3.5 Confirmation bias

Learning objectives

• Describe confirmation bias and its three forms: biased search, biased interpretation and biased memory.

• Explain belief perseverance and attitude polarisation as consequences of confirmation bias.

• Describe Wason’s 2-4-6 task and explain what it shows about hypothesis testing.

• Explain cognitive, motivational, biological and sociocultural explanations of why confirmation bias occurs.

• Describe the aim, method, results and conclusions of Lord, Ross and Lepper (1979).

• Apply confirmation bias to decision-making in forensic science, policing, education, health and relationships.

• Evaluate research on confirmation bias with reference to measurement, bias, causality and responsibility.

• Discuss strategies that reduce confirmation bias.

Introduction

Confirmation bias is the tendency to seek, interpret and remember information in ways that support what a person already believes. It shapes decisions as different as a doctor’s diagnosis, a detective’s choice of suspect, a voter’s reading of the news and a scientist’s interpretation of data, and it helps explain why disagreements often survive exposure to the same evidence.

Confirmation bias is a cognitive bias studied within the cognitive approach, because it concerns how information is selected, processed and stored. It also connects the approaches: its causes include limits on attention and memory, the wish to protect beliefs that matter to a person’s identity, and the information environments that groups and technologies create. It is central to the concept of bias in research, since researchers are as prone to it as anyone else.

What confirmation bias is

Every day people hold expectations and hypotheses: that a new classmate is unfriendly, that a medicine works, that a particular team will win. Testing such ideas properly requires looking for evidence that could show them to be wrong, as well as evidence that fits. Confirmation bias describes a consistent imbalance in this process: information that fits the existing belief is sought more, trusted more and remembered better than information that does not.

The bias is usually unintentional. A person showing confirmation bias is not lying or deliberately ignoring evidence; they experience their conclusion as the honest result of looking at the facts. This is what makes it difficult to correct: people can see it in others far more easily than in themselves.

Confirmation bias is a bias of decision-making because decisions depend on the information that is gathered and how it is weighed. A doctor decides on a treatment, a jury on a verdict and a student on a university after weighing evidence. If the gathering and weighing are tilted towards an early belief, the decision follows the belief rather than the evidence.

Three forms: search, interpretation and memory

Confirmation bias can operate at each stage of processing information. Research usually distinguishes three forms, which often occur together and reinforce one another.

Snyder and Swann (1978) showed biased search in conversation. Participants who were asked to find out whether a person was an extravert chose questions such as what the person would do to liven up a dull party, while those testing whether the person was an introvert chose questions such as what the person found unpleasant about noisy parties. Such questions tend to draw out answers that confirm whichever hypothesis is being tested, since almost anyone can describe both kinds of situation.

Biased interpretation explains why two people can look at the same evidence and both feel more confident in opposite conclusions. Supporting evidence is accepted at face value, while opposing evidence is scrutinised until a flaw is found, and every study has some flaw.

Snyder and Cantor (1979) asked participants to read a description of a woman, Jane, who behaved in both introverted and extraverted ways. Later, some participants were asked to judge whether Jane would suit a job as a librarian and others whether she would suit a job selling property. Those judging her for the librarian job recalled more of her introverted behaviour, and those judging her for the sales job recalled more of her extraverted behaviour. The hypothesis being tested guided what was retrieved from memory.

The cycle explains why confirmation bias is self-reinforcing. A belief directs the search; the evidence found is read in its favour; the supportive parts are remembered; and the belief, now apparently well supported, directs the next search. Over time the person may hold the belief with great confidence on the basis of evidence that, taken as a whole, is weak or mixed.

Belief perseverance and attitude polarisation

Two consequences of confirmation bias are especially important for decision-making, because they explain why beliefs often outlast the evidence that created them.

