Conjunction Fallacy¶

Detect a probability-judgment error by watching for a more detailed scenario being rated more probable than the simpler scenario it is a strict subset of — the signature of resemblance quietly standing in for probability.

Core Idea¶

The conjunction fallacy is the systematic violation, in human probability judgment, of the rule that a conjunction P(A and B) cannot exceed either of its components P(A) or P(B): people rate a specific compound event as more probable than one of its constituents when the compound is more representative of an available description. Tversky and Kahneman (1983) demonstrated it with the Linda problem: subjects given a description matching the stereotype of a politically engaged feminist ranked "Linda is a bank teller and active in the feminist movement" as more probable than "Linda is a bank teller," even though the conjunction is a strict logical subset of the latter. The mechanism is representativeness substitution — the question being answered is silently replaced: rather than estimating the probability of the conjunction, the judge estimates how well the conjunction matches the description, and the conjunction typically matches better because it adds the detail that fits the available stereotype. The substitution is not conscious; the individual has the subjective experience of answering a probability question while actually answering a resemblance question, and the resulting probability order inverts the correct one. The effect is not an artifact of linguistic ambiguity (replications using frequency formats and explicit subset instructions reproduce it, though frequency formats reduce its magnitude), nor of carelessness (it survives incentivization and persists in trained researchers judging within their domain). Its force is greatest when the constituent event (A) is low in representativeness while the conjunction (A and B) is high, producing the largest gap between the resemblance ranking and the probability ranking.

Structural Signature¶

Sig role-phrases:

the probability judge — a human estimating event probabilities, the substrate the fallacy is a property of
the subset pair — two events, a constituent A and a conjunction (A and B), where the conjunction is a strict logical subset of the constituent
the representative cue — a description or stereotype that the added detail B matches more strongly than its absence, the bait for substitution
the conjunction rule — the normative anchor P(A and B) ≤ P(A), the axiom the judgment is checked against
the representativeness substitution — silently answering "how well does this match the description?" instead of "how probable is this?", not consciously noticed
the resemblance-beats-probability inversion — the compound rated more probable than its superset because the added detail fits the cue better, inverting the correct order
the content-free ordering tell — the diagnostic fingerprint: more-specific rated above its less-specific superset, needing neither the stimulus content nor the numbers
the decomposition/frequency repair — restoring the correct order by separating the constituents (so P(A and B) ≤ P(A) becomes inescapable) or natural-frequency framing (making the nesting visible), the fixes that survive incentive and expertise where exhortation does not

What It Is Not¶

Not an artifact of linguistic ambiguity. The error is not that "bank teller" is read as "bank teller and not a feminist": replications using natural-frequency formats and explicit subset instructions reproduce the violation. The inversion survives once the pragmatic reading is controlled, so it is a genuine probability-judgment failure, not a misunderstanding of the question's wording.
Not carelessness curable by incentives or expertise. The effect persists under monetary incentives and in trained researchers judging within their own domain, so motivation- and expertise-based fixes are predicted to fail. Only representational changes — decomposing the conjunction, frequency framing — reliably restore the correct ordering; trying harder does not.
Not a lapse of logic. The judge has not forgotten that a subset cannot be more probable than its superset; they have substituted attributes, competently answering "how well does this match the description?" while believing they answered "how probable is this?" Reframing it as illogic misses that a different, easier question was answered correctly.
Not added detail making a scenario more probable. Each additional stereotype-fitting clause makes a scenario more representative and more vivid while making it a stricter subset and therefore less probable — the joint falls below any conjunct. Vividness should be reread as a probability-lowering move and a warning sign, not as added information that raises likelihood.
Not base-rate neglect. Both are representativeness-driven, but base-rate neglect is a prior ignored in an inference, whereas the conjunction fallacy is a violation of the subset ordering between a compound and its constituent. The discriminating test is whether the offending ranking survives decomposition; a remedy tuned to one is not licensed for the other.
Not a flaw in probability itself. The conjunction rule P(A and B) ≤ P(A) is a theorem that holds in every substrate without exception; an algorithm or well-calibrated estimator that computes probabilities correctly exhibits no fallacy at all. What is violated is the human judgment, not the rule — invoking "the conjunction fallacy" for a sound estimator is a category error.

