Confirmation Bias
People and AI systems both tend to search for, favor, and remember evidence that agrees with what they already believe, while letting contradicting evidence slide by unexamined. The fix is not to distrust everything. It is to go looking for the answer that would prove you wrong before you decide you are right.
Confirmation bias is the pull toward evidence that agrees with an existing belief, and away from evidence that challenges it.
What would change my mind, and did I actually go look for it?
Hiring decisions, strategy reviews, medical and legal judgment, research, investing, and checking AI-generated answers.
Fast overview
Psychologist Peter Wason first demonstrated the pattern in a series of experiments beginning in 1960. In his best known test, people were given a rule and asked which cards to turn over to check it. Most picked cards that could confirm the rule. Very few picked the one card that could actually disprove it, even though disproving it is what a real test requires. The philosopher Francis Bacon had already described the same tendency centuries earlier, writing that the human mind seizes on evidence that supports a favored idea and pushes aside whatever contradicts it.
Confirmation bias operates in three places at once: what evidence you search for, how you interpret evidence once you have it, and what you remember afterward. All three quietly point the same direction.
Mental Model Compass
You feel confident, the stakes are real, and you have not deliberately looked for the opposing case.
A search that only produces agreement, a source you chose because it usually agrees with you, or a decision already made in your head before the evidence arrived.
What is the strongest evidence against this, and where would I find it if it existed?
Treating every disagreement as proof the other side is biased, or treating your own conclusion as settled just because your search kept confirming it.
Learn it progressively
State the specific claim or decision you are about to act on, in one sentence.
Before searching, write down what evidence would prove this belief wrong, and where you would expect to find it.
Deliberately search for that disconfirming evidence, using a source or a person likely to disagree with you.
Give the disconfirming evidence the same scrutiny you gave the confirming evidence, not more skepticism.
Make the call, and write down what would change your mind later, so a future check can compare the belief against what actually happened.
See it in different settings
Hiring
An interviewer forms a strong impression in the first five minutes, then unconsciously steers later questions toward confirming it, missing signs that point the other way.
Investing
An investor who owns a stock reads mostly bullish commentary about it and skims past analyst downgrades, describing them as noise rather than information.
Medicine
A doctor settles on an initial diagnosis, then reads new test results through that lens, explaining away results that would point toward a different condition.
Working with an AI assistant
Someone asks an AI a question that already contains their preferred answer, gets a response that agrees, and stops checking, without ever asking the AI to argue the other side.
Strategy reviews
A team highlights the metrics that improved this quarter and quietly omits the ones that did not, because the metrics that improved feel like the real story.
Worked application
Deciding whether a new marketing channel is working
State the belief: "Our new referral program is driving real growth."
What would prove it wrong: Most referral signups would have happened anyway through organic search, or referral customers convert to paying customers at a lower rate than other channels.
Go get it: Pull the conversion rate for referral signups specifically, and compare it with organic growth in the two months before referrals launched.
Weigh it honestly: Signups are up, but conversion to paying customers from referrals is lower than from search. The growth in raw signups is real. The belief that referrals are the strongest channel needs revising.
Decide and record: Keep the referral program at its current spend, but hold off on expanding it until conversion improves. Revisit in one quarter using the same two numbers.
Common mistakes
Searching, but only in places that already agree with you.
Asking an AI assistant to confirm a decision instead of asking it to challenge one.
Treating a large volume of agreeing sources as strong evidence, when they may all trace back to one original source.
Looking for the counterargument only after the decision is already made, so the search becomes a formality.
Using the label "confirmation bias" to dismiss someone else's evidence instead of examining it.
When not to use it
Not every case of people agreeing is bias. Sometimes the evidence genuinely converges because a claim is well supported. Do not use this model to argue that consensus is always suspect, to justify endless second-guessing that never reaches a decision, or to accuse someone of bias simply because they disagree with your account of events. The model is a check on how evidence was searched for and weighed, not a way to declare every agreement invalid.
Related models
The Map Is Not the Territory
A belief is a map. Confirmation bias is what keeps a map from ever being checked against the territory.
Bayesian Updating
The disciplined alternative: move your confidence in proportion to genuine evidence, rather than searching only for evidence that keeps your confidence right where it already was.
Probabilistic Thinking
Treating a belief as a probability instead of a certainty makes it easier to notice evidence that should move that probability down.
Second-Order Thinking
Ask not just whether the evidence agrees with you, but what someone motivated to disagree with you would say next.
Try It Yourself
Save this as a Skill
This turns the model above into a short, reusable instruction set you can keep and paste into any AI assistant before a decision.