A cognitive bias is a systematic deviation of judgement from a norm (logic, probability, well-understood self-interest). It concerns the person who is reasoning: it is involuntary, and it affects those who know about it too.
A fallacy concerns an argument: its structure or its content cannot establish the conclusion, even though it looks as if it does. French tradition distinguishes the paralogism, committed in good faith, from the sophism in the strict sense, used to deceive.
The two lines are exactly the same length, and knowing it changes nothing about what you see. Biases work the same way: knowing about them is not enough to protect you from them. Measure this illusion →
The two are linked: a bias often makes a fallacy convincing. The argument from authority works because we are inclined to defer to authority; the appeal to popularity, because we tend towards conformity.
Bridges: when a bias makes a fallacy convincing
A fallacy doesn’t persuade by chance: it often leans on a natural tilt of judgement. Knowing the bias helps you spot the argument that exploits it, and vice versa. Each arrow links a bias entry to a fallacy entry.
Belief bias → Fallacy fallacy: both confuse the quality of an argument with the truth of its conclusion.
Two modes of thinking
A common description distinguishes two ways of processing information:
System 1: fast, automatic, effortless, associative. It recognises a face, or answers “10 cents” to the bat-and-ball question.
System 2: slow, deliberate, demanding of attention. It carries out a calculation or checks a piece of reasoning.
In this model, most biases appear when a System 1 answer is accepted without being checked by System 2. The terms come from Stanovich and West (2000) and were popularised by Daniel Kahneman (Thinking, Fast and Slow, 2011). It is a simplifying model, not a description of two regions of the brain.
Biases are the flip side of heuristics, shortcuts that quickly give an answer that is often right. Gerd Gigerenzer stresses the point: in their usual environment, these shortcuts can be as efficient as a full calculation, or even more so.
Buster Benson (2016) sorts about 180 biases by the four problems they answer:
Too much information: we filter, keeping what is recent, repeated, surprising or in line with our expectations.
Not enough meaning: we fill the gaps with patterns, stereotypes and assumptions.
The need to act fast: we cut through, favouring confidence, the status quo and the short term.
Limited memory: we simplify and generalise, at the cost of forgetting and reconstructing.
John Manoogian III drew a widely shared illustration from this classification, the Cognitive Bias Codex.
Valid, true, sound
An argument is valid if its conclusion follows necessarily from its premises. Truth belongs to statements; validity belongs to the link between them. A valid argument whose premises are true is called sound.
“All cats fly; Felix is a cat; therefore Felix flies” is valid, but not sound.
“If it rains, the road is wet; the road is wet; therefore it rained” is not valid, even though its conclusion may be true.
Formal fallacies are errors of structure, which you can spot by replacing the terms with letters (see affirming the consequent). Informal fallacies are errors of content or relevance: the form may be correct, but a premise is false, irrelevant, or assumes the conclusion.
Examining a claim in six questions
What exactly does it say? Restate it, pin down the terms, ask for the absolute numbers behind the percentages. (equivocation, framing)
The principle of charity (Neil Wilson, 1959; Donald Davidson) means interpreting what the other person says in its most coherent version before criticising it. It is the direct defence against the straw man.
The steel man goes further: restate the opposing position in its strongest version, then answer that.
Rapoport’s rules, reported by Daniel Dennett (Intuition Pumps, 2013), for criticising a position:
re-express it so clearly, vividly and fairly that its author says, “Thanks, I wish I’d thought of putting it that way”;
list the points of agreement;
mention anything you have learned from it;
only then, say what you want to criticise.
Finally, naming a fallacy is not enough: a badly built argument can defend a true conclusion (the fallacy fallacy).
Ask rather than guess. In 25 experiments, making an effort to take another person’s perspective did not make people’s guesses about that person’s thoughts and preferences any more accurate; asking them did (Eyal, Steffel and Epley, Journal of Personality and Social Psychology, 2018).
Correcting is still worthwhile. There was a fear that a correction would strengthen the belief it targets: the “backfire effect”. Across 52 topics and more than 10,100 participants, Wood and Porter (Political Behavior, 2019) find none; neither do the three large experiments by Guess and Coppock (British Journal of Political Science, online in 2018), when participants were shown information running against their opinion.