Scope of Application¶

The conjunction fallacy lives in one domain — human probability judgment — and the contexts below are application settings of that single substrate (a judge ranking compound-event probabilities under a descriptive cue), not structurally distinct systems. The conjunction rule it violates is a theorem holding in every substrate, and representativeness is the human mechanism behind the deviation; both belong to those parents, so invoking "the conjunction fallacy" for a sound estimator is a category error and stays out of this map.

Judgment-under-uncertainty research — the canonical home: the Linda problem and its many replications across populations and stimulus/frequency formats.
Forensic, clinical, and intelligence judgment — fleshed-out threat or diagnostic scenarios are rated more probable than their own components, biasing risk and over-allocating to the specific defense the scenario implies.
Behavioral economics — compound-risk evaluation enters prospect-theory and ambiguity-aversion accounts of how people weigh combined events.
Forecasting and scenario-planning practice — the discipline of penalizing vivid detail and pricing each component separately (decompose-and-frequency-frame) is the applied corrective deployed in risk and planning workflows.

Clarity¶

Naming the conjunction fallacy does two things for a judgment researcher. First, it pins a specific normative anchor — the conjunction rule, P(A and B) cannot exceed P(A) — onto a class of judgments where intuition supplies no such constraint, converting a vague sense that "the detailed story felt more likely" into a checkable violation of a known axiom. Second, and more deeply, it separates two things the introspecting mind cannot tell apart: a probability judgment and a resemblance judgment. The whole force of the Linda problem is that the question asked (how probable) and the question answered (how well it matches the description) feel identical from the inside, and the label is what lets the analyst say which one actually drove the ranking. That distinction reframes the error as not a lapse of logic but a substitution of attributes — the judge competently answered the wrong question.

This makes available a sharp, content-free diagnostic that needs no axiom-checking in the moment: whenever a more specific, more detailed scenario is rated more probable than a less detailed one that contains it, the conjunction rule has been violated and resemblance has displaced probability. The practitioner can now read added narrative detail not as added information but as a probability-lowering move that nonetheless raises felt plausibility — so the very vividness that makes a forecast or threat scenario compelling becomes the warning sign. The sharper question is no longer "is this scenario plausible?" but "have I let representativeness stand in for probability, and does the ordering survive decomposing the conjunction into its constituents?"

Manages Complexity¶

Compound-probability judgments are otherwise an open audit: each forecast, threat scenario, or diagnosis can go wrong in its own way, and checking the conjunction rule against every multi-clause story would mean recomputing joint probabilities case by case. The conjunction fallacy compresses that audit to a single content-free ordering test — does adding detail raise the rated probability of a scenario that the detail makes a strict subset? Whenever it does, representativeness has displaced probability, and the judge has answered a resemblance question while feeling they answered a probability one. The analyst no longer needs the stimulus content, the axiom, or the numbers in the moment; the diagnostic collapses the Linda problem, a vivid cyber-attack scenario, a fleshed-out clinical story, and an over-specified intelligence estimate to one regularity (more detail, higher felt plausibility, lower true probability) and one corrective family (decompose into constituents, frequency-frame, penalize narrative detail). It turns "is this scenario sound?" — a high-dimensional plausibility judgment — into a one-bit check on whether the probability ordering survives stripping the added clauses, with added vividness reread as a probability-lowering move rather than evidence.

Abstract Reasoning¶

The conjunction fallacy licenses a set of moves anchored on the subset relation between a conjunction and its constituents. Diagnostic: confronted with a ranking in which a more detailed scenario is judged more probable than a less detailed one that strictly contains it, the analyst infers that a resemblance judgment was substituted for a probability judgment — the judge competently answered "how well does this match the description?" while believing they answered "how probable is this?" The surface signature (added narrative detail raising felt plausibility) is read as the fingerprint of representativeness displacing probability, and crucially this diagnosis needs no access to the stimulus content or the actual numbers: the ordering violation alone, more-specific rated above the less-specific-superset, is sufficient to localize the error. A second diagnostic runs forward from a representative cue: when a constituent is low in representativeness while the conjunction that adds a stereotype-fitting detail is high, the analyst predicts the gap between the resemblance ranking and the probability ranking will be largest, and the inversion most severe.

