Judging two events happening together as more likely than either event alone, violating basic probability.
Imagine you have a big jar of marbles. Some are blue, some are big, and a few are both blue AND big. If I ask you 'Is it more likely you'll pick a blue marble or a blue big marble?', your brain might trick you into saying the blue big marble because it sounds more specific and interesting—but there are always at least as many blue marbles as there are blue big marbles, because every blue big marble is also a blue marble.
The conjunction fallacy occurs when people estimate that a combination of two characteristics or events (A and B) is more probable than one of those characteristics or events alone (A), directly violating the conjunction rule of probability theory. This error is driven by the representativeness heuristic—the more specific and detailed a description is, the more it matches a mental prototype, making it feel more plausible even though each added condition mathematically reduces the overall probability. The fallacy is remarkably robust: even statistically trained professionals, including graduate students in decision science, commit it at rates above 80%. It reveals a deep tension between narrative coherence (how well a story fits together) and extensional logic (how probability actually works), with the human mind strongly favoring coherence.
The same glitch looks different depending on the terrain. Finance, medicine, a relationship, a team — same mechanism, different costume.
Investors and analysts frequently rate detailed market scenarios (e.g., 'a recession caused by rising interest rates and a housing market collapse') as more probable than the broader outcome alone ('a recession'), leading to overconfidence in specific forecasts and misallocation of hedging resources.
Clinicians may judge a detailed diagnostic narrative (e.g., 'the patient has lung cancer caused by asbestos exposure at their former workplace') as more probable than the simpler diagnosis ('the patient has lung cancer'), potentially distorting differential diagnosis and treatment prioritization.
Teachers constructing assessment scenarios may inadvertently create test questions where a detailed answer option seems more correct because it is more descriptively coherent, leading students who rely on narrative fit rather than logical analysis to select probabilistically impossible answers.
People tend to find elaborate explanations for a partner's behavior (e.g., 'they're distant because they're stressed at work AND dealing with a family issue') more believable than a single explanation, leading to overcomplicated interpretations that may miss simpler truths.
Product teams may judge compound risk scenarios for system failures (e.g., 'server outage caused by a DDoS attack during a holiday traffic spike') as more likely than the broader event ('server outage'), leading to over-investment in narrow contingency plans while neglecting general resilience.
In performance reviews, managers may rate a detailed negative narrative about an employee's shortcomings (combining multiple specific failings into one story) as more representative of reality than any single observed weakness, inflating perceived underperformance.
Detailed political predictions (e.g., 'the candidate will lose because of a scandal AND low voter turnout among young people') are rated as more probable than the simple prediction ('the candidate will lose'), making elaborate media narratives disproportionately persuasive to the public.
Amos Tversky and Daniel Kahneman, 1982 (initial report) and 1983 (seminal paper: 'Extensional Versus Intuitive Reasoning: The Conjunction Fallacy in Probability Judgment,' Psychological Review, 90(4), 293–315).
In ancestral environments, pattern-matching and prototype-based reasoning were fast, efficient survival tools. Judging whether a situation 'fit' a known dangerous pattern (e.g., 'predator hiding in tall grass near water') was more adaptive than computing exact probabilities. The brain evolved to prioritize coherent causal stories because scenarios that made narrative sense were often better guides to action than abstract statistical reasoning, which was rarely needed in small-group foraging contexts.
LLMs such as GPT-3 and GPT-3.5 have been shown to exhibit the conjunction fallacy when presented with Linda-type problems, rating conjunctive options as more probable than their constituents. Research indicates that RLHF fine-tuning can actually increase susceptibility to this fallacy, as models trained to produce human-pleasing outputs absorb the representativeness heuristic embedded in human language patterns. More advanced models like GPT-4 show reduced but not eliminated susceptibility, and small changes in entity names or prompt wording can cause models to revert to fallacious reasoning, suggesting pattern-matching rather than genuine logical comprehension.
50 cards are free to preview. Buyers unlock the rest of the deck plus the interactive training — Spot-the-Bias Quiz unlimited, Swipe Deck with spaced repetition, My Blindspots, Decision Pre-Flight, the Printable Deck + Cheat Sheets, and the Field Guide e-book. $29.50$59.