Tone is not the argument. Paul Graham ranks the ways of expressing disagreement, from name-calling to refuting the central point (“How to Disagree”, 2008). It is an essay, not a study; his ranking is a reminder that a polite reply can dodge the substance, and that a sharp one can address it.
A few principles, forged by philosophers, that are useful every day for sorting claims.
Ockham’s razor
Between two explanations that account equally well for the facts, prefer the one that assumes the least. The principle is attributed to William of Ockham (14th century), who never actually wrote the famous formula “entities must not be multiplied beyond necessity”. It is a rule of prudence, not a proof: the simplest explanation can be wrong.
The burden of proof
The one who makes a claim has to supply the evidence; it is not for others to prove them wrong. Bertrand Russell illustrated this with an invisible teapot orbiting the Sun (1952); Christopher Hitchens put it this way: “What can be asserted without evidence can also be dismissed without evidence.”
Evidence proportionate to the claim
“Extraordinary claims require extraordinary evidence” (Carl Sagan, 1979). The idea is already in Laplace in 1814: the more a claim contradicts what we know, the stronger the evidence has to be.
Falsifiability
For Karl Popper (1934), a scientific theory must be open to contradiction by a possible observation. A theory that everything confirms, whatever happens, teaches nothing: that is a defect, not a strength.
The problem of induction
David Hume (1739): no amount of past observation logically guarantees the next one. We generalise all the same, and reasonably so; but a regularity remains a hypothesis, not a certainty.
Methodical doubt
René Descartes (Discourse on the Method, 1637): hold as true only what cannot be doubted. In small doses, this is the heart of critical thinking; pushed to the extreme, it leaves nothing to believe. The realistic goal is not certainty but confidence proportioned to the evidence.
Knowing what you do not know
The line “I know that I know nothing” does not appear as such in Plato. In the Apology, Socrates says rather that he is wiser than another man in this one respect: he does not think he knows what he does not know. Knowing the limits of your knowledge is a starting point.
Eight checks to make before a choice that is costly or hard to undo.
The state of the research
Since the 2010s, psychology has been going through a “replication crisis”. In 2015, the Open Science Collaboration tried to reproduce 100 psychology studies published in 2008: only about a third gave a comparable significant result.
Some widely publicised effects, such as social priming or “ego depletion”, are now strongly contested and are not included here. The entries kept are about well-established effects; when a result is disputed (the Dunning-Kruger effect, the Hofling experiment, the illusion of control, omission bias…), the entry says so.
Glossary
The site’s key words, in three groups; each term links, where there is one, to the entry or workshop that puts it into practice.
Reasoning
Cognitive bias
A systematic error of judgement: it happens to almost everyone, in the same direction, and knowing about it is not enough to avoid it. The bias entries.
Fallacy
Reasoning that looks valid but isn’t, because of its form or its content. The fallacy entries.
Heuristic
A simple, fast rule of judgement, often effective, but a source of systematic errors in certain situations.
Premise, conclusion
The premises are the statements an argument starts from; the conclusion is what it claims to establish.
Valid argument
An argument whose conclusion necessarily follows from its premises: if they are true, so is the conclusion. Validity is about form, not about whether the premises are true.
Sound argument
A valid argument whose premises are true: its conclusion is then established. Valid, true, sound.
Burden of proof
The obligation, for whoever puts forward a claim, to justify it. Appeal to ignorance.
Falsifiability
A criterion proposed by Karl Popper (1934), also called refutability: a scientific claim must be open to contradiction by an observation.
Ockham’s razor
A principle attributed to William of Ockham (14th century): between two explanations that account equally well for the facts, prefer the one that makes fewer assumptions.
Probability and statistics
Base rate
How frequent a phenomenon is in the population, before any particular information about the case at hand. Base rate neglect.
Conditional probability
P(A | B): the probability of A given B. It is generally different from P(B | A): confusing the two is the prosecutor’s fallacy.
False positive, false negative
A false positive is a positive test in someone who doesn’t have what is being looked for; a false negative, a negative test in someone who does.
Sensitivity, specificity
Sensitivity is the share of real cases a test detects; specificity, the share of healthy cases it recognises as such.
