The Allais Paradox

Three ways to read the result

Allais's reading: the theory is wrong. He held that expected utility was a poor description of rational behaviour, not merely of actual behaviour, and that the "American School" — von Neumann, Morgenstern, Savage, Friedman, Marschak — had mistaken a mathematically convenient axiom for a requirement of reason. He proposed instead that people care about the whole distribution of outcomes, including its dispersion, in ways expected utility cannot express. He held this position for decades and it was largely ignored for decades.

Savage's reading: the chooser is wrong. Savage accepted that people answer as Allais said, and treated it as an error of the same kind as a miscalculation. His own account in The Foundations of Statistics (1954, §5.6) is disarming: he gave the paradoxical answers, was shown the drum argument, recognised it, and changed his mind. The sure-thing principle, on this view, is a normative claim, and discovering that people violate it is like discovering that people add up restaurant bills wrongly.

This reading has real force. Anyone who violates independence in a repeated setting can be constructed into a sequence of trades they will accept and which leave them strictly worse off. If you would pay to avoid that, independence has a claim on you.

The descriptive reading: build a better model. The position that won. Whether or not independence is normatively compelling, it is not what people do, and a science of choice needs a model of what people do. This is the road that leads from Allais through prospect theory to behavioural economics.

What the descriptive models keep and drop

Every serious alternative keeps the structure of weighing outcomes against likelihoods and weakens independence somewhere.

• Prospect theory (Kahneman and Tversky, 1979) replaces probabilities with decision weights and utility with a value function defined over gains and losses from a reference point, concave for gains, convex for losses, and steeper for losses than gains.
• Rank-dependent expected utility (Quiggin, 1982) applies the weighting function to cumulative probabilities rather than individual ones, which repairs a defect in the 1979 model — decision weights applied outcome by outcome can lead to choosing a stochastically dominated option.
• Cumulative prospect theory (Tversky and Kahneman, 1992) merges the two, and is the standard model today.
• Regret theory (Loomes and Sugden, 1982; Bell, 1982) explains the pattern differently: what people minimise is anticipated regret, and drawing ticket 1 after refusing a guaranteed million is the most regrettable outcome available.
• Betweenness and disappointment models (Chew, Dekel, Gul) weaken independence in more targeted ways and preserve more of the classical structure.

The models disagree about mechanism and make different predictions elsewhere. What they share is the abandonment of independence, which is the lasting content of Allais's attack.

The objections that were made, and how they fared

"It disappears with real money." Partly true, and worth stating fairly. The original stakes were hypothetical and enormous. Incentivised experiments with real payments generally find the effect at reduced magnitude, and some designs find it weakly or not at all. But the common-consequence effect has been replicated with real money at moderate stakes many times, and reviews of the incentives literature find it attenuated rather than eliminated.

"It disappears with learning." Also partly true. Repeated exposure with feedback reduces violation rates. This supports Savage's reading, and it is a reason to be careful about generalising one-shot laboratory choices to institutional decision-making. It does not help with the many real decisions — insurance, medical treatment, career moves — that are taken once.

"It is a framing artefact." The drum presentation reduces violations somewhat; so does presenting the two problems side by side. This is a real finding and it cuts both ways. That a transparent presentation improves consistency supports the normative view. That the effect survives transparent presentation for a substantial minority undercuts the claim that it is only confusion.

"It is measurement error." The strongest recent version of this objection. Some researchers argue that a portion of apparent axiom violations reflects random noise in responses rather than systematic preference, and that once noise is modelled, the estimated violation rate falls. The debate is technical and unfinished, but the consensus is that noise explains part of the effect and not the direction of it — noise does not predict which way people deviate, and the Allais pattern deviates in a specific direction that replicates.

An honest note on the status

This entry is marked explained rather than resolved deliberately. What is settled is the empirical fact — the choice pattern is robust and widespread — and the formal diagnosis: it violates independence, and models with probability weighting reproduce it. What is not settled is whether people who choose A and D are making a mistake. That is a normative question about rationality, and economists and philosophers still disagree about it. Marking it resolved would be overclaiming.

1738. Daniel Bernoulli resolved the St Petersburg paradox by proposing that people maximise expected utility rather than expected money. This is the ancestor of the theory Allais attacked.

