You believe a new teaching method works.
A student improves dramatically.
You remember the case.
Another student does not improve.
You explain it away.
“They did not practise enough.”
A third student worsens.
“The exam was unusually difficult.”
The belief survives every encounter.
Not because every piece of evidence supports it.
Because supporting evidence is admitted easily while conflicting evidence is made to work much harder to enter.
This is confirmation bias.
Quick Read
Confirmation bias is the tendency to seek, generate, interpret or remember information in ways that favour a belief or hypothesis already under consideration.
The bias can appear at several stages:
- search: looking where confirmation is likely;
- questioning: testing whether the hypothesis is true rather than whether it could be false;
- interpretation: treating ambiguous evidence as supportive;
- scrutiny: demanding stronger proof from opposing evidence;
- memory: recalling confirming examples more easily;
- communication: sharing evidence that strengthens the preferred narrative.
Raymond Nickerson’s influential review described confirmation bias as a broad family of behaviours rather than one single laboratory trick. Earlier work on hypothesis testing, including Peter Wason’s studies, showed how people often search for evidence that could confirm a rule rather than information that could decisively falsify it.
The central question is:
If I wanted this belief to be wrong, what evidence would I have looked for differently?
The One-Sentence Answer
Confirmation bias works when an existing belief changes the path of evidence collection and evaluation, so supporting information receives easier access while disconfirming information is less likely to be searched for, accepted or remembered.
The Confirmation-Bias Chain
initial belief → selective search / positive test → asymmetric interpretation → confirming cases become salient → conflicting cases discounted → confidence rises → future search narrows further
The loop matters.
Confirmation bias can manufacture the impression that evidence is one-sided because the belief influences which evidence enters the sample in the first place.
Confirmation Bias Is Not Simply “Being Wrong”
A person can hold a correct belief and still reason with confirmation bias.
The problem is not the truth value of the conclusion.
The problem is the asymmetry of the evidence process.
If you search only for reasons your correct belief is correct, you may still miss important boundary conditions, exceptions or failure modes.
Confirmation Bias Is Not Research Bias
Research bias is broader.
It includes selection bias, measurement bias, confounding, publication bias, reporting bias and many other distortions that can occur even without a prior preferred belief.
Confirmation bias is specifically directional: the current belief changes how evidence is sought or evaluated.
Confirmation Bias Is Not Availability
Availability makes easy-to-recall examples feel more common.
Confirmation bias makes belief-consistent evidence easier to search for, accept or remember.
The two can reinforce each other.
Once confirming examples are repeatedly noticed, they become more available.
See How The World Works | Availability Heuristic.
Positive Testing
Suppose the hypothesis is:
People who enjoy chess are good at Mathematics.
A positive-testing strategy asks chess players whether they are good at Mathematics.
That sounds sensible.
But a decisive test also needs people who do not play chess, or strong Mathematics students who do not fit the pattern.
A test can look directly at the hypothesis without being diagnostic enough to distinguish it from alternatives.
Wason’s Rule-Testing Problem
Classic experiments asked participants to discover a rule by proposing examples and receiving feedback.
Participants often proposed sequences consistent with their current guess rather than sequences designed to distinguish their guess from competing rules.
The result illustrates a deep principle:
A confirming test is not necessarily an informative test.
Diagnosticity Beats Agreement
When testing a hypothesis, the best evidence is not the evidence most likely to agree with it.
The best evidence is evidence that would look different if a competing hypothesis were true.
That is diagnostic evidence.
A medical symptom present in almost every disease confirms almost nothing.
A test with different likelihood under competing diagnoses is useful.
Good inquiry asks what observation would separate worlds.
Asymmetric Scrutiny
Evidence we like often enters cheaply.
Evidence we dislike is audited aggressively.
A supportive study:
“Interesting. Strong result.”
An opposing study:
“What was the sample size? Who funded it? Was it preregistered? Did they correct for multiple comparisons?”
Those are excellent questions.
The bias appears when we ask them only of the evidence that threatens our position.
Motivated Reasoning
Confirmation bias can be stronger when the belief protects identity, status, money or belonging.
An investor owns the stock.
A manager designed the strategy.
A parent chose the school.
A researcher proposed the theory.
Now disconfirming evidence does not merely threaten a proposition.
It threatens something personally costly.
Motivation changes the price of admitting error.
Confirmation Bias and Loss Aversion
Changing your mind can feel like losing intellectual ownership.
If a belief has become part of identity, abandoning it can feel more costly than acquiring a better belief feels rewarding.
Loss aversion can therefore intensify belief defence.
See How The World Works | Loss Aversion.
Confirmation Bias and Anchoring
An initial hypothesis can become an anchor.
Later evidence is interpreted as adjustment around that starting belief rather than a fresh comparison of competing explanations.
Anchoring owns the pull of the starting point.
Confirmation bias owns the directional evidence process that protects it.
See How The World Works | Anchoring.
Confirmation Bias and Framing
A preferred belief can determine which frame feels natural.
Supporters describe a policy as “protecting jobs.”
