The popular version of cognitive biases is a list: dozens of named errors, each with a neat definition, all of them supposedly yours. The research behind it is narrower and more useful. Amos Tversky and Daniel Kahneman wrote in Science in 1974 that the shortcuts people use to judge how likely or how typical something is are in general quite useful, and that they sometimes produce severe and systematic errors.¹

A bias is the systematic part: an error that leans the same way, for most people, in the same kind of situation. This page takes biases as a class rather than one at a time, alongside the separate account of the psychology behind everyday choices. Taken as a class, they are messier than the popular version admits. Performance on one bias task barely predicts performance on another, and a 2021 review of how these things are measured found correlations so low they suggest no general factor of susceptibility.²

What cognitive biases are, and why the brain uses shortcuts

A cognitive bias is a judgment that bends in a predictable direction because the mind leaned on a cue instead of the full evidence. The cues Tversky and Kahneman named in 1974 were resemblance to a type, the ease with which examples come to mind, and whatever number was already in front of you.¹

The cue comes first for a plain reason: cues are cheap and evidence is expensive. For Gerd Gigerenzer and Wolfgang Gaissmaier, a heuristic is a strategy that leaves out part of the information on purpose, so a decision arrives more quickly, more frugally or even more accurately than a fuller method would manage.³ Ignoring information is not a bug in that definition. It is the definition.

You can watch it happen in a hiring meeting. Someone reads a resume for fifteen seconds and arrives with a view: strong candidate. It came from two or three features that stood out, and it will now shape which questions get asked. In Daniel Kahneman’s 2003 account, that fast, automatic, effortless and often emotionally charged impression is System 1, while the slower, serial, effortful and deliberately controlled check is System 2, and the check runs only some of the time.⁴

1. 1Too much to weigha judgment is needed before the evidence can be worked through
2. 2A cue stands inresemblance, what comes to mind, the number already on the table
3. 3A fast answerusually about as good as a slow one would have been
4. 4A predictable misswhen the cue does not track what matters, everyone errs the same way

That turns the practical question into a different one: not “am I biased?”, but “which cue is my fast answer resting on, and does it still fit this case?”

A shortcut is not a defect

Gigerenzer’s school reads the same evidence as a story about what minds do well. In their 2011 review, Gigerenzer and Gaissmaier report that where predictability is low and samples are small, ignoring part of the information can produce more accurate judgments than weighting and adding it all, and they argue that what deserves study is which environments a given heuristic fits.³

That framing changes what counts as a fix. If a shortcut is accurate in one setting and wrong in another, the error lives in the mismatch, not in the person. A rule like go with the option you have heard of is a fine way to pick a plumber from a list of strangers and a poor way to pick a supplier where the loudest advertiser is not the best.

1. The shortcut: a rule that ignores some of the information on purpose, so an answer arrives fast
2. Where it fits: when the cue tracks what actually matters, the fast answer is about as good as a slow one
3. Where it does not: change the setting and the same rule keeps pointing the old way
4. Why it gets a name: the miss lands on the same side every time, which is what makes it a bias rather than noise

So audit the setting before you audit yourself. Ask what would have to be true about this market or this candidate for your cue to still carry information. That question has an answer. “Am I being irrational?” does not.

Which famous effects survived a bigger, stricter test

Several of the best-known effects have been rerun at scale, and the results split. In a 2014 multisite project, Richard Klein, Kate Ratliff and dozens of colleagues ran 13 classic and newer effects again in 36 samples: anchoring and framing came back clearly, two priming effects did not.⁵

A split like that is not a scandal. We read it this way: a first study is one small measurement, published partly because it came out striking, while a preregistered retest across dozens of labs measures the same thing far better. When the two disagree, we keep the retest.

Two familiar ideas came out badly. A preregistered multisite test of ego depletion found no clear sign that spending self-control on one demanding task leaves less of it for the next, which undercuts the picture of willpower as a fuel tank.⁶ A 2010 meta-analysis of 50 experiments found the average choice-overload effect was virtually zero, with considerable variation between studies.⁷

You meet the difference as a slide with a finding on it: cut the menu because too much choice paralyzes people, or schedule the hard call before lunch because willpower runs down. Both sound like settled physics, and neither holds as a general rule.

So treat any single striking finding as provisional, including the ones on this page, and give weight to effects that survived retesting by people with no stake in their survival.

The long list of named biases is not one tidy science

The catalogue grew faster than the theory under it, and the complaint has come from inside the field for thirty years. Gigerenzer’s 1996 reply to Kahneman and Tversky argued that the heuristics in the programme were too vague to count as explanations, and that one-word labels such as representativeness were being traded as though they were explanations.⁸

Naming an error is not the same as accounting for it, and several names may be pointing at one thing. Aileen Oeberst and Roland Imhoff proposed in 2023 that several biases studied in separate literatures, including the bias blind spot and outcome bias, can be traced back to one prior belief plus a general tendency toward belief-consistent information processing, a more parsimonious account than treating each as its own mechanism.⁹ If they are right, a long list is partly an illusion of bookkeeping.

Then there is measurement. A 2023 inventory by Vincent Berthet and Vincent de Gardelle counted 41 biases measured over 108 studies, and concluded that reliable measures are still needed for some of them.¹⁰ If a task cannot say reliably who is anchored, then “he is the anchoring type” is a sentence about nothing anyone can measure.

The study

Limited evidence

Three small French samples, eight bias tasks, and correlations that stay low

Berthet reviewed the available measures and ran three studies of his own. Correlations between different bias measures were low, which he reads as the absence of any general factor of susceptibility to cognitive biases, and earlier attempts to build a composite bias score had reliabilities far below the usual standard.²

The samples were small and French, two of students and one of paid adults, and one author ran all three, so this is a warning sign rather than a verdict. It agrees with the earlier work it reviews, and nothing since has produced the bias-proneness trait popular accounts assume.