Positive predictive value
The probability of really having what is being looked for when the test is positive. It depends on the base rate: for a rare disease, even a good test gives many false positives. “Positive tests” workshop.
Correlation
A statistical link between two variables that vary together. It is not enough to establish a cause-and-effect relationship. Correlation mistaken for causation.
Confounder
A variable that influences both the supposed cause and the observed effect, creating an apparent link between them.
Regression to the mean
After an extreme measurement, the next one is usually closer to the average, with no cause at work. The entry.
Calibration
Agreement between stated certainty and actual success: someone well calibrated is right about 7 times out of 10 when they say they are 70% sure. “Measuring your confidence” workshops.
Brier score
The average of the squared differences between the stated probabilities and what happened (1 or 0): the lower it is, the better the forecasts.
Studies and evidence
Representative sample
A subset whose make-up reflects that of the population being studied.
Selection bias
A distortion that comes from how the observed cases were chosen: they do not represent the whole. Survivorship bias.
Control group
A group comparable to the one receiving an intervention, but which does not receive it, to measure what would have happened without it.
Randomised trial
A study in which the intervention is assigned by lot: confounders, known or unknown, are spread at random between the groups and balance out on average.
Statistical significance
A result is called significant when a difference at least that large would be unlikely to be observed if the effect did not exist. It says nothing about the size or the importance of the effect.
Replication
Reproduction of a study by other researchers, to check that the result holds. The state of the research.
Meta-analysis
A statistical synthesis of several studies on the same question.
Publication bias
The tendency to publish positive results more readily than null ones, which distorts the overall picture. Cherry-picking.
Further reading
Daniel Kahneman, Thinking, Fast and Slow, Farrar, Straus and Giroux, 2011.
Daniel Kahneman, Olivier Sibony and Cass Sunstein, Noise: A Flaw in Human Judgment, Little, Brown Spark, 2021.
Hans Rosling, Factfulness, Flatiron Books, 2018.
Normand Baillargeon, A Short Course in Intellectual Self-Defense, Seven Stories Press, 2007.
Arthur Schopenhauer, The Art of Being Right (published 1864), various editions.
Gérald Bronner, La Démocratie des crédules, PUF, 2013 (in French).
Gerd Gigerenzer, Gut Feelings: The Intelligence of the Unconscious, Viking, 2007.
Annie Duke, Thinking in Bets, Portfolio, 2018.
Your data
Your progress (cards learned and read, quiz answers, reviews, workshop results, paths, mastery, points and distinctions) is saved only in your browser. By default, it is not sent anywhere: no account required, no audience tracking. If you turn on sync between devices (Profile, Settings), a copy, encrypted on your device, is kept online; no one but you can read it. A copy is made every day, in this same browser (see below). Details: privacy policy. The site is hosted by GitHub Pages, which, like any host, receives requests for pages, but never your answers.
Titles set in Newsreader (Production Type), under the SIL OFL licence.
If you clear your browsing data, your progress disappears. To keep it, or to carry it over to another device (from computer to phone, for instance), save it to a file, then open that file from the other device.
Offline: once loaded, the site works without a connection, in the language you visited. To keep the other language too, use “Keep everything for offline use” in Profile, Settings; the files are only saved on your device.
Challenge links (“Challenge a friend”, custom challenge) contain only numbers: the challenge number, the chosen entries and, if you share it, your score. Never your name or your progress.
How your mastery is calculated
Mastery (from 0 to 1,000) says how well you do in this site’s games. It is calculated entirely from the results already saved in your browser; it can go up or down, but never because you stayed away.
A scored game. For each level played: the average success of your last five rounds, corrected for chance (with four answers, guessing gives 25%: only what is above counts), multiplied by confidence (a single round counts for a third, three rounds or more in full), under a ceiling that rises with difficulty: 500 at level 1 and 1,000 at the game's highest level (750 at level 2 of a three-level game; 625, 750 and 875 at levels 2, 3 and 4 of a five-level game). The game keeps the best of these values. Example: 85% success with four answers, over three rounds, gives 400 at level 1 and 800 at level 3 of a three-level game.
