Ellsberg Paradox

Two urns sit in front of you, each holding one hundred balls. In the first, you can see the count: exactly fifty red and fifty black. About the second you are told only that it contains red and black balls in some unknown mix. You will win a prize if a ball of your chosen color is drawn. Betting on red, nearly everyone prefers to draw from the first urn. Betting on black, nearly everyone prefers the first urn again. Each choice feels obviously sensible, yet together they are impossible to justify: if you think red is likelier in the known urn, you must think black is likelier in the unknown one, and should want to switch. This is the Ellsberg paradox, and it exposes something the Allais paradox does not: not merely that people misuse probabilities they are given, but that for some kinds of uncertainty people act as though no probabilities exist at all. This page presents both classic versions of the problem, shows why no beliefs of any kind can rationalize the majority pattern, separates risk from ambiguity, and follows the debate over whether the aversion the paradox reveals is a bias to be corrected or a reasonable response to missing information. Three interactive tools let you take the choices yourself, measure your own ambiguity premium, and watch three rival theories price the same ambiguous bet.

The Ellsberg paradox is a set of choice problems demonstrating that people's preferences under uncertainty systematically violate the axioms of subjective expected utility, the framework Leonard Savage had recently established as the foundation of rational choice and Bayesian decision theory (Ellsberg, 1961). Its author, Daniel Ellsberg, was a young RAND Corporation analyst, a decade away from the Pentagon Papers fame that would make him a household name, and his paper aimed its title directly at the theory's architect: risk, ambiguity, and the Savage axioms. The paradox turns on the distinction, drawn by Frank Knight, between risk, where probabilities are known, and genuine uncertainty, where they are not (Knight, 1921). Savage's theory had seemed to erase that distinction by deriving personal probabilities from coherent preferences, so that a rational agent treats every uncertainty as risk (Savage, 1954). Ellsberg's problems show that real preferences refuse the reduction: people reliably prefer betting on known chances over unknown ones in a pattern that no assignment of probabilities, however subjective, can produce. That pattern, now called ambiguity aversion, launched one of the largest research programs in decision science, spanning formal models of non-additive belief, the psychology of what people know about their own knowledge, and the neuroscience of uncertainty (Camerer & Weber, 1992; Machina & Siniscalchi, 2014). The sections below present the two experiments, the proof that beliefs cannot rationalize them, the risk-ambiguity distinction, the psychology and boundary conditions of ambiguity aversion, and the theories built to take missing probabilities seriously.

Two Urns and a Question

Ellsberg's first problem is the one in the opening paragraph, and it repays a slow look. Urn one is transparent in the relevant sense: fifty red balls, fifty black, so a bet on either color is a known fifty-fifty gamble. Urn two holds one hundred balls in an unknown mixture of red and black; the proportion could be anything from all red to all black. A ball will be drawn from the urn you choose, and you win a fixed prize if its color matches your bet. Asked to bet on red, most people choose urn one. Asked to bet on black, most people choose urn one again. Ellsberg presented the problems as thought experiments rather than a formal study, and the pattern he described has since been confirmed repeatedly in incentivized experiments with real payoffs (Ellsberg, 1961; Halevy, 2007). The pattern's charm is that each choice alone is easy to defend and the pair is indefensible. Preferring the known urn for red says, in effect, that you treat red in the unknown urn as having less than an even chance. Preferring the known urn for black says the same about black. But red and black are the only colors in the urn: their probabilities must sum to one, so they cannot both be below one half. A chooser with any definite belief about the unknown urn, cautious or hopeful, precise or vague, should favor the known urn for at most one of the two colors, and should strictly want to switch for the other. What the majority pattern reveals is not a belief about the urn but an attitude toward not knowing: the unknown urn is penalized as such, whichever color is at stake.

The Three-Color Urn and the Sure-Thing Principle

Ellsberg's second problem sharpens the point into a direct strike on the axiom at the heart of Savage's system. A single urn contains ninety balls: exactly thirty are red, and the remaining sixty are black and yellow in an unknown mixture. Two pairs of bets are offered, each paying a fixed prize. In the first pair, you may bet on red, a known thirty-in-ninety chance, or on black, whose chance could be anything from zero to sixty in ninety; most people bet on red. In the second pair, you may bet on red-or-yellow, an ambiguous chance, or on black-or-yellow, a known sixty-in-ninety chance; most people bet on black-or-yellow. Figure 1 lays the four bets out as a payoff matrix, and the structure it exposes is exact: within each pair, the two bets pay identically on yellow, zero in the first pair and the full prize in the second, and differ only on red and black. The yellow column is a common consequence, and the sure-thing principle, the central postulate of Savage's theory, says a consequence common to both options cannot affect the choice between them (Savage, 1954). The two pairs are therefore the same choice, red versus black, dressed in different yellow. Yet the majority chooses the known bet in each pair, switching sides between them. The kinship with the Allais paradox is structural and deliberate: both are common-consequence violations of the same axiom family, the independence axiom in its objective and subjective forms (Ellsberg, 1961). What differs, as the next section shows, is how much deeper the Ellsberg version cuts.

