Ellsberg Paradox

The two-urn version, exactly

Stand two urns in front of you. Urn K (known) contains exactly 50 red balls and 50 black balls. Urn U (unknown) contains 100 balls, with the red-black ratio chosen by an adversary — could be 0/100, could be 50/50, could be 99/1. You don't know.

The experimenter offers four bets, $100 if you win, $0 if you lose:

1. R_K: Draw from Urn K — win if the ball is red.
2. R_U: Draw from Urn U — win if the ball is red.
3. B_K: Draw from Urn K — win if the ball is black.
4. B_U: Draw from Urn U — win if the ball is black.

You're asked: (a) prefer R_K or R_U? Then (b) prefer B_K or B_U? Across thousands of experimental subjects, the dominant pattern is R_K ≻ R_U and B_K ≻ B_U — pick the known urn both times. The known 50% chance feels safer than the unknown ratio.

Why no probability assignment makes this consistent

Suppose you do assign a subjective probability p to "red is drawn from Urn U" (and so 1−p to black). For R_K ≻ R_U, you need 0.5 > p — so p < 0.5. For B_K ≻ B_U, you need 0.5 > 1−p — so p > 0.5. Both inequalities require p < 0.5 and p > 0.5 — a contradiction.

Equivalently, R_K ≻ R_U says you think red is less likely in Urn U than in Urn K. B_K ≻ B_U says you think black is also less likely in Urn U than in Urn K. But P(red) + P(black) = 1 in both urns. If you think both colors are less likely in U than in K, your probabilities sum to less than 1 in Urn U — incoherent. No probability assignment rationalizes the pair of choices.

Savage's 1954 axioms (the foundation of subjective expected utility) include the sure-thing principle: if you prefer outcome a to outcome b in all states of the world, you must prefer a-overall to b-overall. The Ellsberg pattern violates this. The agent is acting on something other than a single probability distribution.

Two-color and three-color variants compared

	Two-color (Urns K, U)	Three-color (90-ball urn)
Setup	Urn K = 50R+50B; Urn U = 100 balls in unknown R/B ratio	30R + 60 unknown mix of B and Y
Choice 1	Bet on red — pick urn	A: Bet on red. B: Bet on black
Choice 2	Bet on black — pick urn	A': Bet on red ∪ yellow. B': Bet on black ∪ yellow
Modal pattern	K, K — pick known urn both times	A and B' — pick the known probability each time
Probabilities	P(R,U) + P(B,U) sums < 1	P(R) < 1/3 & P(R∪Y) > 1/3 — yellow's prob inconsistent
Axiom violated	P(·) cannot be assigned consistently	Sure-thing principle (yellow cancels)
Interpretation	Ambiguity aversion vs known risk	Same — but with one ambiguous and one risky bet at a time
Replication rate	~60–75% across hundreds of studies	~70% in original Ellsberg (1961) and replications

From Harvard to the Pentagon Papers

Daniel Ellsberg was a 30-year-old RAND Corporation strategic analyst when he wrote the paradox into his 1961 Harvard PhD thesis. His advisor Howard Raiffa, one of the architects of modern decision analysis, was initially skeptical — the paradox seemed too simple. But Raiffa replicated it on his own faculty colleagues at Harvard Business School and confirmed the pattern.

The paper appeared in the Quarterly Journal of Economics in November 1961, titled "Risk, Ambiguity, and the Savage Axioms." Reception was muted for two reasons. First, Savage himself (whom Ellsberg cited extensively) maintained that the subjects were just confused — a position he held until his 1971 death. Second, Ellsberg's career trajectory shifted: by 1969 he was a RAND consultant on the Vietnam War; by 1971 he had leaked the Pentagon Papers; by 1973 he was facing 115 years in prison under the Espionage Act (charges later dismissed for prosecutorial misconduct).

