is candid that the idea was contested. Critics had argued “all research purported to demonstrate motivated reasoning could be reinterpreted in entirely cognitive, nonmotivational terms” — people might reach convenient conclusions because they fit prior beliefs, not because they were wanted.Kunda, Psychological Bulletin, 1990, read at source 18 Sep 2026: “The major and most damaging criticism of the motivational view was that all research purported to demonstrate motivated reasoning could be reinterpreted in entirely cognitive, nonmotivational terms”. For the checks below it hardly matters: either way, the evidence gathered is lopsided.
The more you want an answer, the harder you should check it. The case you build will convince you first.
Repetition: why familiar feels true
“Repeated statements receive higher truth ratings than new statements, a phenomenon called the illusory truth effect.” It was first reported in 1977, and “cognitive, social, and consumer psychologists have replicated the basic effect dozens of times.”Lisa K. Fazio, Nadia M. Brashier, B. Keith Payne & Elizabeth J. Marsh, Knowledge Does Not Protect Against Illusory Truth, Journal of Experimental Psychology: General, 2015, read at source 18 Sep 2026 (a university-hosted copy of the publisher PDF): “Repeated statements receive higher truth ratings than new statements, a phenomenon called the illusory truth effect.” and “cognitive, social, and consumer psychologists have replicated the basic effect dozens of times.” The 1977 date is the paper’s citation of Hasher, Goldstein and Toppino; this site has not read that original study.
The usual explanation is ease: repetition makes a statement easier to process, “leading people to the (sometimes) false conclusion that they are more truthful”.Fazio et al., Journal of Experimental Psychology: General, 2015, read at source 18 Sep 2026: “Repetition makes statements easier to process (i.e., fluent) relative to new statements, leading people to the (sometimes) false conclusion that they are more truthful”.
It had been assumed that knowing the facts protects you. The 2015 study tested that assumption: “Contrary to prior suppositions, illusory truth effects occurred even when participants knew better.”Fazio et al., Journal of Experimental Psychology: General, 2015, read at source 18 Sep 2026: “Contrary to prior suppositions, illusory truth effects occurred even when participants knew better.”
Keep it in proportion. Each of its two main experiments used 40 undergraduates at one university.Fazio et al., 2015, read at source 18 Sep 2026, in the method section of both experiments: “Forty Duke University undergraduates participated in exchange for monetary compensation.” The authors describe the effect as small: “The illusory truth effects in Experiment 1 represent small shifts along the middle of a 6-point scale.”Fazio et al., 2015, read at source 18 Sep 2026: “The illusory truth effects in Experiment 1 represent small shifts along the middle of a 6-point scale.” A small nudge matters most for claims you meet again and again.Reasoning, September 2026 — an inference from the effect being small per exposure; the study did not measure everyday exposure.
“I have heard this before” is not “I have checked this”. Familiarity is exactly the feeling repetition produces.
Where AI chat comes in
Sycophancy — AI assistants agreeing with you more than the facts warrant — is covered on how AI influences you. This section is about what you bring to the conversation.
A 2024 study compared ordinary web search with AI chat search on controversial topics. Its summary: “Overall, we found that participants engaged in more biased information querying with LLM-powered conversational search, and an opinionated LLM reinforcing their views exacerbated this bias.”Nikhil Sharma, Q. Vera Liao & Ziang Xiao, Generative Echo Chamber? Effects of LLM-Powered Search Systems on Diverse Information Seeking, CHI 2024, arXiv 2402.05880, read at source 18 Sep 2026: “Overall, we found that participants engaged in more biased information querying with LLM-powered conversational search, and an opinionated LLM reinforcing their views exacerbated this bias.”
Two details matter. The obvious fixes did little. Showing sources, and using an AI set up to challenge people’s views, were interventions “both of which had little effect in reducing selective exposure”. And the authors limit their own claim: sessions were short, the task was writing an essay in a closed research system, and “The results may not be fully generalizable to diverse real-world information-seeking settings with search systems.”Sharma, Liao & Xiao, CHI 2024, read at source 18 Sep 2026: “both of which had little effect in reducing selective exposure” and “The results may not be fully generalizable to diverse real-world information-seeking settings with search systems.”
Put the pieces together and a loop appears. A question phrased from inside your view, an answer that tends to go along with the framing, the same point met again in the next answer, and a claim that now feels familiar. No step needs the AI to say anything false.Reasoning, September 2026 — combines the sections above with the framing effect described on how AI influences you; no study has measured this loop as a whole.
The risk is not only what the AI says. It is how you ask, and how often you let it say the same thing back to you.
