Skip to content
Behavioural Science

← All essays and guides

Guide

Cognitive Biases List: 15 Effects Ranked by Evidence

Which cognitive biases hold up? Bias versus heuristic, the Gigerenzer and Kahneman debate, 15 biases rated by evidence, and why 200-item bias lists mislead.

Martin Metzmacher
· 8 min read

In this guide

Most lists of cognitive biases mix three things: effects that replicate, shortcuts that are usually accurate, and labels nobody has tested. This guide sorts fifteen entries into those piles and shows where the evidence is thin.

A bias is a systematic error, and a heuristic is a shortcut

A heuristic is a simple rule that uses little information and little computation. Pick the option you recognise. Take the first option that clears a threshold. A bias is a systematic deviation from a standard of correct judgement. A heuristic can produce a bias, and it can also be the most accurate tool available.

Calling something a bias makes two claims at once. The judgement deviated, and a norm says it should not have. The second claim is where most list entries go wrong. The norm is usually probability theory or expected utility. Those yardsticks are not always the right ones for a decision made under time pressure with incomplete data. The label quietly assumes the yardstick.

Tversky and Kahneman’s 1974 review named three heuristics behind judgement under uncertainty: representativeness, availability and anchoring (Tversky & Kahneman, 1974). Their claim was that these shortcuts produce predictable errors, and that the errors reveal the mechanism. That programme made the bias list possible.

Gigerenzer and Kahneman disagree about the norm

The sharpest critic is Gerd Gigerenzer. His earlier work showed that a simple one-reason rule can predict as well as more elaborate models in some tasks (Gigerenzer & Goldstein, 1996). His reply to Kahneman and Tversky argues that many biases are artefacts of narrow norms applied to problems people do not face in that form, and that heuristics stay vague until their conditions are specified (Gigerenzer, 1996). His best-known demonstration is that base-rate problems become far easier when the same numbers are given as natural frequencies (Gigerenzer & Hoffrage, 1995).

Kahneman and Tversky answered that the illusions are real and appear in ordinary problems, not only in contrived ones (Kahneman & Tversky, 1996). A heuristic account, they argued, has to predict which errors occur and when.

Both sides agree on more than the headlines suggest. Neither believes people are optimisers, and both accept that fast rules work well in some settings. The disagreement has two parts: which norm applies, and whether the errors are the finding or the price of a mechanism that usually pays off. Stanovich and West added that performance varies strongly across individuals, which a single-system story struggles to explain (Stanovich & West, 2000). Evans and Stanovich review the debate over the structure of dual-process theory (Evans & Stanovich, 2013). For that frame, see system 1 and system 2 dual process theory.

When you meet a bias claim, ask two questions. Which norm is being applied? Does the task resemble the lab’s format? Those two questions settle more disputes than the headline does.

Fifteen biases and how well they hold up

BiasWhat it doesEvidence status
AnchoringEstimates drift toward an irrelevant starting numberRobust in the lab; size varies by task
Framing effectSame outcome, different wording, different choiceReplicates in the lab; field effects smaller and less consistent
AvailabilityFrequency judged by how easily examples come to mindRobust in the lab; mechanism still debated
Base-rate neglectCase details override population frequenciesRobust in standard wording; shrinks with natural frequencies
Conjunction fallacyRates “A and B” as likelier than “A” aloneRobust in standard wording; interpretation contested
OverconfidenceIntervals too narrow for what people knowRobust for narrow intervals; “better than average” version weaker (Moore & Healy, 2008)
Planning fallacyUnderestimates the time a task will takeRobust across the tasks studied
Hindsight biasJudges an outcome predictable after learning itRobust in controlled studies
Confirmation biasSeeks and weights evidence that fits a beliefBroadly robust in review; strength varies by task
Sunk costKeeps investing because of past spendingRobust in the lab; field evidence mixed
Status quo biasPrefers the current option when alternatives are addedRobust in the lab; field size not well established
Endowment effectOwners value a good above what buyers will payContested; the gap shrinks under controlled procedures (Plott & Zeiler, 2005)
Loss aversionLosses weigh more than equal gainsWidely cited; coefficient estimates unstable (Gal & Rucker, 2018)
Fundamental attribution errorExplains others’ actions by character and ignores situationRobust as a tendency; size and cultural reach disputed
Dunning-Kruger effectLow performers overrate their skillContested; largely a statistical artefact (Gignac & Zajenkowski, 2020)

These ratings are my reading of the original studies and the replication record. They are not a formal meta-analytic grade. The gap between lab and field is the weakest link in nearly every row.

Why 200-item bias lists mislead

Popular posters and web lists run to 180 or 200 entries. The length is the first warning. Many entries overlap, describing one mechanism under different names. Most carry no effect size, and most were tested in one lab, if at all. A label without a measured effect tells you what to call a mistake, not how big it is.

The second warning is reversal. Gilovich, Vallone and Tversky (1985) found no evidence that basketball shooters have a hot hand, and the “hot hand fallacy” became a staple of bias lists. Miller and Sanjurjo (2018) later showed that the original test contained a subtle selection bias, and that a hot hand is consistent with the data. A list that does not track such reversals is a snapshot.

The third warning is replication. The Open Science Collaboration (2015) attempted to replicate 100 psychology studies and found that fewer than half of the replications reached statistical significance. Not every one of those studies concerned a bias. The lesson still applies: a famous name on a poster is weak evidence.

What reduces bias, and what does not

Telling people to be more careful is the weakest intervention on most lists. Changing the structure of the task does more. Present frequencies instead of probabilities. Ask for a range before a point estimate. Write down the base rate before reading the case. None of these removes a bias from a mind. They change the input so the shortcut works from better data.

Business examples are easy to find. A software team estimates a migration at three weeks and needs eight. The usual fix is to start from how long similar past projects took. A team keeps a failing feature because two years of work went into it. That is sunk cost, and the fix is a rule written before the spending starts. Anchoring, framing and endowment show up in pricing pages, trial conversions and negotiations, each with its own angle. Present bias, social proof and the decoy effect are left out of the table on purpose. They are different kinds of effect, and each needs its own test.

Structural fixes are not free. Each adds time and friction, and few have been tested at scale in business settings. Treat them as hypotheses and measure them the way you would measure a product change.

The verdict: keep fifteen, test them, drop the poster

In my view, a useful bias list is short, and each entry earns its place through a measured effect in a setting like yours. The rest is decoration. Use the table above as a set of hypotheses, not as diagnoses. The next step for research is not another list. It is preregistered field trials that compare structural fixes against simple reminders, powered for the small effects that most field work actually finds.

If your team wants to know which of these effects moves its own numbers, our applied behavioural science practice runs that kind of test. Start with one decision, one metric and one pilot.

Frequently asked questions

Is a heuristic the same as a cognitive bias?

No. A heuristic is a shortcut. A bias is a systematic error measured against a norm. Many heuristics are accurate in the environments they suit, and some produce predictable errors in others.

Who won the Gigerenzer and Kahneman debate?

Neither side won outright. Kahneman and Tversky showed that systematic errors appear in standard problems. Gigerenzer showed that many of those errors shrink when the problem format or the norm changes. Which claim matters depends on the decision in front of you.

Are 200-bias lists useful at all?

As a memory aid, perhaps. As evidence, no. Most entries have no measured effect size and no independent replication. Start from a short list, such as the table above, and check each entry against the original study before you rely on it.

References