Cognitive Bias and Confirmation Bias

A belief forms.

Then something subtle happens.

Evidence supporting it becomes easy to notice.

Evidence against it becomes easy to explain away.

This is confirmation bias.

It is one member of a much larger family of systematic cognitive biases.

What Is a Cognitive Bias?

A cognitive bias is a systematic tendency in judgment that can depart from an appropriate normative standard.

The standard may involve:

• logic,
• probability,
• accuracy.

Not every shortcut is irrational.

Heuristics

A heuristic is a fast rule of thumb.

Heuristics reduce cognitive cost.

They are often useful.

Bias appears when a heuristic produces predictable error.

Bounded Rationality

Humans have limited:

• time,
• memory,
• computation.

We cannot optimize every decision.

Herbert Simon called this bounded rationality.

Ecological Rationality

Some heuristics work well in environments for which they are adapted.

A rule can be biased under one laboratory benchmark yet useful in real life.

Context matters.

Confirmation Bias

Confirmation bias is the tendency to favor information that supports existing beliefs or hypotheses.

It can operate through:

• search,
• interpretation,
• memory.

Selective Search

If you believe:

coffee improves health,

you may search:

“benefits of coffee”

instead of:

“health risks of coffee.”

The query itself shapes evidence exposure.

Positive Test Strategy

People often test hypotheses by looking for cases where the hypothesis predicts:

yes.

This can be rational sometimes.

But it can fail to distinguish alternatives.

Wason Selection Task

Peter Wason’s experiments showed people often test logical rules by seeking confirming cases rather than potential falsifiers.

The exact interpretation is debated, but the task became a classic example.

Example Rule

Suppose cards show:

A, D, 4, 7.

Rule:

“If a card has a vowel on one side, it has an even number on the other.”

Which cards must be turned?

Correct Test

You must inspect:

• A,
• 7.

A could violate the rule with an odd number.

7 could violate it if the other side is a vowel.

Turning 4 is not necessary.

Why People Choose 4

Four feels confirmatory.

But finding a vowel behind 4 does not test the conditional effectively.

Falsification requires looking where the rule could fail.

Interpretation Bias

Ambiguous evidence can be interpreted to support prior belief.

The same event becomes:

proof

for opposing sides.

Biased Assimilation

People may evaluate supportive evidence as:

strong

and opposing evidence as:

flawed.

The asymmetry preserves belief.

Lord, Ross, and Lepper

Classic studies on controversial issues found that people exposed to mixed evidence could become more polarized.

The exact size and generality of backfire effects vary.

But biased evaluation is well documented.

Backfire Effect Caution

The claim that correction usually makes false beliefs stronger is exaggerated.

Backfire can occur, but many corrections work.

We should not turn one bias claim into another myth.

Belief Polarization

When opposing groups process evidence differently, both can become more confident.

Shared evidence does not guarantee convergence.

Memory Confirmation

People remember supportive examples more readily.

Contradictory cases may fade.

Belief shapes memory retrieval.

Motivated Reasoning

Confirmation bias becomes stronger when belief is connected to:

• identity,
• desire.

Reasoning then protects a preferred conclusion.

Directional Motivation

Accuracy motivation asks:

What is true?

Directional motivation asks:

Can I justify the conclusion I want?

The two can produce very different reasoning.

Identity-Protective Cognition

If a belief signals group membership, changing it can threaten social belonging.

The epistemic cost of error competes with the social cost of dissent.

Group Polarization

Discussion among like-minded people can shift the group toward more extreme positions.

Shared arguments reinforce confidence.

Echo Chambers

An echo chamber does not merely reduce exposure.

It can preemptively discredit outside sources.

This protects the internal belief system.

Filter Bubbles

Algorithmic personalization can increase repeated exposure to similar content.

The empirical magnitude varies by platform and behavior.

Still, selective information environments matter.

Availability Bias

We judge likelihood partly by how easily examples come to mind.

Vivid recent events dominate memory.

This can distort risk.

Base-Rate Neglect

People often focus on case details and underweight prior probabilities.

This is base-rate neglect.

It becomes critical in medical testing.

Example

Suppose a disease affects:

1 in 1,000 people.

A test is highly accurate.

A positive result may still have a surprisingly modest probability of actual disease if false positives are not extremely rare.

Bayesian reasoning is needed.

Anchoring

An initial number can influence later estimates.

Even arbitrary anchors can shift judgment.

Adjustments are often insufficient.

Framing Effect

Equivalent choices produce different decisions depending on presentation.

“90% survival”

feels different from:

“10% mortality.”

The factual content is equivalent.

Loss Aversion

Losses often weigh more heavily than equal gains.

This is central to prospect theory.

The effect varies across context.

Endowment Effect

People may value an object more once they own it.

Ownership changes valuation.

This challenges simple stable-preference models.

Status Quo Bias

People often prefer existing arrangements.

Changing requires justification.

Defaults therefore have power.

Sunk Cost Fallacy

Past costs that cannot be recovered should not affect current rational choice.

Yet people continue projects because:

“We have already invested too much.”

This is the sunk cost fallacy.

Survivorship Bias

We often study visible successes and ignore failures.

Example:

successful entrepreneurs tell stories about risk.

Failed entrepreneurs using the same strategy disappear from the sample.

Selection Bias

If the observed sample is not representative, conclusions can be distorted.

