Confirmation bias, and the experiments it comes from

An everyday example

Someone convinced that a new habit is improving their sleep starts noticing the mornings they feel rested and mentally files them as evidence. The mornings they feel groggy get explained away: a late night, a stressful week, the weather. Nobody sat down and decided to grade the good mornings more generously than the bad ones. The noticing and the explaining both happened on the way to a conclusion that was already in place.

That is confirmation bias in its ordinary form: not a refusal to look at evidence, but a habit of looking harder for the evidence that fits, and looking away sooner from the evidence that does not.

The classic experiments

Peter Wason gave people a rule-discovery task in 1960. Participants were told that the triple of numbers 2, 4, 6 fit a rule he had in mind, and that they could test any triple they liked by asking whether it fit, then announce the rule once they were confident. Most people settled quickly on a guess such as "even numbers increasing by two," and tested it by proposing triples that fit their own guess: 8, 10, 12, or 20, 22, 24. Those tests kept coming back yes, which felt like confirmation. The actual rule, in Wason's task, was simply three numbers in increasing order, a rule the guessed one sat entirely inside. A person who never tests a triple built to break their guess, such as 1, 2, 3 or 5, 4, 12, can go on collecting confirming answers indefinitely without learning that their rule is too narrow.

A very different demonstration came from Charles Lord, Lee Ross and Mark Lepper in 1979. Students who already favoured or already opposed capital punishment, sorted by an earlier attitude survey, were each given the same two pieces of research to read: one set of findings suggesting the death penalty deters murder, and one set suggesting it does not. Both sides found the study that agreed with their existing view better designed and more convincing, and picked apart the study that disagreed on points of method. The result reported in the paper is the one that makes the case: after reading the identical, mixed set of evidence, supporters became more supportive and opponents became more opposed. The same facts had pushed both groups further apart, not closer together.

The two studies show different mechanisms. Wason's participants controlled which evidence they gathered and rarely gathered the kind that could prove them wrong. Lord, Ross and Lepper's participants did not choose their evidence at all; they were handed the same page as everyone else and still read it unevenly, crediting the half that agreed with them and discounting the half that did not.

Does it replicate?

Replication grade: Strong: reproduced across many different paradigms and settings

Klayman and Ha's 1987 analysis reframed part of Wason's result. They pointed out that testing cases you expect to fit your hypothesis, what they called a positive test strategy, is a sensible approach to most real hunting-for-a-rule problems, where the true rule and the guessed rule mostly overlap. It only misleads you in a task built like Wason's, where the true rule is broader than any reasonable guess, so every confirming test is uninformative by design. That does not undo the finding: people in Wason's task still announced wrong rules with confidence, having tested only cases built to agree with them. It changes the explanation, from "people irrationally avoid disproof" to "a normally sound strategy fails in this particular kind of task, and people rarely notice the failure."

Nickerson's 1998 review, surveying decades of studies across memory, perception, hypothesis testing and social judgement, treats confirmation bias as several related habits rather than one mechanism: biased search for evidence, biased interpretation of evidence once found, and biased recall of evidence encountered earlier. The attitude-polarization result belongs to the second category; Wason's task belongs to the first. Both have been reproduced in many forms since, across subjects as different as jury reasoning, medical diagnosis and political belief, which is why Nickerson called it one of the most consistently documented patterns in the whole of cognitive and social psychology.

A funnel, not a filter you choose

The diagram on this page is a schematic, not measured data. It draws a scattered field of evidence above a funnel and shows only some of it reaching the box below, where judgement happens. The point of drawing it as a funnel rather than a filter with a switch is that nobody usually decides to exclude the disconfirming pieces; the funnel's shape does that on its own, through which evidence gets noticed, sought out and remembered as worth keeping.

Try it: find the hidden rule

The game below is Wason's task itself. A rule about triples of numbers is hidden; 2, 4, 6 fits it. Test as many triples as you like, then try to name the rule. Watch which triples you choose to test, not just whether you get it right.

How to catch it

The fix for search bias and the fix for interpretation bias are not quite the same, so both belong on this list.

• Before you look, write down what evidence would change your mind. If you cannot name any, you are not really testing the belief.
• When you test a guess, try at least one case built to fail it, not only cases built to fit it.
• When you read something that agrees with you, ask the question you would ask if it disagreed: how big was the sample, who funded it, what would the authors have needed to find to publish it.
• Notice when you are grading two studies by different standards. If a flaw would sink the study you dislike, it should count against the one you like too.
• Ask someone who holds the opposite view what they think the strongest evidence for their side is, and read it before you argue against it.
• Treat a rule or belief that has never once come back "no" as untested, not proven.