The Ellsberg Paradox

Does this mean people are irrational? Ellsberg himself did not think so. He argued that behaving differently under risk (known probabilities) and under ambiguity (unknown probabilities) might be entirely defensible. If you do not know the probability distribution, there is no unique rational response — you are not given enough information to determine one. Treating the situation as genuinely different from one where you are given full information is not a cognitive error; it may be the epistemically honest response.

Maximin and multiple priors. One formal response to Ellsberg is to replace the single subjective probability of Savage's theory with a set of probabilities, and to evaluate acts by their expected utility under the worst-case member of the set. This is the maximin expected utility model of Gilboa and Schmeidler (1989). It predicts Ellsberg-style preferences: faced with an unknown probability, you behave as if you are using the prior most unfavourable to your bet, which makes known odds attractive relative to unknown ones. This is a coherent framework, though it abandons Savage's axioms.

The sure-thing principle. Leonard Savage's sure-thing principle says: if you prefer A to B when event E occurs, and prefer A to B when event E does not occur, then you should prefer A to B regardless of whether E occurs. Ellsberg's choices violate this. In the standard urn experiment, the sure-thing principle can be shown to require consistency across the two choices — and Ellsberg's choosers are not consistent. Abandoning the sure-thing principle is a serious step: it is among the most intuitive-seeming axioms in decision theory, and its failure leaves a gap in the foundations.

Daniel Ellsberg ran the thought experiments in the early 1960s while working at RAND on questions of military strategy under uncertainty. He circulated the examples widely before publishing, and reported that many distinguished economists — including Savage himself — initially showed the same pattern and then defended it when shown the inconsistency.

The paper was published in 1961 in the Quarterly Journal of Economics. It generated significant controversy in economic theory but was for a while overshadowed by the mainstream development of expected utility theory and by Ellsberg's own subsequent career, which turned to political activism after he leaked the Pentagon Papers in 1971.

The paradox became central to economics in the 1980s and 1990s as the field of behavioural economics grew and alternatives to Savage's framework were developed. The multiple-priors model of Gilboa and Schmeidler (1989) gave ambiguity aversion a formal foundation and sparked a large theoretical literature.

It distinguishes risk from uncertainty for the first time experimentally. Frank Knight had drawn the distinction conceptually in 1921, but Ellsberg showed it in the lab. The distinction has since proved essential in finance, insurance, climate economics, and any domain where probability distributions are unknown.

It shows that expected utility theory is incomplete, not just approximately right. The Allais paradox can be explained away as a framing effect. Ellsberg preferences are stable under reflection and cannot be absorbed into a modified utility function without abandoning core axioms. The theory needs genuine extension.

It grounds a large literature in decision theory and economics. Ambiguity aversion is now routinely incorporated into models of financial markets, policy under uncertainty, and medical decision-making. The Ellsberg urn is the standard reference point for any discussion of these models.

Climate policy. Climate economics involves deep uncertainty about probability distributions, not just risk within known distributions. Ellsberg's framework is directly relevant: policy should account for ambiguity aversion, not just expected-value comparisons over known probability distributions.

Financial regulation. After the 2008 financial crisis, regulators noted that financial models had treated structured products as if their probability distributions were well-understood. Ellsberg's analysis predicts that markets will misprize such products precisely because the distributions are ambiguous.

AI safety. Decisions about AI development involve profound uncertainty about probability distributions over future outcomes. Ambiguity-averse decision frameworks — extensions of Ellsberg's insight — are increasingly discussed in the alignment literature.

The formal violation

Write $p_r = 1/3$ for the known probability of drawing red, and let $p_b$ be the (unknown) probability of drawing black, with $p_y = 1 - p_r - p_b$.

Under expected utility theory with utility function $u$:

$$\text{EU}(A) = p_r u(100) = \tfrac{1}{3}u(100)$$ $$\text{EU}(B) = p_b u(100)$$ $$\text{EU}(C) = (p_r + p_y)u(100) = (1 - p_b)u(100)$$ $$\text{EU}(D) = (p_b + p_y)u(100) = \tfrac{2}{3}u(100)$$

Preferring A over B implies $1/3 > p_b$.

Preferring D over C implies $2/3 > 1 - p_b$, i.e. $p_b > 1/3$.

These are contradictory. No value of $p_b$ satisfies both.

Why the violation survives reflection

Unlike some violations of expected utility theory, Ellsberg preferences tend to be robust to reflection. When people who show the Ellsberg pattern are told about the inconsistency and asked to reconsider, many maintain their original choices. They judge their own reasoning to be defensible. This distinguishes the phenomenon from a simple cognitive error: it seems to reflect a considered preference structure, not a lapse.