Findings that did not hold up as told
Research on bias has its own reliability problem. A large team re-ran 100 published psychology studies and reported in 2015: “Ninety-seven percent of original studies had statistically significant results. Thirty-six percent of replications had statistically significant results”.Open Science Collaboration, Estimating the reproducibility of psychological science, Science, 2015, read at source 18 Sep 2026 (a university-hosted copy): “We conducted replications of 100 experimental and correlational studies published in three psychology journals” and “Ninety-seven percent of original studies had statistically significant results. Thirty-six percent of replications had statistically significant results”. That project did not test the effects on this page, but it is a reason to hold any single striking study loosely.
The clearest overstatement in this area is the backfire effect: the idea that correcting people makes them believe a falsehood more. One of the researchers who reported it in 2010 wrote in 2021 that it appeared “In two of the five studies that we conducted”, and that later coverage distorted its generality. His summary of the evidence since: “an emerging research consensus finds that corrective information is typically at least somewhat effective at increasing belief accuracy when received by respondents.” The catch is that the gains “often do not last or accumulate”.Brendan Nyhan, Why the backfire effect does not explain the durability of political misperceptions, PNAS, 2021, read at source 18 Sep 2026: “In 2010, Jason Reifler and I published an article”; “In two of the five studies that we conducted”; “an emerging research consensus finds that corrective information is typically at least somewhat effective at increasing belief accuracy when received by respondents”; “often do not last or accumulate”.
The confirmation bias review itself dates from 1998, long before the 2015 replication project, and many of the studies it summarises are small. So this page treats the broad pattern as well supported and the size of any single effect in daily life as unknown.Reasoning, September 2026 — a judgement about how much weight the sources above can carry.
Trust a finding about thinking by how often independent teams have reproduced it, not by how often it has been retold.
Six checks to run on yourself
- Before you look, say what would change your mind. If nothing would, you are collecting, not investigating. The 1998 review reports that the habit of testing only confirming cases “can be reduced if people are asked to consider alternatives, but that they tend not to consider them spontaneously”. So ask yourself on purpose.Nickerson, Review of General Psychology, 1998, read at source 18 Sep 2026, describing a study this site has not read: “the tendency to use it can be reduced if people are asked to consider alternatives, but that they tend not to consider them spontaneously”.
- Search both ways. Phrase the question the way someone who disagrees would. “Why does this work” and “does this fail” turn up different pages.
- Ask whether you checked it or just heard it. If you cannot say where you learned something, treat it as familiar, not verified.
- Ask the AI before you tell it what you think. Keep your view out of the first question, and do not count its agreement as a second opinion (why, with the evidence). Asking it to argue the other side is worth a try, but in the 2024 study an AI built to challenge people did little on its own. You still have to read the other side, not just request it.
- Go back to the original for the claims you most want to be true. Those are the ones you are least likely to check. How to read a number before you repeat it shows what opening the source catches.
- Write down how sure you are, then find one reason you could be wrong. The review notes that giving reasons against your own position “has reduced overconfidence in some instances” — but “generally overconfidence has only been reduced, not eliminated”.Nickerson, Review of General Psychology, 1998, read at source 18 Sep 2026: “has reduced overconfidence in some instances” and “generally overconfidence has only been reduced, not eliminated”.
Two checks you actually run are worth more than six you only agree with.
Related
The wider habit is on critical thinking. For sources an AI hands you, see how to check an AI citation; for claims that arrive as images, how to check a viral screenshot.
What this page could not verify
- The original 1977 illusory truth study. Its date and the count of replications come from the 2015 paper; the original was not read.
- The individual experiments inside the two reviews. The consider-alternatives and reasons-against findings are reported as the 1998 review describes them. That review says the extent to which training can modify the bias “deserves more research than it has received”.Nickerson, Review of General Psychology, 1998, read at source 18 Sep 2026: “The question of the extent to which the confirmation bias can be modified by training deserves more research than it has received.”
- How large any of these effects are in everyday life, and whether any training reliably reduces them. This site has not found a study that settles either.
- Today’s assistants. The 2024 study used a research system in short sessions.
- Publisher copies. The APA and Science sites refused automated access; both 2015 papers were read from university-hosted copies.
Raymond S. Nickerson, Confirmation Bias: A Ubiquitous Phenomenon in Many Guises, Review of General Psychology (1998) · Ziva Kunda, The Case for Motivated Reasoning, Psychological Bulletin (1990) · Fazio, Brashier, Payne & Marsh, Knowledge Does Not Protect Against Illusory Truth, Journal of Experimental Psychology: General (2015) · Sharma, Liao & Xiao, Generative Echo Chamber?, CHI (2024) · Open Science Collaboration, Estimating the reproducibility of psychological science, Science (2015) · Brendan Nyhan, Why the backfire effect does not explain the durability of political misperceptions, PNAS (2021). All read at source on 18 September 2026.