Interventionist: the corrective moves are predictions about what restores the correct ordering. Decompose the conjunction into its constituents and ask the judge to rank or estimate each alone — the prediction being that the subset constituent must then be rated at least as probable as the compound, because the rule P(A and B) ≤ P(A) becomes inescapable once the parts are separated. Reframe the same problem in natural-frequency form ("of 100 people fitting this description, how many are X; how many are X-and-Y?") — predicted to shrink the effect, since the frequency frame makes the nesting of the smaller set inside the larger one visible. And in scenario-building practice, penalize vivid detail rather than reward it: each added clause is reread as a probability-lowering move (the joint falls below any conjunct) even as it raises persuasiveness, so the intervention is to strip elaboration and price each component separately. To shift the judgment, then, one changes the form in which the compound is presented — decomposed, frequency-framed, detail-penalized — rather than exhorting the judge to be more careful, since the substitution is not a lapse of care.

Boundary-drawing: the concept fixes where it applies and where look-alikes do not. It is the right diagnosis specifically for compound events standing in a subset relation, where the more-specific is rated above the more-general — not for base-rate neglect, where a prior is ignored in an inference, nor for mere narrative plausibility absent the conjunction structure. The diagnostic test that draws the line is whether the offending ordering survives decomposition: if stripping the added clause leaves the constituent rated below the compound, the conjunction rule was violated and representativeness was at work; if the ordering corrects once the parts are separated, the structure is present. Its robustness also bounds the available remedies: because the effect persists under incentivization and in trained researchers judging within their own domain, motivation- and expertise-based fixes are predicted to fail, and only the representational fixes (decompose, frequency-frame) reliably bite. Predictive: wherever an inquiry or forecast leans on a richly specified, stereotype-fitting scenario, the analyst can forecast in advance that its felt plausibility will exceed its true probability and that decomposition will lower its ranking — reading added vividness as a warning sign before any calibration data are gathered.

Knowledge Transfer¶

Within human probability judgment the fallacy transfers as mechanism, intact, with the caveat that its "domains" are application contexts of one substrate — a human judge ranking compound-event probabilities under a descriptive cue — not structurally distinct systems. With that understood, the content-free ordering test (does adding detail raise the rated probability of a scenario the detail makes a strict subset?) and its corrective family (decompose into constituents, natural-frequency framing, penalize vivid detail) carry without translation across forensic, clinical, and intelligence judgment (fleshed-out threat or diagnostic scenarios rated more probable than their components, biasing estimates of risk and over-allocating to the specific defense the scenario implies) and behavioral economics (compound-risk evaluation in prospect-theory and ambiguity accounts). The vocabulary (conjunction rule, representativeness substitution, resemblance-versus-probability), the diagnostic (more-specific rated above its less-specific superset is the fingerprint, needing neither the stimulus content nor the numbers), and the interventions (decompose and price each constituent; frequency-frame to make the nesting visible; reread added narrative detail as a probability-lowering move) all move freely, because the same resemblance-displaces-probability machinery is at work in each.

Beyond the human judge the fallacy does not travel — and the reason is sharp, paralleling base-rate neglect. The conjunction rule it violates — P(A and B) ≤ P(A) — is a logico-mathematical theorem that holds in every substrate without exception, so an algorithm, a formal system, or a well-calibrated market that computes probabilities correctly exhibits no fallacy at all; there is nothing for representativeness to substitute. So what is substrate-general here is the parent — probability and its conjunction axiom, the structure the fallacy violates — not "the conjunction fallacy," which names specifically the human deviation from that structure. The deviation's own mechanism is representativeness / attribute substitution — a second parent that genuinely recurs across human heuristic cognition (it also drives base-rate neglect and other resemblance-for-probability errors) — but it, too, is a feature of heuristic reasoners, not of probability-bearing systems in general. So invoking "the conjunction fallacy" for a non-human estimator is a category error rather than a stretched analogy; when the cross-domain lesson is wanted, carry the conjunction rule (the theorem any sound estimator obeys) and, for the error itself, representativeness substitution (the heuristic any human judge is prone to) — while the fallacy's own cargo (the Linda problem, the vivid-scenario warning, the decompose-and-frequency-frame remedies, the persistence under incentive and expertise) stays bound to human probability judgment. It must also be kept distinct from its representativeness sibling base-rate neglect (a prior ignored in inference) and from mere narrative plausibility absent the subset structure: the discriminating line is whether the offending ordering survives decomposition. So: as mechanism the fallacy stays inside human judgment; the conjunction rule travels everywhere as the violated theorem and representativeness as the human mechanism; carry those parents, not the named fallacy (see Structural Core vs. Domain Accent).