Bias can enter before analysis begins.

Hindsight Bias

After an event occurs, it feels more predictable than it was.

We say:

“I knew it all along.”

Memory of prior uncertainty is reconstructed.

Outcome Bias

We judge decision quality by result.

A good decision can produce a bad outcome by chance.

A bad decision can succeed through luck.

Halo Effect

A positive impression in one domain spills into others.

Someone attractive or charismatic may seem:

more competent

without evidence.

Authority Bias

People give extra weight to statements from perceived authority.

This can be rational when expertise is relevant.

It becomes bias when authority substitutes for evidence.

In-Group Bias

We often trust and favor members of our own group.

This supports cooperation.

It can distort impartial judgment.

Fundamental Attribution Error

We often explain others’ behavior through personality while underestimating situational factors.

For our own behavior, we notice context more readily.

Self-Serving Bias

Success is attributed to:

skill.

Failure to:

bad luck.

This protects self-image.

Optimism Bias

People often underestimate their own risk of negative outcomes.

Some optimism is adaptive.

Too much impairs planning.

Planning Fallacy

Projects often take longer and cost more than expected.

People focus on:

best-case internal plans

rather than reference classes.

Outside View

Daniel Kahneman emphasized the outside view:

compare your case to similar past cases.

Base rates often outperform detailed internal narratives.

Reference Class Forecasting

For projects, ask:

How long did comparable projects actually take?

This counters planning optimism.

Dunning–Kruger Effect

The popular slogan:

“stupid people think they are smart”

is an oversimplification.

The broader issue is that self-assessment is noisy and sometimes systematically miscalibrated.

Bias Blind Spot

People readily recognize biases in others.

They underestimate bias in themselves.

Knowing the names of biases does not immunize us.

Bias Lists Can Backfire

Memorizing dozens of biases can become a rhetorical weapon:

“Your view is biased.”

The useful question is:

What procedure reduces error here?

Debiasing

Effective debiasing often changes process rather than intention.

Examples:

• blind review,
• checklists,
• pre-registration.

Do not rely only on:

“try to be unbiased.”

Consider the Opposite

Ask:

What evidence would make the opposite conclusion plausible?

This forces alternative-model generation.

Steelmanning

State the strongest version of the opposing view.

This reduces straw-manning.

It also exposes which disagreement is substantive.

Disconfirmation Search

Actively search for:

• failed predictions,
• contrary evidence.

This counters selective exposure.

Pre-Mortem

Imagine a project failed.

Ask:

Why?

This encourages identification of risks suppressed by optimism.

Red Teaming

Assign someone to challenge assumptions deliberately.

Organizations use red teams to expose hidden vulnerabilities.

Double-Blind Methods

In experiments, blinding prevents expectations from influencing:

• treatment,
• measurement.

Institutional design can outperform personal virtue.

Randomization

Random assignment helps neutralize confounding.

It prevents researchers from selecting groups in ways that fit expectations.

Replication

A result that survives independent replication is less likely to be one researcher’s bias.

Collective procedures correct individual cognition.

Prediction Markets

Aggregating independent probabilistic judgments can improve forecasting.

They are imperfect but useful for calibration.

Forecasting Tournaments

Repeated prediction with scoring trains:

• calibration,
• updating.

Feedback is essential.

Bayesian Updating

Bayesian reasoning provides a normative model:

update beliefs in proportion to evidence.

It does not eliminate psychological bias.

But it clarifies what ideal updating looks like.

Likelihood Ratios

Ask:

How much more likely is this evidence if my hypothesis is true than if it is false?

This is often better than asking:

Does this evidence fit my hypothesis?

Almost anything can fit a flexible story.

Alternative Hypotheses

Confirmation weakens when evidence is also expected under competitors.

Good reasoning compares hypotheses.

Prediction Before Observation

Make predictions before seeing results.

This prevents hindsight from rewriting expectations.

Calibration

Track:

confidence

against:

accuracy.

A thinker who is 90% confident should be wrong about 10% of such claims.

Without feedback, confidence drifts.

Intellectual Humility

Humility is not:

low confidence everywhere.

It is:

confidence proportional to evidence.

Strong evidence deserves strong belief.

Bias and Expertise

Experts can be less biased in domains with:

• valid cues,
• fast feedback.

Expertise does not generalize automatically outside the trained domain.

Bias and AI

AI systems can inherit biases from:

• data,
• objectives.

They can also amplify systematic patterns at scale.

Machine objectivity is not automatic.

Automation Bias

Humans may overtrust algorithmic recommendations.

A machine output feels precise.

This can reduce independent checking.

Human–Machine Complementarity

Machines and humans make different errors.

A good system combines:

• statistical consistency,
• contextual judgment.

Diversity of failure modes can improve decisions.

The Philosophical Lesson

Cognitive biases are not evidence that humans are hopelessly irrational.

They are side effects of:

• bounded computation,
• adaptive shortcuts,
• social cognition.

Rationality improves when we design procedures that expose our own errors.

The central epistemic virtue is not:

never be biased.

It is:

build systems that make bias correctable.

The Next Question

One of the strongest tools for disciplined belief revision is probability.

Instead of asking only:

true or false?

we can ask:

How confident should I be, given the evidence?

The next essay is:

Probability, Bayesian Reasoning, and Uncertainty.