Ross, Lepper and Hubbard (1975) gave participants false feedback that they had succeeded or failed at a task of telling real suicide notes from fake ones. They were then fully debriefed: told that the feedback had been decided in advance and had nothing to do with their performance. Even so, participants who had received “success” feedback continued to rate their ability higher than those who had received “failure” feedback. A more detailed “process” debriefing, which explained how beliefs persist, greatly reduced the effect. Once a belief exists, people generate their own reasons for it, and those reasons survive when the original evidence is withdrawn.

Polarisation is the group-level result of biased interpretation. If each side accepts the evidence that suits it and rejects the rest, mixed evidence gives both sides more reasons for their existing view, so the disagreement widens.

How confirmation bias is studied: Wason’s 2-4-6 task

The first experimental demonstration of confirmation bias in hypothesis testing came from Peter Wason (1960). It remains the standard illustration because it shows the bias in its purest form: participants had no stake in the answer, so any bias came from how they tested ideas rather than from wanting a particular result.

Wason told 29 university students that the sequence 2-4-6 fitted a rule he had in mind, and asked them to find the rule. They could propose their own sequences of three numbers, and he told them each time whether the sequence fitted the rule. When they felt confident, they announced the rule. The rule was simply “any three numbers in ascending order”.

Most participants quickly formed a narrower hypothesis, such as “numbers increasing by two” or “even numbers increasing by two”, and then tested it with sequences that fitted it, such as 8-10-12 or 20-22-24. Every one of these received a “yes”, which increased their confidence, but none could show the hypothesis to be wrong. A sequence that broke their hypothesis, such as 1-2-3 or 5-17-90, would have told them more, because a “yes” to it would show that their rule was too narrow.

Wason concluded that participants relied on confirming evidence (enumerative induction) instead of seeking evidence that could eliminate hypotheses. Those who announced two or more incorrect rules seemed unable or unwilling to test their ideas against sequences that might fail. He linked the finding to scientific thinking, in which a hypothesis is tested properly only by trying to falsify it.

YouTube

A positive test strategy is not always irrational

Klayman and Ha (1987) argued that the 2-4-6 results do not show a general failure of reasoning. Participants used a positive test strategy: testing cases that their hypothesis predicts will be “yes”. This strategy is efficient in many real situations, in particular when the property being searched for is rare.

Wason’s task is unusual because the true rule is much broader than the hypothesis people form, which is exactly the situation in which positive tests are least informative. Performance also depends on how the task is framed: when a later version told participants that every sequence fitted either a “DAX” rule or a “MED” rule, so that testing a sequence against one rule was also a test of the other, far more of them found the correct rule.

This critique matters for evaluation. Confirmation bias in hypothesis testing may partly reflect a sensible default that fails in particular environments, rather than a fixed flaw. The critique applies less to biased interpretation and biased memory, which distort evidence after it has been collected.

Why confirmation bias happens

Confirmation bias has several causes, and they come from different levels of explanation. A full account combines cognitive limits, motives, brain processes and social context, which is why confirmation bias is a good example of the interaction between approaches.

Cognitive explanations: limited capacity and the way hypotheses are tested

Holding a hypothesis and its alternatives in mind at the same time is demanding. Attention and working memory are limited, so people tend to consider one hypothesis at a time, and the hypothesis being considered is usually the one they already hold. Asking “what would I expect to see if I am right?” comes easily; asking “what would I expect to see if I am wrong?” requires imagining a different state of the world.

In dual-processing terms, the quick, automatic System 1 constructs a coherent story from whatever information is available and treats it as complete. Looking for disconfirming evidence is a slow, effortful System 2 activity that is skipped when people are tired, rushed or confident. Schemas add to the effect: information that fits an existing schema is processed fluently and remembered, while information that does not fit is more easily overlooked or reinterpreted.

Motivational explanations: wanting to be right

People also prefer some conclusions to others. Being wrong is unpleasant and can threaten self-esteem, so evidence against a valued belief is scrutinised more harshly than evidence for it. This is motivated reasoning: reasoning in which the goal of reaching a desired conclusion shapes how evidence is processed.