Why No Beliefs Can Rationalize the Choices

The Allais paradox rules out expected utility; the Ellsberg paradox rules out something more basic. Suppose a chooser in the three-color problem holds any subjective probabilities at all: the red chance is fixed at one third by the known count, and black and yellow share the remaining two thirds in some believed split. Betting on red over black then means the believed chance of black is below one third. Betting on black-or-yellow over red-or-yellow means the believed chance of red-or-yellow, one third plus the yellow belief, falls short of the known two thirds, so the yellow belief is below one third, and black, holding the rest of the two thirds, must be above one third. The two majority choices thus require the belief in black to be simultaneously below and above one third, a contradiction no probability assignment escapes. The two-urn version fails the same way: preferring the known urn for red puts the unknown red below one half, preferring it for black puts the unknown black below one half, and the two beliefs sum to less than one, which additivity forbids (Ellsberg, 1961). Notice what this argument never mentions: utility. The contradiction goes through for every utility function, every attitude toward money, every degree of risk aversion, because it concerns the beliefs, not the values. This is the precise sense in which Ellsberg cuts deeper than Allais: the Allais pattern shows that stated probabilities are not used linearly and can be repaired by transforming them, while the Ellsberg pattern shows that for ambiguous events there are no coherent probabilities available to transform (Machina & Siniscalchi, 2014). The demonstration below lets you take both versions of the choices before seeing your own pattern analyzed.

Risk, Ambiguity, and Knightian Uncertainty

The paradox forced a distinction back into economics that its formal machinery had been built to erase. Frank Knight separated risk, uncertainty that can be measured with known probabilities, from true uncertainty, where the probabilities themselves are unknown, and argued that the two play different economic roles (Knight, 1921). The Bayesian program seemed to dissolve the distinction: Savage showed that an agent whose preferences obey his postulates behaves as if assigning definite personal probabilities to every event, so that all uncertainty is, for a coherent chooser, risk (Savage, 1954). Ellsberg's problems are a counterexample to the dissolution. He used the word ambiguity for the middle territory his urns create: situations where information about probabilities is scanty, unreliable, or conflicting, so that one knows too little to name the odds and knows that one knows too little (Ellsberg, 1961). The behavioral signature is ambiguity aversion, a preference for known-probability prospects over unknown-probability prospects even when nothing suggests the unknown option is objectively worse (Camerer & Weber, 1992). Two features make the phenomenon deep rather than decorative. First, it is a property of the information about probabilities, not of the outcomes or their likelihoods, which is why it survives every relabeling of colors and payoffs. Second, it drives a wedge between two things the standard theory identifies: betting odds and degrees of belief. An ambiguity-averse chooser may report believing the unknown urn is symmetric, yet refuse to bet on it at even odds, holding part of their belief in reserve. What a person will bet no longer reveals, by itself, what they think is true, and that gap is the paradox's most consequential legacy for the study of judgment.

