Cognitive bias

What separates a cognitive bias from an ordinary mistake

A cognitive bias differs from an ordinary mistake in three ways: it has a direction, it repeats, and it is invisible from the inside. Direction means the error leans. Repetition means the same conditions reproduce it in a new group of people. Invisibility means the person making the error feels no different from a person getting it right, which is why you cannot catch a bias by checking whether you feel confident.

The modern research program that made this measurable belongs to Amos Tversky and Daniel Kahneman, whose 1974 paper in the journal Science, "Judgment under Uncertainty: Heuristics and Biases", set out the pattern that a small number of mental shortcuts produce a large number of predictable errors. Their claim was not that people are stupid. It was that the shortcuts are efficient, that they work well in the environments they evolved for, and that they fail in specific, mappable ways. That is a much more useful claim, because a mappable failure can be tested for.

A cognitive bias is therefore not a character flaw and not a synonym for prejudice in the everyday sense. Prejudice about a group is one thing a biased judgment can be about. The category itself is broader and duller: it covers how you estimate a price, how you remember a forecast you made, how confident you are in a skill you have just started learning.

Why cognitive biases exist, and why the mind keeps them

Cognitive biases exist because the alternative is slower and often worse. Four pressures produce them, and none of them is a defect on its own:

• Speed. A judgment that arrives in half a second and is right four times in five beats a calculation that arrives in ten minutes and is right nine times in ten, if you are making the judgment forty times an hour.
• Missing information. Most real decisions are made with a fraction of the relevant facts, so the mind substitutes an easier question for the hard one. Asked how risky something is, you answer how easily an example comes to mind. That substitution is the availability heuristic.
• Limited working memory. Holding six competing considerations at once is genuinely hard, so the mind compresses them into a single impression, and the impression then colors everything attached to it.
• Coherence. The mind prefers a story that hangs together to a set of facts that do not, and it will quietly drop the facts that spoil the story.

Heuristics are the shortcuts themselves. Biases are the shape of the errors those shortcuts make. That distinction matters, because it explains why you cannot simply switch the shortcuts off: the same machinery that misjudges a probability also lets you cross a road, read a face and finish a sentence.

A cognitive bias is not a logical fallacy

A cognitive bias and a logical fallacy are different objects, and the two words are not synonyms even though a great deal of writing treats them as one. A bias is a fact about a mind. A fallacy is a fault in an argument, and it stays a fault whether or not anyone believes it. You can state a fallacy on paper, with nobody in the room, and it is still invalid. You cannot state a bias on paper; you have to observe it in judgments.

Deductive reasoning is the standard the third row is measured against: an argument is valid when the conclusion cannot be false if the premises are true. No amount of psychology settles that question, and no amount of logic settles the second row. Keeping the two apart is the difference between saying "your argument does not work" and saying "you are the kind of person who would think that", and only the first is worth anyone's time.

How many cognitive biases are there, and why the number is not a fact

The number of cognitive biases usually quoted is around 180, which comes from the Cognitive Bias Codex, a widely reproduced diagram built by John Manoogian III from a list compiled on Wikipedia. Treat that figure as a filing count, not a scientific one. The list contains entries that are the same effect under two names, entries that are special cases of a broader effect, and entries whose supporting evidence is a single study that has not been repeated. Nobody has ever produced an agreed taxonomy the way chemistry produced the periodic table.

There is a further reason to hold the number loosely. The Open Science Collaboration, a large group of psychologists who in 2015 published in Science an attempt to repeat 100 studies from three major journals, found that fewer than half of the replications reproduced the original statistically significant result. Some famous effects in this field survived that scrutiny comfortably; others did not. A responsible list of biases is therefore ranked by how well each one has held up, not alphabetized as though all entries carried equal weight.

Ten cognitive biases worth knowing by name

Ten named biases cover most of what a reader meets in practice, and each one has a giveaway you can look for:

• Confirmation bias. Searching for evidence that fits the belief you already hold, and testing it only in ways that could confirm it.
• Anchoring. Letting the first number you hear set the range for every number that follows, including numbers it has nothing to do with.
• The availability heuristic. Judging how common something is by how easily an instance comes to mind, which favors the vivid over the frequent.
• The Dunning Kruger effect. The finding that people who score lowest on a test tend to overestimate their score by the widest margin.
• Survivorship bias. Drawing a lesson from the cases that made it into your sample while the cases that failed are, by construction, absent from it.
• Hindsight bias. Remembering an outcome as more predictable than it was, which quietly destroys your ability to learn from it.
• The halo effect. Letting one favorable trait raise your rating of unrelated traits in the same person, product or organization.
• The sunk cost fallacy. Continuing an unpromising project because of what has already been spent on it, which is money and time no future decision can recover.
• Framing. Answering differently depending on whether the same quantity is presented as a gain or a loss.
• Self serving bias. Attributing successes to your decisions and failures to circumstances, which makes every outcome flattering.

Naming a bias is not the same as proving one. "That is just confirmation bias" is a diagnosis, and a diagnosis of the person is not a reply to the claim. If someone says a bridge design is unsafe and their reasoning is biased, the bridge may still be unsafe. The label tells you where to look for an error. It does not tell you that one is there.

What actually reduces a cognitive bias

To reduce a cognitive bias, change the procedure rather than the effort, because trying harder to be objective has a poor record and changing the conditions of the judgment has a good one. Five procedures do most of the work:

1. Write the prediction down first, with a date and a number, before the outcome is known.
2. Fix the criteria before you see the candidates, whether the candidates are job applicants, suppliers or study designs.
3. Assign someone the disconfirming case, so that at least one person in the room is looking for the evidence that would sink the plan.
4. Count the cases you cannot see, by asking what would be missing from this data set even if the effect were entirely fake.
5. Compare against a base rate, not against the most memorable example you can recall.

Notice that four of the five are things you do to a process, not to your mind. That is deliberate. Bias training that consists of telling people about biases tends to produce people who can name biases, which is a different skill from making better judgments.

The test to run on any judgment you are about to trust

The test to run on any judgment you are about to trust is one question, asked of your own conclusion: what would I expect to see if this were false, and have I looked for it? If you cannot describe the observation that would change your mind, you are not holding a conclusion, you are holding a commitment, and the difference does not show up from the inside.

A second question catches the sampling problems the first misses: who is not in this data, and would they have been left out precisely because they disagree with it? Run both on the next confident claim you meet, including one of your own, and you will have done more than any list of 180 names can do for you.

Where to go next

BIASESThe Dunning Kruger effectThe Dunning Kruger effect, and what the study actually showedBIASESConfirmation biasDefinition and examplesREASONINGDeductive reasoningDefinition and examples