Examples¶

Canonical¶

The defining instance is Tversky and Kahneman's Linda problem (1983). Subjects read a description of Linda: thirty-one, single, outspoken, a philosophy major deeply concerned with discrimination and social justice — a portrait built to match the stereotype of a feminist activist. They then ranked several statements by probability, among them "Linda is a bank teller" and "Linda is a bank teller and is active in the feminist movement." The large majority — around 85 percent in the original naive samples — rated the conjunction as more probable than the single constituent. But the set of feminist bank tellers is a strict subset of bank tellers, so its probability cannot exceed the whole: the ranking inverts the conjunction rule. What subjects actually did was rank the two statements by how well each fits Linda's description, where the feminist detail scores higher.

Mapped back: The two Linda statements are the subset pair, and Linda's activist portrait is the representative cue. Rating the conjunction above its superset is the resemblance-beats-probability inversion, produced by the representativeness substitution — answering how well it matches rather than how probable — in violation of the conjunction rule.

Applied / In Practice¶

Tversky and Kahneman also ran the effect on professional forecasters at an international forecasting congress, a field case with real experts making the kind of judgment their work depends on. One group was asked the probability of "a complete suspension of diplomatic relations between the USA and the Soviet Union, sometime in the coming year"; another group the probability of "a Russian invasion of Poland, and a complete suspension of diplomatic relations between the USA and the Soviet Union, sometime in the coming year." The forecasters rated the second, more detailed scenario as more probable than the first — yet it is a strict subset of it, since every world with both events is a world with the suspension. The added invasion clause supplied a causal, stereotype-fitting story that raised felt plausibility while lowering true probability, and expertise did not immunise against it.

Mapped back: The bare-suspension versus invasion-plus-suspension pair is the subset pair; the causal invasion story is the representative cue. Experts rating the richer scenario higher is the content-free ordering tell — more-specific above its superset — and its persistence among trained forecasters shows why only the decomposition/frequency repair, not expertise, corrects it.

Structural Tensions¶

T1: Content-free diagnostic versus the subset structure it requires (a clean test with a narrow gate). The fallacy's great practical gift is a diagnostic needing neither the stimulus content nor the numbers: whenever a more-specific scenario is rated above the less-specific superset that strictly contains it, the conjunction rule is violated. But the diagnostic bites only where a genuine subset relation holds, and most real vivid-scenario reasoning does not present clean nested conjunctions — the added detail may condition rather than intersect, or the "superset" may be only loosely implied. The tension is that the test's content-freedom, which makes it so portable, depends on a structural precondition that field cases often blur, so the diagnostic either under-applies (missing resemblance-driven errors that lack the subset form) or over-applies (branding any detailed scenario a conjunction fallacy). The discriminating move — does the ordering survive decomposition? — is exactly what separates a true conjunction violation from mere narrative plausibility. Diagnostic: Is the more-detailed scenario a strict logical subset of the simpler one, or is the added detail conditioning information that does not nest inside it?

T2: Fallacy versus competent answer to a different question (is it even an error?). The concept's deepest reframing is that the judge has not committed a logical blunder — they competently answered "how well does this match the description?" while believing they answered "how probable is this?" But that very reframing strains the word "fallacy": if the mind reliably substitutes a tractable resemblance question for an intractable probability one, the output is a correct answer to a substituted question, not a mistake in reasoning. The tension is that labelling it a fallacy imports a normative verdict (you violated an axiom) onto what the mechanistic account describes as an adaptive substitution (you answered the answerable question). Both framings are in the entry, and they pull opposite ways on whether the judge did something wrong or something sensible under bounded resources. Diagnostic: Is the ranking being judged against the probability question the judge meant to answer (a fallacy) or the resemblance question they actually answered (a competent substitution)?