The Psychology of Ambiguity Aversion

If ambiguity aversion were a fixed distaste for vague probabilities, it would be a curiosity; what makes it cognitively interesting is how systematically it moves. Chip Heath and Amos Tversky showed that the same person who avoids betting on an unknown urn will happily bet on their own judgment, at identical stated odds, in a domain where they feel knowledgeable: sports fans prefer betting on games to betting on matched chance lotteries, and the preference reverses in domains of felt ignorance. Their competence hypothesis holds that willingness to act under uncertainty tracks not the vagueness of probabilities but how skilled or informed people feel, an attribution of credit and blame: winning by knowledge is satisfying and losing by ignorance is doubly aversive (Heath & Tversky, 1991). Craig Fox and Tversky then located the effect in comparison itself. In their comparative ignorance experiments, the gap between prices for clear and vague bets was large when both were evaluated side by side, and shrank sharply, sometimes toward nothing, when separate groups priced each bet in isolation: ambiguity aversion flourishes when the missing information is made salient by contrast, with later work confirming the comparative context as a powerful moderator of the effect (Fox & Tversky, 1995; Machina & Siniscalchi, 2014). The attitude is also not uniformly averse. Across likelihoods and domains, people are reliably ambiguity averse for moderate-likelihood gains, the situation the classic urns create, but tend toward ambiguity seeking for losses and for unlikely gains, a fourfold-style pattern that parallels risk attitudes and warns against treating aversion as a universal constant (Kocher, Lahno, & Trautmann, 2018). Behavioral mechanics connect ambiguity to how people handle layered uncertainty: Yoram Halevy replicated the paradox with real incentives and reservation prices and found ambiguity attitudes closely tied to whether people can reduce compound lotteries, two-stage risks, to their simple equivalents, suggesting that ambiguity is processed like a lottery over lotteries that most minds fail to collapse (Halevy, 2007). And the distinction reaches the brain: neuroimaging shows that ambiguous and risky gambles engage distinguishable neural systems, with circuitry associated with vigilance and evaluation responding to the degree of missing information, evidence that risk and ambiguity are separable kinds of uncertainty for the nervous system and not merely for the theorist (Hsu, Bhatt, Adolphs, Tranel, & Camerer, 2005). The demonstration below turns the core quantity, the ambiguity premium, into something you can measure on yourself.

Theories That Take Ambiguity Seriously

Accommodating the paradox formally means abandoning one of the pillars of Bayesian decision theory, and the major models differ in which pillar they choose. David Schmeidler kept a single belief but let it be non-additive: in Choquet expected utility, beliefs are capacities, weights on events that need not sum across a partition, so the unknown urn's red and black can each carry weight three tenths while their union carries weight one, the missing amount expressing reserved confidence; expectations are taken with the Choquet integral, which handles such weights coherently (Schmeidler, 1989). Schmeidler's motivation, it is worth stressing, was normative: he held that it is not obviously more rational to force a single sharp probability onto thin information than to let beliefs reflect the thinness. With Itzhak Gilboa he then axiomatized a second route: maxmin expected utility, in which the chooser entertains a whole set of probability distributions, all those consistent with what is known, and evaluates each act by its minimum expected utility over the set, choosing as a cautious planner would against an unhelpful nature (Gilboa & Schmeidler, 1989). Both Ellsberg patterns fall out immediately: with the unknown urn's red chance anywhere between zero and one, each unknown bet is valued at its worst case and the known fifty-fifty urn wins for either color. A third family, the smooth ambiguity model of Peter Klibanoff, Massimo Marinacci, and Sujoy Mukerji, keeps probabilities everywhere but adds a second layer: the chooser holds beliefs about the possible probability laws and applies a concave transformation to expected utilities across them, so ambiguity aversion becomes curvature over probability uncertainty, exactly as risk aversion is curvature over outcomes, with attitude and information cleanly separated (Klibanoff, Marinacci, & Mukerji, 2005). Table 1 aligns the families, together with the psychological account, and the demonstration that follows lets you watch three of them price the same ambiguous bet as the missing information grows.

Note. The first three are formal decision theories (Schmeidler, 1989; Gilboa & Schmeidler, 1989; Klibanoff, Marinacci, & Mukerji, 2005); the fourth is a psychological account of when the aversion appears (Fox & Tversky, 1995).

Is Ambiguity Aversion Robust?

The paradox has faced seven decades of stress testing, and the honest summary is that the phenomenon is real, replicable, and bounded. On reality: incentivized experiments with actual draws and actual money reproduce the classic pattern, and reservation-price methods show people paying measurably less for ambiguous bets than for objectively equivalent risky ones (Halevy, 2007). The broad experimental literature, surveyed early and influentially by Colin Camerer and Martin Weber, finds ambiguity aversion across procedures, subject pools, and stakes, typically shaving a meaningful fraction off the value of a vague bet relative to a clear one (Camerer & Weber, 1992). On the boundaries: the effect is strongest exactly where Ellsberg placed it, moderate-likelihood gains under side-by-side comparison, and it bends elsewhere. It reverses toward ambiguity seeking for losses and for long-shot gains (Kocher, Lahno, & Trautmann, 2018); it weakens when clear and vague options are priced in isolation rather than jointly (Fox & Tversky, 1995); and it flips into ambiguity preference in domains where people feel expert, which is why the same person can shun the unknown urn and relish a bet on their home team (Heath & Tversky, 1991). These boundary conditions are not embarrassments; they are data. A fixed taste would not care about comparison or competence, whereas an attitude rooted in what people notice about their own information should behave exactly this way. The result is a phenomenon more contingent than the textbook two-line summary suggests, and more psychologically revealing for it (Machina & Siniscalchi, 2014).

