Allais Paradox

The Allais paradox is a pair of choices, first posed to a room of economists in 1952, that almost everyone answers in a way their own principles forbid. Offered a certain fortune or a gamble at a larger one, most people take the sure thing; offered two long shots that differ in exactly the same way, most people switch. Written out side by side, the two answers cannot both come from anyone who values gambles by their expected utility, the reigning theory of rational choice. The paradox is not a trick or a miscalculation: intelligent people shown precisely how their choices contradict the theory often refuse to change them. This page explains the two choices and why they clash with the independence axiom, distinguishes the paradox's two classic forms, shows how overweighting certainty resolves it, and traces the debate it launched, one that is still unsettled, over whether the theory or the chooser should give way. Three interactive tools let you take the choices yourself, watch a probability-weighting curve turn the contradiction on and off, and see the paradox as a geometric fanning of preference.

The Allais paradox is a demonstration that people's choices among risky prospects systematically violate expected utility theory, the standard normative model of decision making under risk (Allais, 1953). It was constructed by the French economist Maurice Allais and first presented at a 1952 colloquium in Paris, as a deliberate challenge to the axioms that Leonard Savage and others had recently established as the foundation of rational choice (Savage, 1954). The paradox takes the form of two paired decisions; the pattern of choice that most people find natural in the two together is logically inconsistent with maximizing expected utility, because it violates the theory's independence axiom. Daniel Kahneman and Amos Tversky later reproduced the effect with modest stakes and identified its psychological source as the certainty effect, the disproportionate weight people attach to outcomes that are certain, making it a cornerstone of the evidence behind prospect theory (Kahneman & Tversky, 1979). What has kept the paradox central for seventy years is its dual edge: it is at once a robust descriptive failure of the standard theory and a live argument, pressed by Allais himself, that the standard of rationality, not the human chooser, is what should be revised. The sections below present the choices, the inconsistency, the two classic forms of the paradox, the cognitive explanation, and the long debate over what it means.

The Choice That Breaks the Rule

The paradox is best met firsthand, in its original large-stakes form. Consider two decisions. In the first, choose between Option A, one million dollars guaranteed, and Option B, a gamble paying one million with 89 percent probability, five million with 10 percent probability, and nothing with the remaining 1 percent. Most people choose A: the certainty of a fortune outweighs a small chance at more shadowed by a real chance of nothing. In the second decision, choose between Option C, an 11 percent chance of one million and otherwise nothing, and Option D, a 10 percent chance of five million and otherwise nothing. Here most people choose D: both are unlikely, so the marginally smaller chance is worth accepting for the far larger prize. Taken one at a time, each choice seems entirely reasonable, and the great majority of people, including trained economists, make exactly this pair, A and then D (Allais, 1953). The trouble, as the next section shows, is that no one who values gambles by their expected utility can consistently make both. Figure 1 lays the four options out to expose their hidden structure.

Why the Two Choices Cannot Both Be Right

The inconsistency follows from the independence axiom, one of the conditions von Neumann and Morgenstern showed to be equivalent to expected utility maximization. Independence requires that if one prospect is preferred to another, then mixing each with the same third prospect in the same proportion leaves the preference unchanged: the shared component is a sure thing either way and should not affect the choice (Savage, 1954). Figure 1 shows that the Allais problems are built to isolate exactly such a shared component. Both problems present the same underlying choice on the differing 11 percent of probability, an 11-in-100 shot at one million versus a 10-in-100 shot at five million alongside a 1-in-100 chance of nothing; the two problems differ only in what fills the common 89 percent, one million in Problem 1 and nothing in Problem 2. By independence, that common block cannot change which side is preferred, so a rational agent must choose A and C, or B and D, but never A and D. The same point can be shown in algebra. Assigning utilities with the utility of nothing set to zero, preferring A to B requires that the utility of one million exceed 0.89 times that utility plus 0.10 times the utility of five million, which simplifies to 0.11 times the utility of one million exceeding 0.10 times the utility of five million. Preferring D to C requires the reverse, that 0.10 times the utility of five million exceed 0.11 times the utility of one million. The two demands are contradictory: no utility function can satisfy both (Machina, 1987). The modal choice is therefore not merely unusual but formally impossible under the theory. Take the choices yourself in the demonstration below, in either of the paradox's two forms.