T3: Penalize vivid detail versus detail as real information (the corrective that can discard evidence). The scenario-planning corrective rereads each added clause as a probability-lowering move — vividness as a warning sign, not information — which correctly counters the fallacy's pull. But taken as a blanket rule it overshoots: in genuine forecasting, added detail sometimes is informative, narrowing a reference class or supplying a real conditioning fact, and a discipline that reflexively strips every elaboration will throw away legitimate evidence along with representativeness bait. The tension is that "detail lowers probability" is exactly true for a conjunction (the joint falls below any conjunct) yet misleading as a general heuristic about scenarios, where some detail earns its keep. The concept's corrective and sound conditioning both operate on "added detail," and confusing the two either reintroduces the fallacy or over-flattens real analysis. Diagnostic: Does the added clause make the scenario a stricter subset (lowering probability) or supply conditioning information that legitimately updates it — and is the corrective penalizing the right one?

T4: Robust to motivation versus responsive only to representation (the finding's double edge). The effect's persistence under incentives and in trained experts judging their own domain is a major strength of the result — it shows the fallacy is not carelessness and cannot be exhorted away. But the same robustness is a discouraging boundary on remedy: motivation, training, and expertise are exactly the levers organisations reach for, and all are predicted to fail, leaving only representational fixes (decompose, frequency-frame) that must be built into the task form rather than the judge. Even those only reduce the effect — frequency framing shrinks but does not abolish it. The tension is that what makes the finding scientifically clean (immunity to incentive and skill) is precisely what makes it practically stubborn: you cannot fix it by trying, only by re-presenting, and even then incompletely. Diagnostic: Is the proposed fix changing the form of the judgment (decomposition, frequency frame — can work) or the effort/expertise of the judge (predicted to fail)?

T5: Autonomy versus reduction (its own named bias, the theorem it violates, and the heuristic that violates it). The conjunction fallacy is a named, canonically studied bias with proprietary cargo — the Linda problem, the vivid-scenario warning, the decompose-and-frequency-frame remedies, the persistence under incentive and expertise — all bound to a human judge. But two different things travel beyond it, in opposite directions. The conjunction rule P(A and B) ≤ P(A) is a logico-mathematical theorem holding in every substrate, so a sound estimator or calibrated market exhibits no fallacy at all — invoking "the conjunction fallacy" for an algorithm is a category error, not a stretched analogy. The deviation's mechanism is representativeness / attribute substitution, a parent that recurs across human heuristic cognition (also driving base-rate neglect, its sibling). The tension is between a standalone bias that earns its own name and the recognition that its portable content splits into the theorem any estimator obeys and the heuristic any human is prone to. Diagnostic: Resolve toward the conjunction rule (the violated theorem) and representativeness substitution (the human mechanism) when asking what carries beyond human judgment; toward the named conjunction fallacy when diagnosing a specific person's compound-probability ranking in situ.

Structural–Framed Character¶

The conjunction fallacy sits at the framed-leaning end of the structural–framed spectrum — kept off the pure pole by a genuine structural credential, that it names a real and reliable regularity of heuristic cognition rather than a social convention, but held on the framed side by an unusually heavy evaluative load, a strict binding to a cognitive substrate, and portable content that lives in its parents rather than in the named fallacy. On evaluative weight it points framed, and hard: "fallacy" is one of the most normatively charged labels available, a flat verdict that a judgment is wrong, an axiom violated — the whole apparatus is organized around a normative anchor (the conjunction rule) against which the human judgment is convicted, a far cry from the value-neutral description a mechanism like feedback supplies. On human-practice-bound it points framed: the fallacy is by definition a property of a human judge, so it dissolves the instant the cognitive substrate is removed — a sound estimator or a calibrated market that computes probabilities correctly exhibits no fallacy at all, and invoking "the conjunction fallacy" for such a system is a category error, not a stretched analogy; the effect is pinned to a heuristic reasoner, though (unlike a rhetorical figure) not to any social institution. Institutional origin is mixed but leans framed: the underlying substitution of resemblance for probability is a real fact about how minds under bounded resources answer the easier question, which no survey invented — but the concept, with its Linda problem, its frequency-format replications, and its decompose-and-frequency-frame remedies, is furniture of judgment-under-uncertainty research. On vocab travels it points framed: the operative vocabulary (representativeness substitution, the subset pair, the vivid-scenario warning) is keyed to a human judge ranking compound-event probabilities and does not float free of that substrate. And on import versus recognize it patterns as import/category-error when stretched: what recurs off-substrate is not the fallacy but its parents, so any off-domain "conjunction fallacy" imports the label rather than recognizing the named phenomenon.

