Judging the whole as less likely than the sum of its parts — individual pieces adding up to more than 100%.
Imagine someone asks you how many toys you have. You say 'maybe 20.' But then they ask, 'How many action figures? How many stuffed animals? How many Legos? How many puzzles?' When you count each type separately, you end up saying numbers that add to 35. The pieces somehow feel like more than the whole, just because you thought about each one by itself.
When people estimate the likelihood of a broad category (e.g., 'dying from natural causes'), they tend to assign it a lower probability than the combined probabilities they would give to its specific subcategories (e.g., cancer, heart disease, stroke, other natural causes) when those subcategories are listed explicitly. This violates basic probability axioms, since the whole must logically equal the sum of its mutually exclusive parts. The effect is driven by the fact that explicitly listing subcategories—'unpacking' them—increases their psychological salience and makes each feel more concrete and imaginable, thereby inflating their individual probability estimates. The degree of subadditivity is robust across domains and has been replicated consistently in studies involving health risks, financial forecasts, legal judgments, and everyday frequency estimates.
The same glitch looks different depending on the terrain. Finance, medicine, a relationship, a team — same mechanism, different costume.
Investors and analysts tend to assign higher cumulative risk to a portfolio when individual risk factors (interest rate changes, currency fluctuation, sector downturns, regulatory changes) are listed separately than when asked about 'overall market risk.' This leads to over-hedging or excessive diversification in response to itemized risks that seem larger in aggregate than the packed whole.
Patients asked about the likelihood of 'side effects' from a medication give lower estimates than patients who are presented with a list of specific side effects (nausea, headache, dizziness, fatigue). This can lead to inflated risk perception and medication non-adherence when side effects are itemized in detail on packaging or during informed consent.
Students estimating the probability of failing a course give a lower figure than when asked separately about failing due to poor exam scores, missed assignments, low participation, and a bad final project. This can lead to underestimation of overall academic risk when threats are considered holistically, or panic when they are itemized.
When someone considers the general question 'Will this relationship end?' they estimate low probability. But when they separately consider incompatible life goals, communication problems, financial stress, and family conflicts, the itemized probabilities sum to a much higher figure, potentially amplifying anxiety about the relationship.
Product teams listing specific failure modes (server crash, data corruption, API timeout, authentication error) during risk reviews produce inflated total risk estimates compared to assessing 'system failure' as a single category. Conversely, presenting users with a single 'security' label feels less comprehensive than listing 'encryption, two-factor authentication, fraud detection, and privacy controls,' influencing perceived product value.
During performance reviews, listing specific shortcomings (missed deadlines, poor communication, low initiative, weak technical skills) makes an employee's overall performance seem worse than a summary statement of 'needs improvement.' Managers who unpack negatives may unknowingly inflate the perceived severity of an employee's issues.
Political campaigns and media outlets exploit this effect by itemizing specific threats (terrorism, cyberattacks, pandemics, economic collapse) rather than referring to 'national security risks' in the abstract. The unpacked list makes the total threat seem more imminent and severe, which can be used to justify policy positions or generate engagement.
Amos Tversky and Derek J. Koehler, 1994, in their paper 'Support Theory: A Nonextensional Representation of Subjective Probability' published in Psychological Review.
In ancestral environments, survival depended on detecting and responding to specific, concrete threats rather than abstract categories of danger. A brain that becomes more alert when hearing 'snake, spider, and scorpion' than when hearing 'dangerous animals' was better at mobilizing targeted defensive responses. This bias toward concreteness ensured that vivid, identifiable risks received disproportionate attentional resources, even if it distorted holistic probability assessment.
AI systems trained on human-generated probability estimates can inherit subadditive patterns if training data reflects human judgments rather than calibrated statistical outputs. Recommendation and risk-scoring algorithms that decompose categories into subcategories may produce inflated cumulative risk scores. Additionally, when LLMs are asked to estimate probabilities of events, they may exhibit subadditivity if they generate estimates for parts independently without enforcing normalization constraints, producing outputs where subcategory probabilities sum to more than the whole.
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.