1944 and 1947. John von Neumann and Oskar Morgenstern published Theory of Games and Economic Behavior. The second edition (1947) contains the axiomatic derivation of expected utility from a small set of conditions on preferences. The result made expected utility the default model of rational choice.

May 1952. The Centre National de la Recherche Scientifique held a colloquium in Paris on the foundations and applications of the theory of risk. Savage, Marschak, Friedman, Samuelson and de Finetti were among those present, and Allais circulated his questionnaire to the participants. The best-known part of the story is that he put the questions to Savage over lunch and Savage gave the inconsistent answers.

1953. Allais published "Le comportement de l'homme rationnel devant le risque: critique des postulats et axiomes de l'école américaine" in Econometrica, volume 21, pages 503–546. The paper is polemical, long, and written in French in a journal whose readership was increasingly Anglophone, which is part of why it took so long to land.

1954. Savage published The Foundations of Statistics, giving his own axiomatisation of subjective expected utility and, in §5.6, his account of taking Allais's test and failing it.

1970s. The result stayed marginal for two decades. It was cited as a curiosity and generally answered with some version of Savage's reply. Allais continued to argue his case and was largely outside the mainstream.

1979. Kahneman and Tversky published "Prospect Theory: An Analysis of Decision under Risk" in Econometrica, volume 47, pages 263–291. They replicated the Allais pattern with a systematic battery of problems, named the certainty effect, and supplied a model. This is the point at which the finding became central rather than marginal. The paper is among the most cited in the history of economics.

1982. Quiggin introduced rank-dependent utility, fixing the dominance problem.

1988. Allais received the Nobel Memorial Prize in Economic Sciences. The citation was for his work on market equilibrium and the efficient use of resources, but the recognition of the risk work was explicit and, by then, overdue. Much of his early work had been published only in French, and Paul Samuelson had written that had Allais's earliest writings been in English, a generation of economic theory would have taken a different course.

1992. Tversky and Kahneman published cumulative prospect theory.

2002. Kahneman shared the Nobel Memorial Prize with Vernon Smith. Tversky had died in 1996 and the prize is not awarded posthumously; Kahneman said repeatedly that the work was joint.

It broke the monopoly of expected utility. For thirty years after von Neumann and Morgenstern, expected utility was both the model of how people choose and the standard of how they should. Allais showed the two could not be the same thing. Everything in behavioural economics follows from taking that gap seriously.

It changed what counts as evidence in economic theory. Allais's method — write down the axioms, construct a choice pair that separates them, ask people — is now standard. Before 1953 the dominant methodological position, associated with Milton Friedman's 1953 essay, was that the realism of assumptions did not matter provided predictions were good. Allais's design attacks that directly: the assumption generates a testable prediction about choices, and the prediction fails.

It located the failure precisely. This is what makes the result useful rather than merely destructive. Allais did not show that people are irrational or that utility theory is generally wrong. He showed that one specific axiom fails, and he showed it with a pair of gambles that isolates that axiom and nothing else. That precision is why the repair was possible.

It gave prospect theory its foundation. The probability weighting function — the single most consequential idea to come out of this — exists because distributions like the Allais pair demanded it. Overweighting of small probabilities explains why the same person buys both insurance and lottery tickets, a combination expected utility struggles with, and it is the mechanism behind a large share of documented anomalies in insurance, gambling and risk communication.

Insurance markets. People over-insure against small, salient, near-certain-to- be-avoided losses — extended warranties, low-deductible policies, flight insurance — and under-insure against large, low-probability ones such as flood and earthquake. The certainty effect explains the first pattern directly: the appeal of eliminating a risk entirely is out of proportion to the size of the risk eliminated. Insurers price against this, and consumer protection regulators cite it.

Medical decisions. Patients choosing between a treatment with a high certain survival rate and one with a higher expected but uncertain outcome show the Allais pattern reliably. The framing of survival statistics — "a 90% chance of surviving five years" versus "a 10% chance of dying within five years" — interacts with it, and the informed consent literature treats both effects as things clinicians must actively manage rather than hope to avoid.

Vaccine and safety communication. Messaging that promises a risk will be "eliminated" is far more persuasive than messaging that promises it will be reduced by a large factor, even when the second is objectively better. The certainty effect is the reason, and public health communication research has been explicit about it since the 1980s.