Two things pull structural here, and neither is the fallacy itself. The mechanistic skeleton it instantiates is representativeness / attribute substitution — silently answering "how well does this match the description?" in place of "how probable is this?", a resemblance-for-probability swap that recurs across human heuristic cognition (it also drives base-rate neglect, its sibling); and the normative anchor it deviates from is the conjunction rule P(A and B) ≤ P(A), a logico-mathematical theorem that holds in every substrate without exception and is as structural as content gets. Both are genuinely portable, which is what tempts a structural reading. But neither pulls the conjunction fallacy off the framed side, because they are precisely the parents — the heuristic any human is prone to and the theorem any sound estimator obeys — not what makes "the conjunction fallacy" itself travel: the cross-domain reach belongs to representativeness (as mechanism) and to the conjunction rule (as math), while the fallacy's distinctive content — the Linda problem, the reread of vivid detail as a probability-lowering warning sign, the decompose-and-frequency-frame remedies, and the persistence under incentive and expertise — is exactly the human-judgment accent that stays home. Its character: a heavily normative, cognition-substrate-bound error label, real enough to occur observer-free in a reasoning mind yet structural only as the gap between a substrate-independent probability theorem and the representativeness heuristic that overrides it, with its own furniture bound fast to human probability judgment.

Structural Core vs. Domain Accent¶

This section decides why the conjunction fallacy is a domain-specific abstraction and not a prime, and it carries the case for its domain-specificity — so it is worth being exact about what could lift and what stays home. Its skeleton is genuinely doubled, and the two halves travel in opposite directions: the fallacy deviates from the conjunction rule, a probability theorem, while the deviation is produced by representativeness / attribute substitution, a heuristic mechanism.

What is skeletal (could lift toward a cross-domain prime). Two abstract structures survive stripping the human judge. First, the conjunction rule — P(A and B) ≤ P(A), that a set cannot be less probable than a superset it is contained in — a logico-mathematical theorem that holds in every substrate without exception and is as structural as content gets; any sound estimator, formal system, or calibrated market obeys it. Second, the mechanism the fallacy instantiates: representativeness / attribute substitution — silently answering an easier resemblance question ("how well does this match the description?") in place of the intended one ("how probable is this?"). That substitution is a genuine, portable feature of bounded-resource cognition, recurring across human heuristic reasoning (it also drives base-rate neglect, the fallacy's sibling). Both are portable; but they are the cores the fallacy shares with its parents, not what makes it the conjunction fallacy.

What is domain-bound. Almost everything that makes the entry the conjunction fallacy in particular is human-judgment furniture, and none of it survives extraction. It is by definition a property of a probability judge: the Linda problem and its frequency-format replications; the subjective experience of answering a probability question while actually answering a resemblance one; the reread of added narrative detail as a probability-lowering move that nonetheless raises felt plausibility; the decompose-and-frequency-frame remedies; and the finding that the effect persists under incentive and expertise while yielding only to representational change. The decisive test: remove the human judge and there is nothing left to err — an algorithm that computes P(A and B) correctly exhibits no fallacy at all, so invoking "the conjunction fallacy" for a sound estimator is a category error, not a stretched analogy. What remains without the judge is only the theorem it should have obeyed and, one rung out, the heuristic a judge is prone to.