Two Forms: Common Consequence and Common Ratio

Allais's construction actually yields two distinct violations, and careful treatments keep them apart. The version in Figure 1 is the common-consequence effect: the two problems share a common outcome, the 89 percent block, and the paradox is that changing that shared consequence from a good outcome to a bad one reverses people's willingness to gamble on the rest. The second form is the common-ratio effect, which Kahneman and Tversky used in their most famous demonstration. There, one problem offers a certain 3,000 against an 80 percent chance of 4,000, and most people take the certain amount; a second problem offers a 25 percent chance of 3,000 against a 20 percent chance of 4,000, and most people now take the larger prize. The probabilities in the second problem are exactly the first problem's probabilities scaled down by the same ratio, one quarter, so independence again demands the same choice in both, yet preferences reverse (Kahneman & Tversky, 1979). The two effects have slightly different anatomies, one turning on a shared outcome and the other on proportional scaling of probabilities, but they share a diagnosis. In each, the reversal appears precisely when one option changes from certain to merely very probable, or the reverse, which points to a single culprit: the special psychological pull of certainty.

The Certainty Effect and Probability Weighting

The cognitive explanation of the paradox is that people do not treat probabilities linearly, as the theory requires, but weight them, and the weighting is steepest near the endpoints. Reducing a probability from 100 percent to 99 percent feels like a large loss, the surrender of a guarantee, whereas reducing it from 50 percent to 49 percent barely registers, even though the two changes are numerically identical. Kahneman and Tversky named this the certainty effect and made it the engine of the Allais reversal: in Problem 1 the certain million is overweighted, so giving up certainty for Option B feels costly, while in Problem 2, where both options are already uncertain, no such premium applies and the larger prize wins (Kahneman & Tversky, 1979). Prospect theory formalizes this by replacing raw probabilities with decision weights drawn from a nonlinear weighting function that overweights small probabilities, underweights moderate to large ones, and changes fastest as certainty is approached; the cumulative version of the theory refines the function so that it never leads a chooser to accept a dominated gamble, and supplies parameter estimates from experimental data (Tversky & Kahneman, 1992). Under such a weighting function the Allais choices are no longer contradictory: they are the correct output of a system that prices certainty at a premium. The demonstration below lets you adjust a probability-weighting curve and watch the paradox switch on as the curve bends away from the straight line of expected utility, and switch off as it straightens. The relationship of the paradox to the descriptive theory it helped inspire is developed on the Expected Utility Theory page.

Fanning Out: The Geometry of the Violation

The paradox has an illuminating geometric form. Any gamble over three fixed outcomes, a worst, a middle, and a best, can be plotted as a point in a triangle whose horizontal axis is the probability of the worst outcome and whose vertical axis is the probability of the best, a device known as the Marschak-Machina triangle. In this triangle, expected utility theory's indifference curves, the sets of gambles among which an agent is indifferent, are parallel straight lines: independence forces them to be both straight and parallel. The four Allais options plot as the corners of a parallelogram, so the segment from A to B and the segment from C to D point in the same direction and have the same length. If indifference curves were parallel lines, the two segments would cross them identically and the two choices would have to agree. The modal Allais pattern instead requires the indifference curves to be steeper toward the lower right of the triangle and shallower toward the upper left, so that they fan out rather than run parallel. Mark Machina showed that this fanning-out property, far from being mere disorder, can be captured by a coherent generalized theory in which preferences remain smooth and well behaved but abandon the strict linearity that independence imposes (Machina, 1982). The demonstration below lets you place the two Allais segments in the triangle and rotate a family of indifference lines, so you can confirm that no single parallel family ranks both choices the way people actually do.

The fanning-out account is one of several theories built to accommodate the paradox by relaxing independence while preserving as much of the theory as possible. Table 1 sets out the main families. They differ in mechanism but agree on the target: each replaces the independence axiom with something weaker that still yields orderly, non-arbitrary choice (Starmer, 2000).

Note. Rank-dependent utility was introduced by John Quiggin (Quiggin, 1982) and regret theory by Graham Loomes and Robert Sugden (Loomes & Sugden, 1982); prospect theory's weighting is the most widely used (Tversky & Kahneman, 1992).

Is the Paradox Robust?