Regulatory cost-benefit analysis. Formal analysis uses expected values, which means it systematically undervalues the elimination of risks relative to how the public values it. This is a persistent source of friction between analysts and the public in environmental and safety regulation, and it is not simply public innumeracy — it is a disagreement about whether the independence axiom should govern collective decisions.

Risk models in finance and AI. Any system that optimises expected utility — portfolio optimisation, reinforcement learning agents with scalar rewards, automated decision systems — is committed to independence and will make choices that human stakeholders find unacceptable in exactly the Allais configuration. Work on risk-sensitive reinforcement learning, distributional value functions and rank-dependent objectives is the response, and it reaches for the same mathematical tools Quiggin and Tversky developed.

The independence axiom

Let $\Delta(X)$ be the set of simple lotteries over a set of outcomes $X$, and $\succsim$ a preference relation on $\Delta(X)$. For lotteries $p, q, r$ and $\alpha \in (0,1]$, write $\alpha p + (1-\alpha) r$ for the compound lottery.

Independence. For all $p, q, r \in \Delta(X)$ and all $\alpha \in (0,1]$: $$p \succsim q \iff \alpha p + (1-\alpha) r \succsim \alpha q + (1-\alpha) r$$

Von Neumann–Morgenstern theorem. If $\succsim$ is complete, transitive, continuous and satisfies independence, there exists $u : X \to \mathbb{R}$, unique up to positive affine transformation, such that

$$p \succsim q \iff \sum_{x} p(x),u(x) ;\ge; \sum_{x} q(x),u(x)$$

Independence is the axiom with content. Completeness, transitivity and continuity are structural; independence is what makes the representation linear in probabilities.

The Allais pair as a violation

Write $M_0 = 0$, $M_1 = 100$, $M_5 = 500$ (millions). Define the lotteries:

$$ \begin{aligned} L_{1A} &= \delta_{M_1} \ L_{1B} &= 0.89,\delta_{M_1} + 0.10,\delta_{M_5} + 0.01,\delta_{M_0} \ L_{2A} &= 0.11,\delta_{M_1} + 0.89,\delta_{M_0} \ L_{2B} &= 0.10,\delta_{M_5} + 0.90,\delta_{M_0} \end{aligned} $$

Let $\alpha = 0.11$, $p = \delta_{M_1}$, and $q = \tfrac{10}{11}\delta_{M_5} + \tfrac{1}{11}\delta_{M_0}$. Then:

$$L_{1A} = \alpha p + (1-\alpha),\delta_{M_1}, \qquad L_{1B} = \alpha q + (1-\alpha),\delta_{M_1}$$ $$L_{2A} = \alpha p + (1-\alpha),\delta_{M_0}, \qquad L_{2B} = \alpha q + (1-\alpha),\delta_{M_0}$$

Check the second pair: $\alpha q = 0.11 \cdot \tfrac{10}{11}\delta_{M_5} + 0.11\cdot\tfrac{1}{11}\delta_{M_0} = 0.10,\delta_{M_5} + 0.01,\delta_{M_0}$, and adding $0.89,\delta_{M_0}$ gives $L_{2B}$ exactly.

Each pair is ${ \alpha p + (1-\alpha) r,; \alpha q + (1-\alpha) r }$ for the same $p$, $q$ and $\alpha$, differing only in the common consequence $r \in {\delta_{M_1}, \delta_{M_0}}$. Independence therefore requires $L_{1A} \succ L_{1B} \iff L_{2A} \succ L_{2B}$. The modal pattern $L_{1A} \succ L_{1B}$ together with $L_{2B} \succ L_{2A}$ contradicts it directly.

The name common consequence effect comes from this construction: the two decisions share a consequence occurring with probability $1 - \alpha$, and only its identity differs.

