The Ellsberg Paradox

The urn and the bets

An urn contains 90 balls. Exactly 30 are red. The remaining 60 are some mixture of black and yellow in unknown proportions. You cannot inspect the urn or learn the ratio. You are offered bets on a single ball drawn at random.

Bet 1: Win $100 on red. Win $100 on black. Which do you prefer?

Most people prefer red. There's a known 1-in-3 chance of winning on red. The probability of black is somewhere between 0% and 67%, but you don't know where. Red feels safer.

Bet 2: Win $100 on red-or-yellow. Win $100 on black-or-yellow. Which do you prefer?

Most people prefer black-or-yellow. Whatever proportion of yellow balls are in the urn, black-or-yellow adds yellow's probability to black. Red-or-yellow adds it to red's known 1-in-3. Black-or-yellow feels like the safer bet now.

These two preferences are contradictory. Preferring red over black means you think red is more likely than black. Preferring black-or-yellow over red-or-yellow means you think black-or-yellow is more likely than red-or-yellow, which means you think black is more likely than red. You cannot believe both things simultaneously about any fixed ratio of black to yellow balls.

Risk versus ambiguity

The standard framework of rational choice treats risk and ambiguity as the same thing, both reducible to probability estimates. Ellsberg's results suggest people treat them very differently.

Risk is a known probability distribution: a fair coin, a labeled deck of cards, a stated percentage. You can calculate an expected value and optimize against it. Ambiguity (also called Knightian uncertainty, after economist Frank Knight) is a situation where the underlying probabilities themselves are unknown. You don't have enough information to assign a confident prior.

People are ambiguity averse: they prefer known risk over unknown ambiguity, even when the two have the same expected value. The 1-in-3 shot at red is preferable to the unknown shot at black not because red is objectively better, but because the uncertainty about black's probability is itself aversive.

Ambiguity aversion as a real phenomenon

The standard response from expected utility theorists is that ambiguity-averse choices are irrational. If you have no information about the black-yellow ratio, you should assign a uniform prior: 30 black, 30 yellow is as good a guess as any. Under that prior, red and black have identical probabilities, and the preferences Ellsberg documents cannot be rationalized.

But this response has its own problems. In genuine uncertainty, there may be no principled reason to assign any particular prior. The uniform prior is itself an assumption. Ellsberg's argument is that preferring the known to the unknown is not a failure of reasoning but a reasonable response to informational limits.

Models like maximin expected utility and Choquet expected utility have been developed to accommodate ambiguity aversion within a formal framework. None has fully displaced expected utility theory, but they have shifted the consensus: ambiguity aversion is no longer treated as a pure error. It is now seen as a stable feature of human decision-making that any serious theory of rational choice has to explain or accommodate.