A pattern this consequential invites the question of whether it is real, stable, and general, and the evidence is layered. On persistence, Paul Slovic and Amos Tversky ran the decisive early test: they presented people with the choices, then with a clear statement of the normative argument for the independence axiom, and asked them to choose again. Violations persisted as the majority pattern even after the argument was understood, which undercut the hope that the paradox is a mere lapse that reflection dissolves (Slovic & Tversky, 1974). On generality, Steffen Huck and Wieland Müller moved the paradox out of the seminar room and put it to a large, demographically representative sample rather than the usual convenience pool of students, and found the effect held across the population (Huck & Müller, 2012). But the picture is not one of uniform strength. A large synthesis by Pavlo Blavatskyy, Andreas Ortmann, and Valentyn Panchenko, drawing on 81 experiments across 29 studies, found the Allais paradox to be a fragile empirical finding whose appearance depends heavily on presentation: it emerges clearly with high, hypothetical payoffs and when the middle outcome is close to the best one, and it can even reverse when the probability of the extreme outcomes is arranged in particular ways (Blavatskyy, Ortmann, & Panchenko, 2022). And on the descriptive theories built to explain it, Michael Birnbaum has shown that the tools are not the last word: across a battery of new tests, prospect theory and its cumulative successor generate eleven fresh paradoxes of their own, false predictions and self-contradictions that a different family of configural-weight models anticipated (Birnbaum, 2008). The paradox, in short, is durable but conditional, and the search for the theory that best captures it is not finished.

Criticisms and Open Questions

The most important dispute the paradox raises is not whether people violate the axiom but whether they are wrong to. Allais never intended the paradox as a catalogue of human error; he presented it as evidence that a reasonable person may legitimately prefer certainty and attend to the whole distribution of outcomes, and that the independence axiom, not the chooser, is the questionable element (Allais, 1953). The persistence of violations under full reflection gives this position force, since it is hard to call irrational a preference that survives a clear explanation of the contrary argument (Slovic & Tversky, 1974). The episode that crystallizes the debate is autobiographical: when Allais posed his choices to Leonard Savage, the very architect of the sure-thing principle, at the 1952 Paris meeting, Savage made the modal, axiom-violating choices, then on reflection judged his own choices to be the error and revised them to conform, taking the paradox as a spur to correct his intuitions rather than his theory (Savage, 1954). Two thinkers of the first rank thus drew opposite lessons from the same choices, and both positions remain defensible. Defenders of the axiom add a pragmatic argument, that agents who abandon independence can, in principle, be led step by step into a series of choices that leaves them worse off with certainty. A second open question is empirical rather than normative: given that the effect's magnitude shifts so much with framing, real versus hypothetical stakes, and the spacing of outcomes, it is unclear how large and how universal the underlying tendency truly is, which complicates any attempt to fix a single descriptive model (Blavatskyy, Ortmann, & Panchenko, 2022). These are not signs of a defective phenomenon but of a deep one whose interpretation careful researchers still contest.

Worked Example

The contradiction can be made concrete with any utility numbers at all, which is what makes it so sharp. Suppose a person's preferences did satisfy expected utility theory, and set the utility of nothing to zero, the utility of one million to some value u1, and the utility of five million to some value u5, with u5 greater than u1 because more money is better. Now impose the two modal Allais choices. Preferring the certain million in Problem 1 to the gamble means u1 is greater than 0.89 times u1 plus 0.10 times u5; subtracting 0.89 times u1 from both sides gives 0.11 times u1 greater than 0.10 times u5. Preferring the five-million gamble in Problem 2 means 0.10 times u5 is greater than 0.11 times u1, since the certain-zero portions contribute nothing. Placing the two results together yields 0.11 times u1 greater than 0.10 times u5 greater than 0.11 times u1, which asserts that a number is strictly greater than itself. There is no escape through the particular values, because u1 and u5 never had to be specified; the impossibility holds for every utility function whatsoever (Machina, 1987). The lesson is that the Allais pattern does not reflect an unusual utility curve or an eccentric attitude toward money. It reflects a treatment of probability that expected utility theory does not permit at all, which is exactly why accommodating it required new theories rather than new utility functions (Kahneman & Tversky, 1979).

Why It Matters

The Allais paradox matters because it converted a philosophical worry into a reproducible fact and, in doing so, helped redirect the study of choice. For the founders of decision theory the axioms were self-evident canons of reason; the paradox showed that one of them is routinely and knowingly violated, forcing a separation between expected utility theory as a normative ideal and as a description of behavior that organizes the whole field of behavioral decision research (Machina, 1987). It supplied the central anomaly that prospect theory and its relatives were built to explain, and the certainty effect it isolated remains one of the best-documented regularities in the psychology of risk, with consequences reaching from insurance pricing to the design of public lotteries to the framing of medical and financial decisions (Kahneman & Tversky, 1979; Starmer, 2000). Just as importantly, it is a standing lesson in scientific humility about rationality itself: a preference that a great theorist endorsed on reflection, that survives explanation, and that appears across the general population cannot be dismissed as simple confusion, and deciding what it teaches about reason is a question economics and psychology still share (Allais, 1953; Slovic & Tversky, 1974). The demonstrations on this page let you find the pattern, and its resolution, in your own choices.