Geometry: the Marschak–Machina triangle

Fix three outcomes $x_1 < x_2 < x_3$. A lottery over them is determined by $(p_1, p_3)$ with $p_2 = 1 - p_1 - p_3$, so the lottery space is a triangle in the plane. Expected utility with $u(x_1) = 0$, $u(x_3) = 1$ gives indifference curves

$$p_1 \cdot 0 + (1-p_1-p_3),u(x_2) + p_3 = \text{constant}$$

which are parallel straight lines, with slope determined by $u(x_2)$ alone. Parallelism is the geometric content of independence, and it is what the Allais pair tests. The four lotteries $L_{1A}, L_{1B}, L_{2A}, L_{2B}$ form a parallelogram in this triangle: the segment from $L_{1A}$ to $L_{1B}$ is parallel to the segment from $L_{2A}$ to $L_{2B}$, and equal in length. Preferring $L_{1A}$ and $L_{2B}$ means the indifference curves through the two segments have different slopes — they fan out toward the lower right, the certainty corner.

Machina's "fanning out" hypothesis (1982) generalises this: keep smooth preferences over lotteries and local expected-utility behaviour, but drop global linearity. It accommodates the Allais pattern and several others while retaining enough structure to do comparative statics.

Probability weighting

Prospect theory replaces $\sum_i p_i u(x_i)$ with

$$V = \sum_i \pi(p_i), v(x_i)$$

where $v$ is a value function over gains and losses relative to a reference point and $\pi$ is a decision weight. In the 1979 formulation, $\pi$ satisfies:

• Overweighting of small probabilities: $\pi(p) > p$ for small $p$.
• Subcertainty: $\pi(p) + \pi(1-p) < 1$ for $0 < p < 1$.
• Subproportionality: for fixed $r$, $\pi(pq)/\pi(p)$ decreases in $p$.
• $\pi(0) = 0$, $\pi(1) = 1$, with discontinuities near the endpoints.

Subcertainty is what generates the certainty effect. If $\pi(0.89) + \pi(0.11) < 1$ then $1 - \pi(0.89) > \pi(0.11)$, and the Allais pair becomes consistent: preferring $L_{1A}$ requires $(1 - \pi(0.89)),v(M_1) > \pi(0.10),v(M_5)$ while preferring $L_{2B}$ requires $\pi(0.10),v(M_5) > \pi(0.11),v(M_1)$, and both hold whenever

$$\pi(0.11) ;<; \frac{\pi(0.10),v(M_5)}{v(M_1)} ;<; 1 - \pi(0.89)$$

Under expected utility that interval is empty, because $\pi(0.11) = 0.11 = 1 - 0.89$. Subcertainty opens it.

Cumulative prospect theory (1992) applies the weighting to cumulative probabilities, which guarantees respect for stochastic dominance. Tversky and Kahneman's estimated one-parameter form is

$$w(p) = \frac{p^{\gamma}}{\left(p^{\gamma} + (1-p)^{\gamma}\right)^{1/\gamma}}$$

with $\gamma \approx 0.61$ for gains and $\gamma \approx 0.69$ for losses. The function is inverse-S shaped: it crosses the diagonal near $p \approx 0.35$, overweighting probabilities below that and underweighting those above. The companion value function is

$$v(x) = \begin{cases} x^{\alpha} & x \ge 0 \ -\lambda(-x)^{\beta} & x < 0 \end{cases}$$

with estimates $\alpha = \beta \approx 0.88$ and loss aversion $\lambda \approx 2.25$.

The normative counter-argument

The strongest case for independence is dynamic. Consider a decision-maker who strictly prefers $L_{1A}$ to $L_{1B}$ and $L_{2B}$ to $L_{2A}$. Present the choice as a two-stage lottery: first a draw determines whether the common consequence occurs; only if it does not do you get the sub-lottery $p$ or $q$.

If the agent evaluates the tree by backward induction and satisfies consequen- tialism (the choice at a node depends only on what remains) and dynamic consistency (a plan made ex ante is carried out), then independence is forced. Violating independence therefore requires giving up one of those two.

That is a real cost, and it is why the normative question stays open. An agent who abandons dynamic consistency can be led through a sequence of individually acceptable choices to a strictly dominated position. An agent who abandons consequentialism must let the branches that did not occur influence the choice at the node reached — which is exactly what regret theory says people do, and which is not obviously irrational if regret is a real cost.

The modern literature on resolute choice and sophisticated choice explores both routes. Neither has produced a consensus, and the honest statement is that Allais's questionnaire opened a normative question that seventy years of work has sharpened without closing.