Taber and Lodge (2006) found that participants with strong prior attitudes to gun control and affirmative action rated arguments that agreed with them as stronger, spent their effort counter-arguing the opposing arguments, and chose to read supporting arguments when they could select sources freely. These effects were strongest among the participants who knew most about politics, and they led to more extreme attitudes. The link with cognitive dissonance is direct: evidence against a belief that is part of a person’s identity creates discomfort, and discounting the evidence removes it.

A biological explanation: how the brain weighs agreement

Kappes, Harvey, Lohrenz, Montague and Sharot (2020) asked pairs of participants to judge the prices of properties and then showed each person their partner’s judgement and how confident the partner was. When the partner agreed, participants became more confident the more strongly the partner held their view. When the partner disagreed, the strength of the partner’s view made almost no difference. Functional MRI showed a matching pattern in the posterior medial prefrontal cortex, which tracked the strength of others’ opinions when they confirmed the participant’s judgement but much less when they disconfirmed it.

The study shows that the bias is visible in how the brain represents other people’s opinions, with disconfirming opinions coded more weakly. It does not show that a brain region causes confirmation bias. The finding is consistent with the idea that the bias involves more than a conscious choice to ignore inconvenient evidence.

Sociocultural explanations: groups, identity and information environments

Many beliefs are shared with groups that matter to a person, such as a political party, a religion, a fan club or a profession. Rejecting a group belief can mean conflict with the group, so there is social as well as personal pressure to defend it. People also choose friends, media and online communities that share their views, and recommendation systems on social media tend to show more of what a user already engages with. The result can be an echo chamber, an information environment in which supporting evidence is easy to find and opposing evidence is rare, so that biased search is built into the environment itself.

Confirmation bias in real-world decisions

Confirmation bias matters most where an early hypothesis guides an expensive or irreversible decision. Research has studied it in forensic science, police interviews, classrooms, clinics and science itself.

Forensic science: experts who see what they anticipate

In 2004, after the Madrid train bombings, FBI fingerprint analysts declared a “100 percent” match between a print found on a bag of detonators and Brandon Mayfield, a lawyer in Oregon. Spanish police disputed the match and later identified the print as belonging to an Algerian man, Ouhnane Daoud. Mayfield was released after about two weeks in custody, and the US government later apologised and settled his lawsuit. The case showed that expert judgement can be steered by expectations about a suspect once an initial match has been declared.

Dror, Charlton and Péron (2006) tested this directly. They took fingerprints that experienced fingerprint experts had themselves identified as matches earlier in real casework, and presented them again to the same experts with context suggesting that the prints were the ones wrongly matched in the Mayfield case. Most of the experts now judged the prints as not matching or inconclusive, contradicting their own earlier identifications. The evidence was identical; only the expectation had changed.

Policing: behavioural confirmation in interrogation

Confirmation bias can change the evidence itself as well as how it is read. Kassin, Goldstein and Savitsky (2003) had 52 student “suspects” either commit a mock theft or not, and then be questioned by 52 student “interrogators” who had been led to expect that most suspects were guilty or that most were innocent. Interrogators who expected guilt chose more guilt-presumptive questions, used more persuasive techniques, put more pressure on suspects and were more likely to judge them guilty, especially when the suspects were in fact innocent. Neutral observers who later listened to recordings judged suspects in the guilty-expectation condition as more defensive.

Education: interpreting ambiguous performance

Darley and Gross (1983) showed university students a video of a girl, “Hannah”, playing in either a poor urban neighbourhood or a comfortable suburban one, which suggested a low or high socioeconomic background. Some participants then watched her answer achievement test questions inconsistently, getting some hard questions right and some easy ones wrong. Participants who saw only the background rated her ability at about her grade level in both conditions. Those who also watched the test rated her above grade level if they believed she came from a well-off family and below grade level if they believed she came from a poor one.