Criticisms and Open Questions

The deepest question the paradox raises is the same one Allais raised for risk: when preferences and axioms collide, which should yield? Here, unusually, the theory's builders split toward the choosers. Ellsberg himself argued that deliberate, reflective people confronted with his problems often decline to conform to the sure-thing principle even after the conflict is explained, and that their caution about acting on flimsy probability information is not obviously an error (Ellsberg, 1961). Schmeidler and Gilboa built that position into mathematics: maxmin expected utility is an axiomatic theory, derived from postulates a careful person could endorse, in which ambiguity aversion is coherent rather than confused, and Schmeidler's stated view was that forcing a sharp Bayesian prior onto thin information is itself the questionable move (Gilboa & Schmeidler, 1989). The classic rejoinder was pressed by Howard Raiffa in a comment published in the same journal issue as Ellsberg's paper: a chooser can hedge ambiguity away by randomizing, flipping a fair coin to decide which color to bet on, which converts the ambiguous bet into an objective fifty-fifty gamble and makes any premium for the known urn look incoherent; whether real and rational choosers can or should treat randomized bets this way remains actively contested (Machina & Siniscalchi, 2014). A second front is internal to the repair shop: the ambiguity models disagree with one another, face their own paradoxes and calibration problems, and complicate dynamic choice, since cautious preferences can make a chooser's plans and later selves fall out of agreement, so no successor theory yet commands the consensus Savage once did (Machina & Siniscalchi, 2014). A third is diagnostic: because competence and comparison move the effect, it is genuinely hard to say where belief ends and taste begins, whether the ambiguity-averse chooser holds pessimistic beliefs, incomplete beliefs, or ordinary beliefs plus a dislike of feeling ignorant, and the answer matters for whether policy should correct the pattern or respect it (Heath & Tversky, 1991; Fox & Tversky, 1995). Seventy years on, the paradox remains what it was on arrival: a small, sharp fact that large theories must organize themselves around.

Worked Example

The whole paradox, and its leading resolution, fits in a few lines of arithmetic. Take the two-urn problem with a prize of 100 and, for simplicity, value amounts at face value. The known urn's bets are each worth 50: half a chance at 100. Now suppose the chooser assigns the unknown urn's red some definite probability p. The red bet is worth 100p and the black bet 100(1 − p); whatever p is, the two values sum to 100, so at least one unknown bet is worth at least 50, and a believer in any p should take the unknown urn for at least one color. Strictly preferring the known urn for both colors requires 100p < 50 and 100(1 − p) < 50, that is, p below one half and above one half at once. No p works. Now give the chooser a maxmin mind instead: not one belief but a set, say every red probability from 0.3 to 0.7, with each bet valued at its worst case over the set (Gilboa & Schmeidler, 1989). The unknown red bet is now worth its minimum, 100 times 0.3, which is 30, and by symmetry the unknown black bet is also worth 30, while the known bets remain worth 50 each. Preferring the known urn for both colors is no longer a contradiction; it is the model's calm prediction, with an ambiguity premium of 20 on either bet. The arithmetic also exposes the signature of non-additive belief: the two unknown bets together are worth 60, not 100, and the missing 40 is confidence held in reserve, the numerical shadow of knowing that one does not know (Schmeidler, 1989). Shrink the belief set toward the single point one half and the premium melts to zero; widen it to total ignorance and the unknown bets are worth nothing at all. Ambiguity aversion, in this light, is not a mistake about the urn but a price on the width of one's own uncertainty.

Why It Matters

The Ellsberg paradox matters first for what it did to theory: it re-divided uncertainty. After Savage, rational choice recognized only risk; after Ellsberg, decision theory had to build a formal home for situations where probabilities are not merely unknown but unknowable on the available information, reviving Knight's distinction with mathematical teeth and producing the modern economics of ambiguity, whose models now shape work on insurance, portfolio choice, and robust policy under deep uncertainty such as climate change (Knight, 1921; Machina & Siniscalchi, 2014). It matters just as much for cognitive psychology, because it isolates a pure signal of metacognition in choice: the urns hold constant everything about outcomes and vary only what the chooser knows about their own information, and behavior moves, tracking comparison, competence, and the ability to collapse layered uncertainty (Fox & Tversky, 1995; Heath & Tversky, 1991; Halevy, 2007). The brain honors the same distinction, processing ambiguous and risky prospects through distinguishable systems (Hsu, Bhatt, Adolphs, Tranel, & Camerer, 2005). And it matters as a lesson in intellectual humility running in both directions: the majority pattern is a genuine violation of compelling axioms, yet the field's eventual verdict has been closer to Ellsberg than to his critics, treating caution about missing information as something to model rather than to cure. There is a fitting coda in the author's own life: the analyst who showed how deeply choices depend on the quality of one's information spent his later decades, after releasing the Pentagon Papers, arguing that the public's information about its gravest risks had been deliberately kept ambiguous. The paradox and the whistleblowing were, in that sense, one career.

