The Conjunction Fallacy
Learn it → Apply it → Save as Skill
The conjunction fallacy is the mistake of judging a detailed, two-part story (A and B) as more probable than one of its own parts (A alone), even though logic guarantees the opposite: every case where both A and B are true is also a case where A is true, so "A and B" can never be more probable than "A" by itself. Psychologists Amos Tversky and Daniel Kahneman documented the pattern in their 1983 paper "Extension versus intuitive reasoning: The conjunction fallacy in probability judgment," published in Psychological Review (volume 90, issue 4, pages 293 to 315). The practical lesson for anyone reading a forecast, a pitch, a diagnosis, or a plan is that a more specific, more vivid story is not automatically a more likely one. Specificity can make a claim feel more convincing while quietly making it less probable.
Dr. Wallace's Favorite Mental Models turns one well-chosen way of thinking at a time into something you can actually use. Each entry gives you a fast overview, an interactive walkthrough, varied real and practical examples, and an AI-ready Markdown skill you can keep.
Instead of only reading about a mental model, you can build your own AI skills library over time: a personal set of portable reasoning tools that help ChatGPT, Claude, and other AI assistants apply the thinking you value.
The conjunction fallacy shows up whenever two conditions are joined with "and": Linda is a bank teller and a feminist; the stock misses earnings and drops hard; the suspect had a motive and was seen nearby. Adding the second condition narrows the set of cases that satisfy the claim, so the probability of the pair can only stay the same or go down, never up. The fix is not more intuition. It is checking the conjunction rule directly: write out the single, broader condition on its own, and compare.
Important: this is the documented result of Tversky and Kahneman's original Linda study and its frequency-format follow-up, not an invented illustration. Asking people how many of 100 women fitting Linda's description are bank tellers, versus how many are bank tellers who are also active feminists, pushes most people toward the correct comparison and sharply cuts the error rate. The fallacy is sensitive to how the question is framed, not only to raw reasoning ability.
Use the two axes below, how fast and intuitive the thinking is, and whether the pattern is a broad, general-purpose heuristic or one specific, provable logical error, to see how the conjunction fallacy relates to three commonly confused neighbors. Select a dot or a chip to read how each one compares.
Each stage builds on the one before it. Use the numbered buttons or Next to move through the walkthrough.
Open an example to see the conjunction fallacy in a specific context. The first four are documented research; the rest are common hypothetical scenarios used to teach the method, clearly labeled as such.
Linda is described as 31, single, outspoken, and very bright. She majored in philosophy, and as a student was deeply concerned with discrimination and social justice, and took part in anti-nuclear demonstrations. Asked which is more probable, "Linda is a bank teller" or "Linda is a bank teller and is active in the feminist movement," about 85% of respondents in the standard wording chose the second option, even though every feminist bank teller is also a bank teller, so the conjunction cannot be more probable than the single condition it contains.
Asked to picture tennis player Bjorn Borg reaching a Wimbledon final, participants on average rated "Borg will lose the first set but win the match" as more probable than simply "Borg will lose the first set," even though the first outcome is a narrower case nested entirely inside the second. If he loses the first set and wins the match, he has, by definition, also lost the first set.
Shown sequences of red (R) and green (G) outcomes from repeated rolls of a die, around 65% of participants judged the longer sequence GRGRRR as more likely to appear than the shorter sequence RGRRR, even though GRGRRR is simply RGRRR with one more roll (a G) added in front of it. Any run of rolls that matches GRGRRR necessarily contains a run that matches RGRRR, so the longer, more "pattern-like" sequence cannot be more probable than the shorter one it depends on.
Professional political and intelligence analysts were asked to rate the probability of "a Soviet invasion of Poland and a subsequent break in U.S.-Soviet diplomatic relations" versus the probability of "a break in U.S.-Soviet diplomatic relations" alone. The conjunction was rated at about 4% on average; the single broader condition was rated at only about 1%. Trained experts making real professional judgments proved just as vulnerable as the lay participants in the Linda study.
A hiring manager pictures a strong candidate so clearly that the compound story, she accepts and stays more than two years, feels more certain than the plain fact that she accepts at all. The detailed story is emotionally persuasive precisely because it is specific, which is also exactly why it cannot be more probable than the single condition it depends on.
An analyst's thesis, "the company misses earnings and the stock falls more than 20% within a week," reads as more rigorous than the bare claim "the stock falls more than 20% within a week," because it supplies a reason. The reason adds narrative confidence while mathematically it can only hold the probability flat or push it down, never up.
A product manager champions a feature because "power users will adopt it and it will cut churn," a specific, satisfying story, over the plainer, broader claim that "it will cut churn," which covers more ground and is less demanding to be true.
A clinician favors "this patient has condition X and it explains both the rash and the fatigue" over the plainer, broader "this patient has condition X," because the first story feels like it accounts for more of the picture, even though it is a stricter, less probable claim than the diagnosis alone.
A juror reasons that "the defendant had a motive and was seen near the scene" feels more like guilt than either fact alone, when formally the conjunction of two pieces of circumstantial evidence is never more probable than either piece taken by itself.
A marketer pitches a campaign around "a 34-year-old new parent who switches brands after one bad delivery experience," a vivid conjunction of traits, treating it as though it describes a more common buyer than the broader, single-trait segment "customers who switch brands after a bad delivery," which it logically cannot exceed.
This is an illustrative, invented scenario designed to teach the method. The figures are the analyst's own gut estimates, not documented research statistics.
Expand each item for the research and reasoning behind the model.