The practical upshot is unglamorous. “She is a biased thinker” is not a description of anyone: being pulled by an anchor tells you little about whether the same person will chase a sunk cost. In a performance review that label does the work of an insult rather than a diagnosis. Anchoring, confirmation bias and the sunk cost fallacy each have their own evidence.

Further reading

• 

  Gut Feelings: The Intelligence of the Unconscious

  by Gerd Gigerenzer

  Gigerenzer's case, cited here, that shortcuts which ignore information are often accurate, and how to tell when one fits the setting.

As an Amazon Associate WiserHours earns from qualifying purchases.

Knowing about a bias does not switch it off

Learning the name of a bias gives you a better vocabulary, not better judgment. Emily Pronin, Daniel Lin and Lee Ross showed it in 2002: across three surveys, people rated themselves as less subject to various biases than the average American, their classmates and fellow travelers, and students who had just displayed the better-than-average effect insisted their self-assessments were accurate even after reading a description of the bias they had just shown. Knowing about a bias, the authors concluded, neither prevents you from succumbing to it nor makes you aware that you have.¹¹

Myth

Once you know about a bias, you can correct for it.

Fact

In Pronin and colleagues' 2002 studies, people denied being biased immediately after showing the bias and reading about it; in a 1996 set of anchoring experiments, anchors still pulled judgments when participants were forewarned.

The anchoring work points the same way. Timothy Wilson and colleagues reported in 1996 that anchoring appears to operate unintentionally and nonconsciously, in that it is difficult to avoid even when people are forewarned.¹² A warning arrives as information. The pull arrives as a feeling about what a reasonable number looks like, and a feeling does not announce where it came from.

Worse, the people most certain they are unaffected are the hardest to help. In a 2015 experiment, US online participants were randomly assigned to read either an article explaining a common attribution error and how to correct it, or a control article of the same length. The training reduced the error only among people who scored lower on a measure of seeing bias in others but not themselves, and the same scale predicted ignoring advice.¹³

In a meeting it sounds ordinary: “I have already accounted for that.” It reports a feeling and leaves nothing anyone can check.

What this means for you: treat your own sense of having corrected for something as the least reliable signal in the room, and put the correction somewhere it can be inspected.

Where the evidence actually stands

A few effects are solid, the catalogue around them is loose, and the fixes are weaker than the diagnosis. Three numbers set the scale.

The table below is a map of confidence, not a scoreboard.

Read the middle column before the label. A row marked limited is not one to ignore, but one to act on gently.

Fix the process, not the brain

Since you cannot reliably retrain the judgment, change the conditions it is made under. Teaching helps a little: pooling 54 randomized education trials in 2025, Swaryandini and colleagues measured a small average improvement on bias tests, judged every included study to carry a risk of bias that was unclear or high, and left open whether the gain shows up outside the classroom.¹⁵ One field study is more encouraging. At HEC Paris, graduate students played a bias-training game, then met an unannounced business case: 59 percent of the trained group chose the inferior, hypothesis-confirming solution against 72 percent of the untrained, a 19 percent relative reduction. Training dates followed the school timetable rather than random assignment.¹⁷

Three process habits have better-tested roots than “try harder”.

Write the criteria before you see the options, and collect judgments independently before anyone speaks. Anchoring is difficult to avoid even when people are forewarned, so the defence that works is sequence rather than vigilance: fix the standard while there is nothing to anchor to.¹² In a meeting the first number spoken becomes everyone’s starting point, and asking each person to write an estimate first costs about two minutes.

Argue the other side on purpose. Charles Lord, Mark Lepper and Elizabeth Preston found in 1984, in two experiments with 150 undergraduates, that instructing people to consider the opposite reduced biased judgment more than simply telling them to be fair and unbiased.¹⁶ The instruction is specific for a reason: “be objective” leaves you doing the same reasoning with a sterner face, while “what would make the other option the right one?” sends you looking for evidence you have not collected.

Start from how comparable cases turned out. In a geopolitical forecasting tournament that ran for two years, forecasters who were randomly assigned to a brief probability-training module, one that taught reference classes and corrected common biases, produced more accurate forecasts than the untrained.¹⁸ Public budgets use the same move. Bent Flyvbjerg’s 2008 paper describes the first practical use of reference class forecasting, on cost projections for large transport infrastructure investments in the UK, a method that sets the forecast from how comparable past projects actually performed rather than from the plan in front of you.¹⁹

None of this asks you to feel unbiased, which is the point. Each step leaves a trace a colleague can inspect: a written criterion, estimates made before the discussion, a list of cases that ended badly. A habit that survives checking beats a resolution that does not.

WiserHours

A decision process that does not rely on being unbiased

• Criteria written down before the options were seen
• Everyone wrote an estimate before anyone spoke
• Five comparable past cases looked up, with how they ended
• One person asked to argue the case against the favorite
• A date set for when the decision gets made, win or lose

Find this checklist online:
https://wiserhours.com/decision-making/cognitive-biases-explained/#checklist-cognitive-biases-explained-process

The bottom line

Cognitive biases are what mental shortcuts look like when the setting has changed underneath them, and mostly they are doing their job. Treat the famous names with care: anchoring and framing have survived large retests, willpower-as-fuel and broad choice overload have not, and nobody has found a general trait of being bias-prone. Since knowing about a bias does not disarm it, spend the effort on the process instead. Criteria before options, independent estimates before discussion, comparable cases before your own optimism, and a decision date set in advance.