Why this does not clear the prime bar. A prime is a relational structure whose vocabulary travels and whose cross-domain transfer is recognition of the same mechanism, not analogy. The fallacy's transfer is bimodal. Within human probability judgment it travels intact — the content-free ordering test (more-specific rated above the superset it strictly contains) and its corrective family carry without translation across forensic, clinical, and intelligence judgment and behavioral economics, because the same resemblance-displaces-probability machinery runs in each. Beyond the human judge the fallacy does not travel at all: what recurs off-substrate is not the fallacy but its two parents, so any off-domain "conjunction fallacy" imports the label rather than recognizing the phenomenon. When the cross-domain lesson is needed it is already carried, in more general form, by those parents: the conjunction rule travels everywhere as the theorem any sound estimator obeys, and representativeness substitution travels across human heuristic cognition as the mechanism any judge is prone to. The cross-domain reach belongs to the parents; "the conjunction fallacy," as named, is the human deviation whose distinctive content — the Linda paradigm, the vivid-detail warning, the decompose-and-frequency-frame remedies — stays bound to human probability judgment.

Relationships to Other Abstractions¶

Current abstraction Conjunction Fallacy Domain-specific

Parents (3) — more general patterns this builds on

Conjunction Fallacy is a kind of Bias Prime

Conjunction fallacy is bias specialized to a systematic probability ranking displaced toward representative detail and away from set inclusion.
Conjunction Fallacy is part of Representativeness Heuristic Domain-specific

Conjunction fallacy contains representativeness because the more stereotype-matching scenario is ranked by resemblance instead of probability.
Conjunction Fallacy presupposes Probability Prime

Conjunction fallacy presupposes probability because its decisive error is assigning a strict subset more probability than an event that contains it.

Not to Be Confused With¶

Base-rate neglect. Its representativeness sibling: both are driven by resemblance standing in for probability, but base-rate neglect is a prior ignored in an inference, whereas the conjunction fallacy is a violation of the subset ordering between a compound and its constituent. The discriminating test differs — the conjunction fallacy is diagnosed by whether the offending ranking survives decomposition, base-rate neglect by whether a neglected prior is restored. Tell: Is a background rate being left out of an update (base-rate neglect), or is a more-detailed scenario rated above the simpler one that contains it (conjunction fallacy)?
The conjunction rule (the theorem it violates). The logico-mathematical fact that P(A and B) ≤ P(A), holding in every substrate without exception. The fallacy is not the rule but the human deviation from it; a sound estimator or calibrated market that computes probabilities correctly obeys the rule and exhibits no fallacy at all. Tell: Are you naming the axiom any correct system honours (the conjunction rule — this is what travels), or a judge's failure to honour it (the fallacy)?
Representativeness / attribute substitution (the mechanism parent). The heuristic of silently answering an easier "how well does this match?" in place of the intended "how probable is this?" This is the engine the fallacy runs on, and it recurs across human heuristic cognition (it also drives base-rate neglect). The conjunction fallacy is the specific case where that substitution produces a subset-ordering violation. Tell: Is the point the general resemblance-for-probability swap (representativeness — the portable mechanism), or its particular expression as a conjunction rated above its constituent (the fallacy)?
Mere narrative plausibility. The ordinary sense that a richly specified, coherent story "feels likely." Absent a genuine subset relation, a detailed scenario judged compelling is not the conjunction fallacy — the fallacy requires that the added detail make the scenario a strict subset of the simpler one, so that rating it higher violates the rule. Tell: Does the more-detailed scenario strictly nest inside the simpler one (conjunction fallacy), or is it just a vivid story with no superset it logically falls under (mere narrative plausibility)?
A sound estimator computing a joint probability. An algorithm or well-calibrated model that computes P(A and B) correctly is not "committing a small conjunction fallacy" — it obeys the theorem exactly, and invoking the fallacy for it is a category error, not a stretched analogy. The fallacy is a property of a human judge, not of any probability-bearing system. Tell: Is there a human experiencing a resemblance judgment as a probability judgment (fallacy), or a system mechanically computing the joint correctly (no fallacy — category error)?

Neighborhood in Abstraction Space¶

Conjunction Fallacy sits in a moderately populated region (49^th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.

Family — Unclustered & Miscellaneous (309 abstractions)

Nearest neighbors

Base Rate Fallacy — 0.87
Subadditivity Effect — 0.85
Belief Bias — 0.85
Cherry Picking — 0.84
Conservatism Bias — 0.83

Computed from structural-signature embeddings · 2026-07-12