Key Researchers

Daniel Ellsberg (1931–2023). American economist and RAND strategic analyst who posed the paradox in the doctoral research behind his 1961 paper, arguing that ambiguity is a real and rational concern; a decade later he became world-famous for releasing the Pentagon Papers.

Leonard J. Savage (1917–1971). American mathematician and statistician whose subjective expected utility theory and sure-thing principle are the paradox's explicit target; Ellsberg's title names his axioms.

David Schmeidler (1939–2022). Israeli mathematician and economic theorist who created Choquet expected utility with non-additive beliefs and, with Gilboa, maxmin expected utility, holding that the Ellsberg pattern reflects reasonable caution rather than error.

Itzhak Gilboa. Professor of economics and holder of the AXA Chair in Decision Sciences at HEC Paris; Schmeidler's doctoral student and co-author, he axiomatized maxmin expected utility and remains a leading theorist of rationality under uncertainty.

Amos Tversky (1937–1996). Israeli cognitive psychologist whose work with Heath and with Fox uncovered the competence and comparative-ignorance effects, relocating ambiguity aversion from the urn to the chooser's sense of their own knowledge.

Colin F. Camerer. Robert Kirby Professor of Behavioral Economics at the California Institute of Technology; co-author of the field's defining survey of ambiguity research and of the neuroimaging study showing that the brain distinguishes ambiguity from risk.

Key Terms

Frequently Asked Questions

What is the Ellsberg paradox?
The Ellsberg paradox is a set of choice problems in which most people prefer betting on known probabilities over unknown ones in a pattern that no assignment of subjective probabilities can justify. It shows that people treat ambiguity, the absence of probability information, as a feature of a gamble in its own right, violating the axioms of subjective expected utility (Ellsberg, 1961).

What is the difference between risk and ambiguity?
Risk is uncertainty with known probabilities, like a draw from an urn whose contents you can count, while ambiguity is uncertainty about the probabilities themselves, like a draw from an urn whose mix is unknown. The distinction goes back to Frank Knight, and the Ellsberg paradox showed that it matters for actual choice (Knight, 1921).

Why do the Ellsberg choices violate expected utility theory?
Because no coherent beliefs can produce them. Preferring the known urn for both colors implies that each color in the unknown urn has probability below one half, which is impossible since the two must sum to one, and the three-color version violates the sure-thing principle by letting a consequence common to both bets reverse the choice (Ellsberg, 1961).

How is the Ellsberg paradox different from the Allais paradox?
The Allais paradox concerns risk: probabilities are stated, and people fail to use them linearly. The Ellsberg paradox concerns ambiguity: probabilities are missing, and no beliefs of any kind can rationalize the majority pattern, which makes it a challenge to the existence of subjective probability rather than to how probabilities are weighted (Allais, 1953).

What is ambiguity aversion?
Ambiguity aversion is the preference for options with known probabilities over options with unknown probabilities, even when there is no reason to think the unknown option is objectively worse. It is the behavioral signature of the Ellsberg problems and one of the most studied phenomena in decision research (Camerer & Weber, 1992).

Is ambiguity aversion irrational?
That remains contested. Ellsberg argued that caution about thin probability information is reasonable, and maxmin expected utility gives such caution a rigorous axiomatic foundation, while critics reply that ambiguity can in principle be hedged away by randomizing which side of a bet to take (Gilboa & Schmeidler, 1989).

What theories explain the Ellsberg paradox?
Three formal families lead: Choquet expected utility, which lets beliefs be non-additive; maxmin expected utility, which evaluates acts by their worst case over a set of possible probability distributions; and the smooth ambiguity model, which adds beliefs about probabilities and a cautious attitude toward that second layer (Schmeidler, 1989).

Is ambiguity aversion universal?
No. It is reliable for moderate-likelihood gains, the situation the classic urns create, but people are often ambiguity seeking for losses and for unlikely gains, and the effect weakens when ambiguous options are evaluated in isolation or in domains where people feel competent (Kocher, Lahno, & Trautmann, 2018).

References

The three interactive tools on this page — the Ellsberg-choice, ambiguity-premium, and model-comparison demonstrations — generate their figures and compute their results live in your browser; no dataset is bundled with the page. The empirical claims in the text are sourced to the references above.