Key Researchers

Maurice Allais (1911–2010). French economist and 1988 Nobel laureate who devised the paradox and argued, against the emerging orthodoxy, that its violations of the independence axiom reflect a defensible concern for certainty rather than an error of reason.

Leonard J. Savage (1917–1971). American mathematician and statistician whose sure-thing principle the paradox targets; his own axiom-violating response to Allais's choices, and his decision to revise it, became the paradox's defining anecdote.

Amos Tversky (1937–1996). Israeli cognitive psychologist who, with Kahneman, reproduced the paradox at modest stakes, isolated the certainty effect, and showed that violations persist even after the normative argument is explained.

Daniel Kahneman (1934–2024). Israeli-American psychologist and 2002 Nobel laureate who, with Tversky, made the certainty effect the foundation of prospect theory, the descriptive model that accommodates the paradox through probability weighting.

Mark J. Machina. Professor emeritus of economics at the University of California, San Diego, who represented the paradox geometrically as a fanning out of indifference curves and built a generalized expected utility analysis that retains smooth preferences without strict independence.

John Quiggin. Professor of economics at the University of Queensland whose theory of anticipated, or rank-dependent, utility gave the weighting of probabilities a rigorous form that avoids the pitfalls of the original prospect theory.

Michael H. Birnbaum. Professor of psychology at California State University, Fullerton, whose experiments revealed new paradoxes that the prospect theories cannot handle and advanced configural-weight models as an alternative.

Key Terms

Frequently Asked Questions

What is the Allais paradox?
The Allais paradox is a pair of decisions whose typical answers are inconsistent with expected utility theory. Most people choose a certain prize in one problem but a long-shot gamble in a second problem that differs only by an outcome common to both options, which violates the theory's independence axiom (Allais, 1953).

Why do the two Allais choices contradict expected utility theory?
Because the two problems differ only in a consequence common to both options, and the independence axiom says a common consequence cannot change which option is preferred. Written in algebra, the popular pair of choices requires one quantity to be both greater than and less than another, which no utility function can satisfy (Machina, 1987).

What is the difference between the common-consequence and common-ratio effects?
They are the two classic forms of the paradox. In the common-consequence effect, two problems share an outcome and changing that shared outcome reverses choice; in the common-ratio effect, one problem's probabilities are a proportional scaling of another's, yet preferences still reverse. Both trace to the certainty effect (Kahneman & Tversky, 1979).

What is the certainty effect?
The certainty effect is the disproportionate weight people place on outcomes that are certain. A drop from certainty to high probability feels much larger than a numerically equal drop between two uncertain probabilities, which is what drives the Allais reversal (Kahneman & Tversky, 1979).

How does prospect theory resolve the Allais paradox?
Prospect theory replaces raw probabilities with decision weights from a nonlinear function that overweights small probabilities and changes fastest near certainty. Under this weighting the Allais choices are no longer contradictory but are the expected output of a system that prices certainty at a premium (Tversky & Kahneman, 1992).

Does the Allais paradox mean people are irrational?
Not necessarily. Allais argued the choices are reasonable and that the independence axiom is the questionable element, and violations persist even after people are shown the normative argument, so many theorists treat the pattern as a challenge to the axiom rather than proof of error (Slovic & Tversky, 1974).

Is the Allais paradox a robust finding?
It is durable but conditional. Violations persist under reflection and appear in representative samples, but a large synthesis found the effect's size depends strongly on presentation, emerging clearly with high hypothetical payoffs and even reversing under some arrangements of probability (Blavatskyy, Ortmann, & Panchenko, 2022).

Who discovered the Allais paradox?
The French economist Maurice Allais, who presented it at a 1952 colloquium in Paris and published it in 1953 as a deliberate challenge to the axioms of rational choice then being established by Leonard Savage and others (Allais, 1953).

References

The three interactive tools on this page — the Allais-choice, probability-weighting, and fanning-out 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.