Opponents describe it as “raising costs.”
Both may cite true facts while selecting the representation that reinforces their prior position.
See How The World Works | Framing.
Confirmation Bias and Base Rates
A person wants a treatment to work.
They find three testimonials.
The testimonials confirm the belief.
The base rate of improvement among all users remains unknown.
Confirmation bias selects the examples.
Base-rate neglect overweights them.
See How The World Works | Base-Rate Neglect.
Science Is Designed to Fight This Problem
Science does not assume scientists are unbiased.
It builds procedures because scientists are human.
- control groups;
- blinding;
- preregistration;
- replication;
- peer criticism;
- transparent methods;
- registered outcomes;
- adversarial testing;
- independent data.
These mechanisms make it harder for one preferred explanation to control the whole evidence pipeline.
See How Scientific Research Works and How Preregistration Works.
Preregistration Separates Prediction From Explanation
After seeing results, humans are very good at explaining why the result makes sense.
That flexibility can turn surprise into apparent confirmation.
Preregistration records hypotheses, outcomes and analysis plans before the data are known.
Exploration remains valuable.
It is simply labelled honestly as exploration rather than retroactive prediction.
The Investment Example
You buy a stock because you believe a new product will dominate its market.
From that moment, every news item has two jobs.
Inform you.
Threaten or protect the position you already own.
Positive sales data feel meaningful.
Negative data become “temporary noise.”
A disciplined investor writes the sell thesis before buying:
What evidence would tell me this thesis is wrong?
The Business Strategy Example
A management team launches a new strategy.
Revenue rises in one region.
“The strategy is working.”
Revenue falls elsewhere.
“Implementation was weak.”
Perhaps both statements are true.
But if every positive result is attributed to strategy and every negative result to execution, the strategy becomes impossible to falsify.
An unfalsifiable strategy cannot learn.
The Education Example
A teacher believes a student is lazy.
Late homework confirms laziness.
Excellent oral answers are treated as exceptions.
Then the teacher learns the student is caring for a younger sibling every evening.
The same evidence can now support a different causal model.
Labels are dangerous because they become search instructions.
Once “lazy” is the hypothesis, behaviour is scanned for laziness.
Mark the Work Before Reading the Old Grade
If a teacher sees that a student usually scores 55 before marking a new essay, the old belief can influence interpretation.
Blind or partially blind marking can protect the new evidence from the old model.
The technique is simple:
Let the current work speak before the prior reputation does.
The News Example
Two readers with opposing views read the same report.
Each notices the sentence that confirms their prior model.
Each dismisses the inconvenient paragraph as weak or biased.
Both leave more confident.
Polarisation can therefore arise even when people receive overlapping information if their evidence filters differ.
Algorithms Can Reinforce Confirmation
A recommendation system learns what a user clicks.
The user clicks belief-consistent content.
The system supplies more of it.
Availability rises.
The user infers that “everybody is saying this.”
Personal preference becomes information environment.
This is not a proof that recommendation systems necessarily create ideological bubbles, but it is a mechanism they must actively manage.
The Search Engine Problem
Search queries can bake conclusions into the question.
Search:
Why does policy X fail?
You will retrieve failure arguments.
Search:
Evidence for and against policy X; evaluations; outcomes; limitations
The information environment changes.
Neutral search is difficult, but symmetric search is possible.
Confirmation Bias and Common Knowledge
Groups can develop shared beliefs that shape which evidence members feel safe presenting.
If everyone knows the organisation is committed to a strategy, bad news becomes socially expensive.
People self-censor.
Leaders then see fewer negative reports and infer the strategy has broad support.
Confirmation becomes institutional rather than merely cognitive.
See How The World Works | Common Knowledge.
Hierarchy Makes Disconfirmation Expensive
A junior employee sees evidence that the senior leader’s plan is failing.
Speaking up risks status.
The hierarchy therefore changes the evidence pipeline.
Good organisations create channels where disconfirming information is rewarded, protected and escalated.
See How Hierarchy Works.
Red Teams Exist for a Reason
A red team is assigned to attack a plan, model or security posture.
The role changes incentives.
Finding failure is no longer disloyalty.
It is the job.
This is an institutional solution to confirmation bias: make disconfirmation somebody’s explicit responsibility.
Premortems
Imagine the project has failed badly one year from now.
Ask every participant to write why.
The premortem temporarily changes the frame.
Instead of searching for reasons the plan will work, the group is authorised to search for failure routes.
It does not prove the plan will fail.
It repairs the asymmetry in search.
Steelman the Opposition
Before criticising an opposing view, state the strongest version of it that a competent opponent would accept.
This prevents the mind from selecting the weakest evidence against its preferred belief.
If you can only defeat a caricature, you have learned little.
Seek Evidence That Could Change Your Mind
Before investigating, write:
- What do I currently believe?
- What evidence would increase my confidence?
- What evidence would decrease my confidence?
- What result would make me abandon the hypothesis?
If the last question has no answer, you may not be testing a belief.
You may be defending one.