Special measures. Calibration: the average gap between your confidence and your success (no gap: 1,000; 30 percentage points or more: 0). Anchoring: the gap between your estimates after a high and a low anchor (no gap: 1,000; 50% or more: 0). Forecasts: the Brier score of your resolved forecasts, that is, the average of the squared gaps between the probability you gave and what happened (1 or 0). A score of 0 gives 1,000; 0.25, the score of someone who always says “50%”, gives 0.
Concepts: 80% for the quiz review boxes, 20% for duels between close concepts. Each correct answer in the quiz moves the concept up one box, from 1 to 5, unless you answered “Guessing”; every concept in box 5 gives 800.
Six domains: the world in numbers, measuring your confidence, perceiving, reasoning (with philosophising), debating, and concepts. A domain’s mastery is the average of its games, a game never played counting as zero: you cannot inflate your score by playing only what you are good at. The axes of your latest full Grand test each count as a game of the nearest domain, except impartiality, an exploratory indicator. Overall mastery is the average of the six domains; it gives your rank, and a rank once reached is never lost.
Its limits. It measures success in this site’s games, not intelligence nor critical thinking in life. It depends on how many games you play. It compares you with no one: the site has no one else’s data. And a number you aim for soon stops being a good measure: playing for the score rather than to understand would raise it without teaching you anything.
How this site tries to motivate you
This site teaches you to spot the mechanisms that influence us. It uses a few of them itself to make you want to come back: here is which ones, and what it refuses to do.
What it uses. Clarity points, a daily goal, a streak of days, ranks and distinctions. They reward what the site teaches: stating your true confidence (the quiz scoring makes “Guessing” pay better than “Sure” when you are in doubt), reviewing at the right time, spotting a concept in everyday life. A ring that fills up plays on the goal gradient (people speed up as they approach a goal: Kivetz, Urminsky and Zheng, 2006); points granted for past progress play on the endowed progress effect (a head start increases effort: Nunes and Drèze, 2006). A streak you do not want to lose plays on loss aversion; points you have accumulated, on the endowment effect and sunk costs.
What research says. A meta-analysis of 40 experiments finds positive but modest effects of gamification: on learning (g = 0.49), motivation (0.36) and behaviour (0.25). The g measures the gap with groups without gamification, in standard deviations: 0.49 is roughly half of one. Its authors present this as a provisional picture of the question (Sailer and Homner, 2020). Another meta-analysis concludes that rewards promised for an activity already found interesting can reduce interest in it, while positive feedback increases it (Deci, Koestner and Ryan, 1999); this result is still debated.
What the site refuses to do. No notifications or reminders. No punishing streak: one missed day a week does not break it, and there is no reproach. No points for time spent, sharing or creating an account. No ranking between people. And a plain mode, in the profile settings, hides all this: the site becomes what it was, and everything is still counted in case you change your mind.
Sailer, M. and Homner, L. (2020). The gamification of learning: a meta-analysis. Educational Psychology Review, 32(1), 77-112. Deci, E. L., Koestner, R. and Ryan, R. M. (1999). Psychological Bulletin, 125(6), 627-668. Nunes, J. C. and Drèze, X. (2006). The endowed progress effect. Journal of Consumer Research, 32(4), 504-512. Kivetz, R., Urminsky, O. and Zheng, Y. (2006). The goal-gradient hypothesis resurrected. Journal of Marketing Research, 43(1), 39-58.
How this site is secured
What is protected. Your progress, which may contain personal notes, and the integrity of the site: a tampered page could mislead those who trust it.
How. The site loads only its own files, no script from any other site. It never inserts text as code: your browser refuses to do so anyway (Trusted Types). An imported backup file is checked item by item before being used. The site refuses to be displayed inside another site’s page, and does not pass on the address of the pages you read to the sites its sources link to. Before each release, automatic checks verify all this, and nothing is published if they fail.
Sync is optional: your progress is encrypted on your device before being sent, with a key derived from your code, which the server never sees. No one else can read it, not even the site’s owner. The downside: a lost code cannot be recovered.
What is not protected. What is saved in your browser is not encrypted there: a malicious extension or someone using your device can read it. An exported backup file is in plain text; it is in your keeping. And the site’s host, like any host, sees visitors’ IP addresses.
The details, for those who want to check: the site’s threat model (file SECURITE.md in the source code). To report a vulnerability: use GitHub’s private reporting.