Confusing the probability of the evidence given innocence with the probability of innocence given the evidence.
Imagine you know that almost all dogs wag their tails when happy. You see a creature wagging its tail in the bushes. 'It must be a dog!' you say. But lots of other animals wag their tails too, and there are way more non-dogs in the world than dogs. Just because happy dogs almost always wag doesn't mean that tail-wagging almost always means dog. Mixing up those two things is the Prosecutor's Fallacy.
The Prosecutor's Fallacy occurs when someone conflates two fundamentally different conditional probabilities: the probability of observing certain evidence assuming a hypothesis is true, and the probability that the hypothesis is true given that the evidence has been observed. In legal settings, this manifests as equating the rarity of a forensic match among innocent people with the probability that a matching defendant is guilty. The fallacy ignores the prior probability (base rate) of the hypothesis being true before the evidence was introduced, which can drastically change the correct conclusion. Although named for its prevalence in prosecution arguments, the error appears across medicine, epidemiology, machine learning, and everyday interpersonal reasoning whenever conditional probabilities are transposed without applying Bayes' theorem.
The same glitch looks different depending on the terrain. Finance, medicine, a relationship, a team — same mechanism, different costume.
Fraud detection systems flag transactions based on patterns that match known fraud, but because fraudulent transactions are rare compared to legitimate ones, the vast majority of flagged transactions are false positives. Investigators who treat every flag as near-certain fraud waste resources and may freeze innocent accounts, confusing the probability of matching a fraud pattern given innocence with the probability of fraud given a match.
Clinicians and patients frequently misinterpret positive screening results for rare conditions, assuming a positive test means near-certain disease. Because the base rate of many screened conditions is very low, even highly accurate tests produce far more false positives than true positives in the general population, leading to unnecessary anxiety, invasive follow-up procedures, and overtreatment.
When standardized tests identify students as 'gifted' or 'at-risk' based on score thresholds, educators sometimes treat the classification as definitive. They confuse the test's sensitivity (how well it detects the trait if present) with the predictive value (how likely the trait is present given the score), leading to misplacement of students, especially in large or diverse populations where base rates vary.
People invert conditional probabilities in interpreting social signals — for example, assuming that because distant behavior is common when someone is upset, any distant behavior must mean that person is upset. This ignores the many other explanations for the behavior and the low base rate of actual conflict, creating unnecessary anxiety and accusations.
Machine learning classifiers are often evaluated by accuracy or recall, but when deployed to detect rare events (spam, malware, anomalous behavior), even small false-positive rates generate floods of false alerts. Teams that focus on the model's accuracy rate without considering the base rate of the target event can grossly overestimate the reliability of positive predictions, leading to alert fatigue and wasted engineering effort.
HR analytics that flag employees as flight risks based on behavioral patterns may confuse the probability of matching the pattern given an employee stays with the probability of leaving given a match. Because most employees stay, flagging systems can overwhelm managers with false alarms, damaging trust and wasting retention resources.
Media coverage of rare but dramatic events (terrorism, mass shootings) can lead the public to commit this fallacy: because nearly all such events involve a certain profile, people assume that anyone matching the profile is dangerous, ignoring that the overwhelming majority of people fitting the profile are harmless. This drives discriminatory profiling policies and disproportionate fear.
William C. Thompson and Edward L. Schumann coined the term in their 1987 paper 'Interpretation of Statistical Evidence in Criminal Trials: The Prosecutor's Fallacy and the Defense Attorney's Fallacy,' published in Law and Human Behavior.
In small ancestral environments with low population sizes and high base rates for threats, the difference between 'if predator then tracks' and 'if tracks then predator' was negligible — both directions of inference were roughly equivalent because the relevant populations were tiny and threats were common. The brain evolved to process simple covariation rather than formal probability inversion, which was rarely needed when base rates were intuitively known through direct experience.
Machine learning classifiers trained to detect rare events (fraud, disease, security threats) are highly susceptible to this fallacy when their outputs are interpreted. A model with 99% accuracy detecting a 0.01% base-rate event will generate overwhelmingly false positives, yet operators routinely treat positive predictions as near-certain. Additionally, when ML models are evaluated retrospectively on datasets where the target event has already been selected for, the resulting performance metrics can be inflated — a direct analog of selecting only cases that experienced a transition and then claiming predictive power.
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.