Bayesian Updating as a Discipline
Bayesian reasoning provides a useful ideal.
New evidence should move belief according to how much more likely the evidence is under one hypothesis than another.
The evidence should not receive a different likelihood ratio merely because we prefer one hypothesis.
See How Bayesian Inference Works.
Confirmation Bias Can Sometimes Look Rational
Not every positive test is irrational.
Sometimes confirming tests are efficient because the hypothesis predicts distinctive evidence.
Sometimes search costs make exhaustive falsification impossible.
Sometimes one error is much more costly than another.
Modern accounts therefore distinguish a general positive-test strategy from a genuine confirmation bias.
The question is whether the search strategy discriminates effectively among live alternatives.
The Confirmation-Bias Audit
- Write the current belief. Make the hypothesis explicit.
- List live alternatives. What else could explain the same observations?
- Define disconfirming evidence. What result would reduce confidence?
- Search symmetrically. Look for evidence for and against.
- Use diagnostic tests. Prefer observations that distinguish competing hypotheses.
- Audit search language. Does the query presuppose the conclusion?
- Check scrutiny symmetry. Are supportive and opposing studies held to the same standard?
- Check availability. Are confirming examples easier to remember because you notice them more?
- Check base rates. Are selected anecdotes crowding out population evidence?
- Separate identity from hypothesis. What personal cost comes from admitting error?
- Use blind review where possible. Hide labels that reveal which outcome you expect.
- Invite adversarial review. Give someone the job of finding failure.
- Run a premortem. Imagine the preferred plan failed and explain why.
- Record predictions before outcomes. Prevent retrospective reinterpretation.
- Update visibly. Write what changed your mind and by how much.
When the Confirmation-Bias Lens Fails
The lens fails when disagreement is automatically diagnosed as bias.
People can inspect the same evidence and rationally differ because they have different priors, values or models.
It fails when positive testing is assumed to be inherently irrational.
It fails when “be open-minded” becomes an excuse to give weak evidence equal weight with strong evidence.
And it fails when the label is weaponised against opponents while one’s own evidence process remains unaudited.
A Better Question Than “What Evidence Supports Me?”
Ask:
What evidence would be easier to see, search for or accept if I believed the opposite—and have I given that evidence a fair route into the decision?
How Confirmation Bias Connects to the Rest of the World
- Evidence: belief changes which evidence is admitted and weighted.
- Research bias: confirmation can contaminate design, analysis and reporting.
- Availability: repeatedly noticed confirming examples become easier to retrieve.
- Base-rate neglect: selected cases can crowd out population prevalence.
- Anchoring: the first hypothesis can become the reference point for later interpretation.
- Framing: preferred beliefs encourage preferred representations of the same facts.
- Loss aversion: abandoning an owned belief can feel like loss.
- Bayesian inference: likelihood-based updating is a useful normative counterweight.
- Hierarchy: power can suppress disconfirming information before leaders see it.
- Common knowledge: shared commitments change what evidence group members voice.
- Scientific research: blinding, preregistration, replication and critique institutionalise disconfirmation.
Frequently Asked Questions
What is confirmation bias?
Confirmation bias is the tendency to search for, interpret, remember or test information in ways that favour an existing belief or hypothesis.
Is confirmation bias the same as motivated reasoning?
No. Motivated reasoning is broader and emphasises how desired conclusions or identity-related goals shape reasoning. Confirmation bias can occur without strong personal motivation, though motivation often strengthens it.
Is seeking confirming evidence always bad?
No. Positive tests can be efficient when they are genuinely diagnostic. The problem is a search strategy that fails to distinguish the preferred hypothesis from plausible alternatives.
How can organisations reduce confirmation bias?
Use independent estimates, preregistration, red teams, premortems, blind review, symmetric evidence standards, explicit disconfirmation criteria and protected channels for bad news.
Research Basis and Further Reading
- Raymond S. Nickerson, “Confirmation Bias: A Ubiquitous Phenomenon in Many Guises”, a broad review of the phenomenon across reasoning domains.
- Peter Wason’s classic hypothesis-testing studies showed how people often seek confirming rather than falsifying cases.
- Joshua Klayman and Young-Won Ha’s work on positive test strategies added an important nuance: positive testing is not identical to irrational confirmation and must be evaluated by diagnosticity.
What to Read Next on eduKateSG
- How Evidence Works — how claims become more or less believable.
- How Research Bias Works — the wider family of evidence distortions.
- How The World Works | Base-Rate Neglect — why preferred cases can overwhelm population evidence.
- How Preregistration Works — how predictions are separated from explanations made after seeing data.
The Larger Idea
Beliefs are not passive files stored in the mind.
They are search instructions.
They tell attention what to notice.
They tell memory what feels relevant.
They tell ambiguity which direction to lean.
That is why intelligence is not merely the ability to build a strong argument.
It is the ability to build an evidence process strong enough to survive our preference for the argument we already have.
Confirmation bias is what happens when a belief stops waiting for evidence and quietly begins managing which evidence is allowed to arrive.