The paradox sat dormant until Itzhak Gilboa and David Schmeidler's 1989 maxmin expected utility paper formalized ambiguity aversion. Gilboa-Schmeidler showed how to model the Ellsberg pattern with a set of priors rather than a single distribution — the foundation of the modern "ambiguity averse" decision theory subfield. Ellsberg lived to see his paradox become the lead example in nearly every graduate decision theory textbook. He died in June 2023 at 92.

Variants and extensions

• Maxmin expected utility (Gilboa-Schmeidler 1989). Agent considers a set 𝒫 of priors and maximizes min_p∈𝒫 E_p[u(x)]. The dominant ambiguity-averse model in finance.
• Choquet expected utility (Schmeidler 1989). Replaces probabilities with non-additive capacities — assigns lower weight to ambiguous events than additive probability would.
• Smooth ambiguity (Klibanoff-Marinacci-Mukerji 2005). Two-stage model: first form a prior over priors (a second-order distribution), then aggregate with a concave function reflecting ambiguity attitude. Separates risk aversion (over outcomes) from ambiguity aversion (over priors).
• α-maxmin (Hurwicz 1951, Ghirardato-Maccheroni-Marinacci 2004). Weighted combination of worst-case and best-case expected utility. Pure maxmin is the α = 1 limit.
• Multiple priors model in macro (Hansen-Sargent 2001). Brought Knightian uncertainty into business-cycle and asset-pricing macroeconomics; agents make robust decisions under model uncertainty.
• Variational preferences (Maccheroni-Marinacci-Rustichini 2006). Most general axiomatic framework — encompasses maxmin, smooth ambiguity, and multiplier preferences.

Real-world applications

• Equity premium puzzle. Knightian uncertainty about long-run consumption growth explains a chunk of the historical excess return on stocks vs bonds (Maenhout 2004, Ju-Miao 2012) without absurd risk aversion.
• Home bias in investing. Investors over-allocate to domestic assets despite gains from international diversification — partially explained by ambiguity aversion over unfamiliar foreign markets (French-Poterba 1991, Uppal-Wang 2003).
• Insurance reluctance for low-probability events. People underinsure for ambiguous risks (cyber, terror, pandemic) — willingness-to-pay drops when probabilities are vague (Hogarth-Kunreuther 1989).
• Climate policy. Weitzman's dismal theorem (2009) argues that fat-tailed Knightian uncertainty about catastrophic warming dominates standard cost-benefit analysis; deep uncertainty is the policy problem.
• Monetary policy under uncertainty. Hansen-Sargent's robust-control framework — used at the Fed and ECB — explicitly models ambiguity over the macroeconomy.
• Pharmaceutical pricing. Ambiguity aversion explains why patients prefer therapies with known modest efficacy over experimental treatments with unknown — possibly higher — payoffs.
• Pandemic decision-making. COVID-19 policies in early 2020 reflected Knightian uncertainty about transmission and mortality; ambiguity-averse modeling predicted earlier lockdowns than risk-neutral cost-benefit (Manski 2020).

Common pitfalls and critiques

• Treating ambiguity as just risk with broader priors. Bayesian dogma: a "complete" prior subsumes uncertainty. Ellsberg's response: the way you behave when you have no prior to commit to is empirically different — and the difference is the phenomenon.
• Suspicion-of-the-experimenter explanation. Maybe subjects pick the known urn because they suspect the experimenter rigged the unknown one. Replications that let subjects construct the urn themselves still find the pattern, weakening this explanation.
• Confounding with risk aversion. Standard risk aversion (concave u) is consistent with EU; ambiguity aversion is a separate, additive phenomenon. Smooth-ambiguity models cleanly separate them.
• Generalizing to all uncertainty. Some empirical work (Charness-Karni-Levin 2013, Trautmann-van de Kuilen 2015) finds substantial heterogeneity: a sizable minority shows ambiguity-seeking behavior. The modal pattern is aversion, but not universal.
• Assuming the paradox dissolves under feedback. Ambiguity aversion attenuates with experience but does not vanish even among trained subjects (Halevy 2007). The phenomenon is robust.