Cognitive Biases
Human reasoning relies on heuristics — fast, frugal shortcuts that work well most of the time but produce systematic errors under predictable conditions. Cognitive biases are those systematic errors. They are not random noise; they are patterned departures from normative reasoning that can be anticipated, diagnosed, and partially corrected. This skill catalogs the twelve most consequential biases, each with a mechanism, a diagnostic signal, and a mitigation strategy.
Agent affinity: tversky (heuristics and biases tradition), kahneman-ct (System 1 / System 2 framing), paul (integration with elements of reasoning)
Concept IDs: crit-confirmation-bias, crit-availability-anchoring, crit-intellectual-humility, crit-calibrated-confidence
The Bias Catalog at a Glance
| # | Bias | Mechanism | Diagnostic signal |
|---|---|---|---|
| 1 | Confirmation bias | Seek and weight supporting evidence more than disconfirming | "I knew it" for every matching case; disconfirming cases feel like "exceptions" |
| 2 | Availability heuristic | Judge probability by how easily examples come to mind | Vivid recent events dominate risk estimates |
| 3 | Anchoring | First number or idea biases subsequent estimates | Second guess is close to the first even with new information |
| 4 | Representativeness | Judge by resemblance to a stereotype, ignoring base rates | Ignoring how rare the category actually is |
| 5 | Framing effects | Same content, different phrasing, different choices | Preferences flip when the same option is described as gain vs. loss |
| 6 | Hindsight bias | Past events feel inevitable after the fact | "It was obvious" retrospectively; no one predicted it |
| 7 | Overconfidence | Confidence intervals are too narrow relative to accuracy | 90% confidence intervals contain the truth ~50% of the time |
| 8 | Motivated reasoning | Conclusion drives evidence evaluation, not vice versa | Evidence quality is judged leniently for favored conclusions |
| 9 | Sunk cost fallacy | Past investment justifies continued investment | Continuing a failing project because "we've already spent so much" |
| 10 | Fundamental attribution error | Others' behavior attributed to character; one's own to circumstance | "They're incompetent" vs. "I was having a bad day" |
| 11 | In-group favoritism | Judgments tilt toward one's own group | Same behavior is praised in allies, criticized in opponents |
| 12 | Base rate neglect | Ignoring prior probabilities in favor of new information | Updating too strongly on a single diagnostic result |
Bias 1 — Confirmation Bias
Mechanism. Once a hypothesis is entertained, the mind preferentially searches for, attends to, remembers, and weighs evidence that supports it. Disconfirming evidence is overlooked, dismissed, or explained away.
Worked example. A researcher convinced that a particular herb cures headaches runs a trial. When the trial shows no effect, she attributes this to "impure samples." When a later trial shows a small effect, she counts this as confirmation. Over twenty trials, the balance is null — but her belief strengthens throughout.
Mitigation — the disconfirmation pass. Before committing to a conclusion, list three observations that would falsify it, then actively search for each. If you cannot find a way the claim could be wrong, you do not understand the claim.
Mitigation — the pre-mortem. Imagine the project has failed. What is the most likely reason? This reframes confirmation into disconfirmation without feeling threatening.
Bias 2 — Availability Heuristic
Mechanism. Probability is estimated by the ease with which instances come to mind. Vivid, recent, or emotionally charged events are over-weighted; common but unremarkable events are under-weighted.
Worked example. After a high-profile plane crash, travelers estimate air travel as more dangerous than driving, even though per-mile statistics show driving is approximately 60 times more dangerous. The crash is vivid; driving fatalities are individually unmemorable.
Mitigation — consult base rates. Before judging probability from memory, check whether actual frequency data exists. The memory is a sample from attention, not from reality.
Mitigation — reverse the question. Instead of "how risky is X?" ask "how many times did X happen last year per million attempts?"
Bias 3 — Anchoring
Mechanism. Numerical estimates are biased toward any anchor value available at the moment of estimation, even when the anchor is arbitrary or irrelevant.
Worked example. A classic study asked participants to write down the last two digits of their social security number, then estimate the price of a bottle of wine. People with higher digits consistently gave higher estimates. The digits had zero informational value, yet they shifted estimates by 50-100%.
Mitigation — multiple independent estimates. Generate the estimate twice, using different starting points, then average.
Mitigation — consider the opposite. Explicitly ask "why might my estimate be too high?" and "why might it be too low?" before committing.
Bias 4 — Representativeness Heuristic
Mechanism. Probability is judged by how much an instance resembles a stereotype of the category, rather than by the base rate of the category.
Worked example (Tversky & Kahneman, 1983). Linda is 31, single, outspoken, and deeply concerned with issues of discrimination and social justice. Which is more probable?
- Linda is a bank teller.
- Linda is a bank teller and active in the feminist movement.
Most people pick (2), but (2) is a conjunction of (1) with another claim, so P(2) <= P(1) by the conjunction rule. Representativeness overrode probability.
Mitigation — apply the conjunction rule. A more specific claim is always less probable than its components.
Mitigation — consult base rates. How many bank tellers are there? How many active feminists? The base rates are what actually determine the answer.
Bias 5 — Framing Effects
Mechanism. Preferences change when the same choice is described differently — typically, emphasizing gains vs. losses leads to different risk preferences (loss aversion).
Worked example. A disease will kill 600 people without intervention. Option A saves 200 for sure. Option B has 1/3 chance of saving all 600 and 2/3 chance of saving none. Most people pick A. Now reframe: Option C kills 400 for sure, Option D has 2/3 chance of killing all 600 and 1/3 chance of killing none. Most people pick D. But C = A and D = B.
Mitigation — translate between frames. Whenever you encounter a choice, restate it in both gain and loss frames. If your preference changes, the framing is driving you, not the content.
Mitigation — state outcomes in absolute terms. "200 saved out of 600, 400 die" is frame-neutral.
Bias 6 — Hindsight Bias
Mechanism. After an event occurs, the mind reconstructs prior probabilities to make the actual outcome seem more likely than it was. "I knew it all along."
Worked example. Before an election, experts give a candidate a 40% chance of winning. After the candidate wins, the same experts remember themselves as having predicted 65%. The revision is unconscious.
Mitigation — write down predictions in advance. Any calibration system (prediction markets, forecast journals, probability logs) defeats hindsight bias by making the original prediction concrete and audit-able.
Mitigation — study outcomes with the ex ante information only. When analyzing a past decision, ask what was known at the time, not what we know now.
Bias 7 — Overconfidence
Mechanism. People assign higher confidence to their beliefs than accuracy warrants. 90% confidence intervals contain the truth in roughly 50% of cases in typical studies.
Worked example. An expert is asked to provide a range for the population of a foreign country such that they are 90% sure the true value is in