Recipe · serves one confident error
Conjunction Fallacy
The conjunction fallacy is the error of judging a conjunction of two events (A and B) as more probable than one of its constituents (A alone), in violation of the conjunction rule of probability that P(A and B) can never exceed P(A).
Difficulty to avoid: strong grounding
Adapted · results varyIngredients · what the mind brings
Method · how they combine
Tversky and Kahneman (1983) explained the fallacy via the representativeness heuristic: people judge probability by how well the target (a feminist bank teller) resembles the stereotype evoked by the description, and the detailed conjunction resembles 'Linda' better than the bland constituent, so it feels more probable.
uses Representativeness heuristic
Availability and the coherence/plausibility of a detailed scenario contribute, especially in forecasting.
uses Base rate neglect
Note · At least four rival or supplementary accounts exist. (1) Conversational/semantic: Hertwig and Gigerenzer (1999) argue 'probable' and 'and' are polysemous and that respondents make sensible pragmatic inferences, so the 'error' partly reflects task ambiguity, not faulty reasoning. (2) Inductive confirmation: Tentori, Crupi and colleagues argue people track degree of confirmation (how much the evidence supports a hypothesis) rather than probability, which can rationally favor the conjunction. (3) Quantum-probability models (Pothos and Busemeyer) model order/interference effects; Tentori and Crupi (2013) argue these inadequately explain the data. (4) Probability-theory-plus-noise accounts treat the fallacy as a by-product of noisy probability estimates. Mellers, Hertwig and Kahneman (2001), an adversarial collaboration, found frequency formats alone did not eliminate the effect, partially adjudicating the dispute.
Tested variations
Linda problem (naive subjects). About 89% of statistically naive respondents ranked the conjunction (T&F) as more probable than its constituent (T) in the direct-comparison version, violating the conjunction rule.
Amos Tversky, Daniel Kahneman, 1983 · University of British Columbia and Stanford undergraduates (statistically naive group)
Linda problem (statistically sophisticated subjects). Statistical sophistication did not protect against the error; the conjunction violation rate among sophisticated subjects was about the same as among naive subjects.
Amos Tversky, Daniel Kahneman, 1983 · Graduate students in decision science / sophisticated respondents
Diplomatic-relations forecasting study. The more detailed conjunction (invasion + suspension) was rated significantly more probable than the simpler suspension alone, despite logically being a subset, showing the fallacy in real-world expert forecasting of unique events.
Amos Tversky, Daniel Kahneman, 1983 · 115 professional analysts employed by industry, universities, or research institutes
Frequency-format reduction. Conjunction violation rates fell sharply under frequency phrasing, which the authors argued reflected polysemy of 'probable' and conversational inference rather than a stable reasoning error.
Ralph Hertwig, Gerd Gigerenzer, 1999 · University student samples (multiple experiments)
Tested rating — replication
The basic phenomenon replicates very robustly under the classic single-event probability framing across student, expert, and statistically trained samples. However, its size and interpretation are genuinely contested: Hertwig and Gigerenzer (1999) showed frequency phrasing dramatically reduces violation rates, and the Mellers, Hertwig and Kahneman (2001) adversarial collaboration found frequency formats by themselves did not eliminate conjunction effects, though effects disappeared with certain disambiguated phrasings once filler items were removed. The error is reliable but partly format-dependent, and the mechanism remains disputed.
Substitution that changes the dish
Reframe the judgment in explicit frequencies and check set inclusion before estimating. Ask 'out of 100 people like this, how many are bank tellers? how many are bank tellers AND feminists?' and draw the nested sets: the conjunction is always inside the constituent, so it can never be larger. Frequency framing and forcing direct comparison of the two options measurably reduces the error.
Hertwig and Gigerenzer (1999) showed frequency formats sharply lower violation rates; the Mellers, Hertwig and Kahneman (2001) adversarial collaboration confirmed format and disambiguation matter, though they do not fully eliminate the effect.
Reader submissions · seen in the wild
Russia-invades-Poland forecasting demonstration, Second International Congress on Forecasting · 1982
In July 1982, 115 professional forecasters at the Second International Congress on Forecasting rated a detailed conjunction ('a Russian invasion of Poland, and a complete suspension of US-Soviet diplomatic relations in 1983') as significantly more probable (p < .01) than the broader event ('a complete suspension of US-Soviet diplomatic relations in 1983') that logically contained it. This is a real, dated demonstration that professional analysts commit the fallacy on genuine future events.
Oil-price scenario forecasting demonstration · 1982
In 1982, Kahneman presented academic forecasters with a prediction. One group judged 'oil consumption will decrease by 30 percent'; another judged 'a dramatic rise in oil prices will lead to a 30 percent reduction in oil consumption.' The second, more detailed causal-conjunction scenario was found more compelling/probable, despite being logically narrower, illustrating the fallacy in economic forecasting. (Widely recounted, e.g., by Dobelli; the underlying conjunction-scenario effect is documented in Tversky & Kahneman's forecasting work.)
Tasting note
Watch for moments when adding vivid, specific detail to a forecast or scenario makes it feel MORE likely. If a richer story ("the company fails because a key executive quits and a lawsuit hits") seems more probable than the plainer version it is nested inside ("the company fails"), you are likely committing the conjunction fallacy. The tell is that the more detailed, more representative-sounding option is rated higher even though it can only be a subset of the simpler one.