Background alone had little effect on the ratings. The stereotype acted as a hypothesis, and the ambiguous test performance was then interpreted as confirming it. Teachers face the same situation whenever a student’s performance is mixed, which is one reason for anonymous marking and for judging work against explicit criteria.

Health: diagnosis and treatment

A clinician who forms an early diagnosis may ask questions that would confirm it and interpret ambiguous symptoms as consistent with it, a pattern that overlaps with anchoring and with premature closure in diagnosis. Patients show confirmation bias too: a person convinced that a treatment works will tend to remember the days on which they felt better after taking it, and someone anxious about a vaccine may search online for reports of side effects and find many. Confirmation bias also helps explain why people sincerely report links that controlled studies do not find, such as between their joint pain and the weather.

Science: researcher bias and peer review

Scientists are not immune. A researcher who expects a result may stop checking an analysis once it gives the expected answer, while checking carefully when it does not. Mahoney (1977) sent journal reviewers manuscripts that were identical except for whether the results supported or contradicted the reviewers’ own theoretical position. Reviewers rated the methods of the study as better and were more likely to recommend publication when the results agreed with their views. Confirmation bias of this kind is one source of researcher bias and of publication bias in the scientific literature.

Relationships: first impressions

In relationships, an early impression of a new colleague, flatmate or partner acts as a hypothesis. Ambiguous behaviour, such as a short reply to a message, is read in line with the impression (“she is unfriendly” or “she is busy”), and memories that fit the impression are recalled when the relationship is discussed. Through behavioural confirmation, treating someone coldly because they seem unfriendly can produce the coldness that confirms the impression.

Key study: Lord, Ross and Lepper (1979), biased assimilation and attitude polarisation

If people weighed evidence fairly, giving two opposing groups the same mixed evidence should bring their views closer together, or at least leave them unchanged. Charles Lord, Lee Ross and Mark Lepper tested whether the opposite happens when people hold strong beliefs on a contested social issue.

Aim

The researchers aimed to investigate whether people with strong opposing views on capital punishment would evaluate the same mixed evidence in a biased way, rating studies that supported their view as more convincing, and whether reading that evidence would make their views more extreme.

Method

The participants were 48 undergraduates at Stanford University in the United States. They were selected because, on a questionnaire completed earlier, they had expressed strong views on capital punishment: half believed it deters murder and supported it, and half believed it does not and opposed it.

Each participant read about two fictitious studies on whether the death penalty deters murder. One compared murder rates in US states before and after they introduced capital punishment; the other compared murder rates in neighbouring states with and without it. One study appeared to show a deterrent effect and the other showed none. Which design produced which result was counterbalanced, so that any difference in ratings could not be due to one design being genuinely better. Participants first read a short summary of each study’s results, then a fuller description with the procedure, criticisms of the study and the authors’ replies.

After each part, participants rated how convincing and how well conducted each study was, and reported how their own attitude to capital punishment, and their belief in its deterrent effect, had changed. The independent variable was participants’ prior attitude (a quasi-experiment variable, since it could not be assigned), together with whether each study supported or opposed that attitude; the dependent variables were the ratings and the self-reported attitude change.

Results

Both supporters and opponents rated the study that agreed with their own position as more convincing and better conducted than the study that disagreed with it, and this held whichever research design the agreeing study used. Participants accepted the confirming study at face value and found many flaws in the disconfirming one. After reading all the material, both groups reported that their attitudes had become more extreme in their original direction. Supporters reported becoming more in favour of capital punishment and opponents more opposed, although both had read the same evidence.

Sources: Lord, C. G., Ross, L. and Lepper, M. R. (1979). Biased assimilation and attitude polarization: the effects of prior theories on subsequently considered evidence. _Journal of Personality and Social Psychology, 37(11), 2098–2109._

Conclusions

People with strong beliefs assimilate new evidence in a biased way: they accept confirming evidence uncritically and examine disconfirming evidence until they find reasons to dismiss it. As a result, mixed evidence strengthens both sides, and exposing opposing groups to the same body of evidence can increase polarisation instead of reducing disagreement.