Amos Tversky and Daniel Kahneman documented it in their 1983 paper "Extension versus intuitive reasoning: The conjunction fallacy in probability judgment," published in Psychological Review, volume 90, issue 4, pages 293 to 315. The paper's centerpiece is the "Linda problem," but the finding generalized across many other experiments using unrelated material, from tennis predictions to die rolls to geopolitical forecasts.
The conjunction rule states that for any two conditions A and B, the probability of both being true, P(A and B), can never exceed the probability of A alone, P(A). Every situation satisfying "A and B" also satisfies "A" by itself, so the set of A-and-B cases is always a subset of, or equal to, the set of A cases. It is a rule with no exceptions, which is why the conjunction fallacy can be proven wrong outright rather than merely argued against.
Largely, yes. When the same Linda problem is reframed in frequency terms, asking people to imagine 100 women who fit Linda's description and estimate how many are bank tellers versus how many are bank tellers who are also active feminists, the error rate drops sharply, from roughly 85% choosing the conjunction under the standard wording to roughly 20% under the frequency wording. That gap shows the fallacy is sensitive to how a question is posed, not only to how carefully a person reasons.
No. A Cold War-era study of professional political and intelligence analysts found the same pattern in real forecasting work: experts rated "a Soviet invasion of Poland and a subsequent break in U.S.-Soviet diplomatic relations" at an average of about 4%, higher than their average rating of about 1% for "a break in U.S.-Soviet diplomatic relations" alone. Trained professionals making real judgments were vulnerable too.
No. The same research program found it with a 1980 study on predictions about tennis player Bjorn Borg, where a two-part outcome ("loses the first set but wins the match") was rated more likely than its own first half ("loses the first set"), and with sequences of die rolls, where about 65% of participants rated a longer six-roll sequence as more likely than the shorter five-roll sequence it required. Neither of those examples involves a social stereotype at all.
The representativeness heuristic is the general-purpose shortcut: judging how probable something is by how well it matches a mental prototype or stereotype, rather than by applying probability rules. The conjunction fallacy is one specific, distinct consequence of applying that shortcut to a joined "A and B" claim. Representativeness explains why the Linda conjunction feels right (it matches the stereotype better than "bank teller" alone does); the conjunction fallacy names the provable logical error that results. Not every use of representativeness produces a conjunction fallacy, and the conjunction rule, P(A and B) is always less than or equal to P(A), is what gives this particular error its own distinct normative fix.
Detail and vividness make a claim easier to picture and easier to believe. They do not make it more probable; adding conditions can only hold probability flat or push it down.
Representativeness is the broad shortcut of judging by resemblance to a stereotype. The conjunction fallacy is the specific, provable error that shortcut produces when a joined claim is compared to one of its own parts. They are related, not identical.
The Cold War forecasting study found trained political and intelligence analysts making the identical error in real professional judgments, not just students answering a survey.
The rule only guarantees P(A and B) is less than or equal to P(A) when B is genuinely an additional condition layered onto A. Two separate, non-nested claims being compared is a different question and the rule does not mechanically settle it.
Asking "how many out of 100" sharply cuts the error rate on that one question, but the underlying intuition to favor a vivid story returns on the next question unless the habit of checking is repeated.
The gap between roughly 85% errors under standard wording and roughly 20% under frequency wording, on the same people reasoning about the same facts, shows that how a question is framed matters as much as how carefully someone thinks.
A headline or pitch that silently adds a second condition, "and it will also...", can quietly turn a plausible broad claim into a narrower one, without flagging that the odds have necessarily gone down.
the conjunction rule compares "A" to "A and B," where every A-and-B case is also an A case. Comparing two separate, non-nested claims, say, two different single conditions, is an ordinary probability comparison, not a conjunction-fallacy question.
sometimes added detail is genuine evidence that should change your estimate of the broader claim too, not just decoration on top of it. The fallacy is about treating the conjunction as more probable than its own component, not about detail being inherently misleading.
naming a mechanism, timeline, or named cause is often exactly what makes a forecast useful and falsifiable. The lesson is to keep the probability bookkeeping honest alongside the specificity, not to avoid specificity altogether.
This browser tool structures your thinking. It does not send your information anywhere or perform AI analysis.
Save as Skill converts the model into portable Markdown instructions that teach an AI how to spot a hidden conjunction, apply the conjunction rule, ask the right follow-up questions, and avoid the common mistakes above.
Save the skills you find useful and organize them into your own reusable library. The Markdown is intentionally model-neutral, so it can be used wherever an AI system accepts uploaded files, project knowledge, custom instructions, or reusable text context. Platform capabilities differ, so this is a portable skill file rather than a claim of one-click installation.
Describe a specific claim, forecast, or decision you are facing. This builds a separate, personalized prompt that asks an AI to apply the conjunction fallacy and the conjunction rule directly to your situation, distinct from the prompt builder above.
No. It means a detailed, joined claim needs its probability checked against the plain, broader claim it depends on, rather than judged by how convincing or specific it feels. Detail can be useful and even diagnostic. It is just not, on its own, evidence of higher probability.
Asking "how many out of 100" forces a set-based, almost visual way of thinking, people picture 100 cases and have to sort them into bins, which makes the subset relationship ("bank tellers who are feminists" sits inside "bank tellers") much harder to miss. The standard probability wording invites a resemblance judgment instead, which is exactly where representativeness takes over.
Whenever you notice the word "and" joining two conditions in a forecast, pitch, diagnosis, or argument, write out the plain, broader claim with the second condition dropped, and ask whether you would still rate the joined claim as more probable than that plain version. If the joined claim is winning, the conjunction rule says your numbers cannot both be right.