Evaluation

Measurement: was polarisation measured or only reported?

Attitude change was measured by asking participants how much their views had changed, not by measuring their attitudes before and after with the same scale. Self-reported change may reflect what participants believed had happened, or what they felt a committed supporter should say, rather than real change.

Later research separated the two. Kuhn and Lao (1996) found that genuine polarisation, a shift from a moderate to a more extreme view, was a real but infrequent outcome of reading mixed evidence, and that simply thinking about the topic produced much of the attitude change. Biased assimilation is therefore the more secure finding, and polarisation is less general than the original study suggested.

Bias: who was studied?

The sample was small and consisted of students at one elite US university, selected for holding unusually strong views. This is a deliberate choice, because the theory concerns strongly held beliefs, but it limits generalisation to people with moderate views and to other cultures. Later studies have broadened the evidence. Taber and Lodge (2006) found similar biased evaluation of arguments about gun control and affirmative action, strongest among politically knowledgeable participants, and the effect has been found for many issues since. Most of this research, however, still comes from the United States and other Western countries, where politics is highly polarised.

Responsibility: deception and belief perseverance

Participants were deceived: the studies they read were invented. Deception was necessary, because real studies would differ in quality and the experimenters needed identical evidence of equal strength. The issue, however, is not trivial. Ross, Lepper and Hubbard (1975) showed that beliefs created by false information can survive a standard debriefing, so participants who left the study with stronger views on capital punishment, based on fictitious evidence, might not have fully reverted when told the studies were invented. A careful debriefing that explains how such beliefs persist is the responsible way to run studies of this kind.

Reducing confirmation bias

Because the bias is largely unintentional, telling people to “be fair” does little. Lord, Lepper and Preston (1984) compared instructions to be as fair and unbiased as possible with instructions to consider the opposite: to ask whether they would have rated a study the same way if it had produced the opposite result. Considering the opposite reduced biased assimilation more than the instruction to be unbiased. It works by making alternative possibilities accessible, which is the step that people skip.

Other strategies change the situation rather than the person:

• Blind procedures. Fingerprint analysts can study a print before learning anything about the suspect, and markers can grade work without names. Removing the information that creates the expectation removes the hypothesis to be confirmed.

• Pre-registration in research. Researchers state their hypotheses and analysis plan before collecting data, which limits the freedom to analyse until the expected result appears.

• Seeking out disagreement. Groups can assign one member to argue against the favoured option, and individuals can deliberately read sources from the other side of a debate.

• Falsification as a habit. Asking “what evidence would show that I am wrong?” turns a positive test strategy into a test that can fail, as in the 2-4-6 task.

YouTube

Evaluating confirmation bias as an explanation of decision-making

Confirmation bias is among the best-documented cognitive biases, but the label covers several different processes, and the research behind it varies in quality. Evaluation organised through the concepts brings out both its strength as an explanation and its limits.

Measurement: one bias or several?

Confirmation bias is measured in very different ways: the sequences chosen in the 2-4-6 task, ratings of study quality, the questions chosen in an interview, the details recalled from a description. These tasks measure different processes (search, interpretation, memory), and a person who shows the bias on one may not show it on another. Some researchers argue that “confirmation bias” is an umbrella term for several related tendencies rather than one mechanism. A strength of the research is that many of these measures are behavioural, such as which sequences people test or which arguments they choose to read, rather than self-reports.

Causality: experiments and real cases

Laboratory studies such as Wason (1960), Darley and Gross (1983) and Kassin et al. (2003) manipulate the expectation and measure its effect, so they support a causal claim: holding a hypothesis changes how evidence is collected and judged. Real cases such as the Mayfield misidentification fit the same pattern but cannot show causation on their own, because many factors contributed to the error. Dror et al. (2006) bridge the gap by manipulating context for real experts using real casework prints, although with a small number of experts.

Bias: is the research itself biased?

Most studies use Western university students, and many concern issues that are especially polarised in the United States, such as capital punishment and gun control. The size of the bias may differ in cultures where open disagreement is less valued or where people are more comfortable holding apparently contradictory ideas. There is also a risk of confirmation bias in confirmation-bias research: striking demonstrations are published and cited, while Kuhn and Lao’s (1996) finding that genuine polarisation is infrequent receives less attention.

Perspective: bias or reasonable strategy?

From a normative perspective, any unequal treatment of evidence is an error. From another perspective, some of it is reasonable. Klayman and Ha (1987) showed that positive testing is efficient in many environments, and it is rational to examine surprising evidence more carefully than evidence that fits a well-established belief: a single study suggesting that smoking is harmless deserves more scrutiny than one confirming that it is harmful. The bias becomes a problem when the prior belief itself is weak, as with an early impression of a person, and when disconfirming evidence is dismissed for reasons that would never be applied to confirming evidence.

Responsibility: applying the findings

Research on confirmation bias has led to real reforms. Forensic laboratories have introduced procedures in which analysts study evidence before seeing information about the suspect, and pre-registration is now common in psychology. There are also responsibilities in using the idea. Labelling an opponent’s view as “just confirmation bias” can be a way of avoiding their arguments, and research on politically charged topics must be careful not to present one side as more biased than the other without evidence.

Exam practice

Confirmation bias is an approach term, so it can appear by name in Paper 1 Section A, and in Paper 1 Section B as the term to apply to a scenario from any context. It also supplies evidence for a Paper 1 Section C essay on cognitive biases in decision-making, and it is a natural example when discussing the concept of bias in Paper 2 and in the internal assessment.

Summary

• Confirmation bias is the tendency to search for, interpret and remember information in ways that support existing beliefs, while underweighting contradicting information.

• Its three forms (biased search, biased interpretation, biased memory) reinforce one another in a cycle, so beliefs can feel well supported by weak or mixed evidence.

• Belief perseverance (Ross, Lepper and Hubbard, 1975) and attitude polarisation are consequences: beliefs outlast discredited evidence, and mixed evidence can push opposing groups apart.

• In Wason’s (1960) 2-4-6 task, only 6 of 29 participants found the rule without first announcing a wrong one, because most tested only sequences that fitted their hypothesis; Klayman and Ha (1987) argued that this positive test strategy is often sensible.

• Causes operate at several levels: limited capacity and System 1 processing, motivated reasoning, weaker neural coding of disconfirming information, and group identity and echo chambers.

• Lord, Ross and Lepper (1979) showed that supporters and opponents of capital punishment rated the same evidence by different standards and reported more extreme views afterwards.

• Confirmation bias affects forensic science, as Dror et al. (2006) showed, and also police interrogation, education, medicine and science itself, as shown by Kassin et al. (2003), Darley and Gross (1983) and Mahoney (1977).

• Considering the opposite, blind procedures and pre-registration reduce the bias; simply telling people to be fair does little.

Self-check

• To do:

  I can describe confirmation bias and its three forms, with an example of each.

• To do:

  I can explain belief perseverance and attitude polarisation and how they follow from confirmation bias.

• To do:

  I can describe Wason’s 2-4-6 task and explain what it shows, including the positive test strategy critique.

• To do:

  I can explain cognitive, motivational, biological and sociocultural explanations of confirmation bias.

• To do:

  I can describe the aim, method, results and conclusions of Lord, Ross and Lepper (1979).

• To do:

  I can apply confirmation bias to a scenario in any of the four contexts, showing at least two of its forms.

• To do:

  I can evaluate confirmation bias research with reference to measurement, causality, bias and responsibility.

• To do:

  I can discuss strategies that reduce confirmation bias and explain why they work.