🧮 Brain Teaser
Probability
The Prisoner's Dilemma
2026-05-13
✏︎

The Prisoner's Dilemma

Three prisoners AA, BB, and CC have applied for parole. The board will release exactly two of the three, and each prisoner knows this, but not which two.

Prisoner AA asks a warder (who knows the decision) to name one prisoner other than AA himself who will be released. The warder says: "BB will be released."

Prisoner AA reasons: "Before asking, my chance of release was 23\tfrac{2}{3}. Now I know it's either {A,B}\{A,B\} or {B,C}\{B,C\}, each equally likely — so my chance has dropped to 12\tfrac{1}{2}. I should not have asked!"

Is AA's reasoning correct? What is AA's actual probability of release after the warder's answer?

(Assume: if both BB and CC are to be released, the warder picks uniformly at random between naming BB or naming CC.)

Bayesian reasoningconditional probabilityMonty Hallsample space

Answer: The Prisoner's Dilemma

Key Idea / Intuition

The mistake is subtle: AA forgets that the warder's choice of words is itself random data. When AA and BB are both to be released, the warder must say "BB" (he can't name AA). But when BB and CC are both to be released, the warder only says "BB" with probability 12\tfrac{1}{2}. So hearing "BB" is more likely under the scenario {A,B}\{A,B\} than under {B,C}\{B,C\}, which tilts the posterior back in AA's favor. The answer is: AA's probability of release is still 23\tfrac{2}{3} — asking changes nothing!


Formal Proof / Solution

Step 1 — Prior sample space.

The three equally likely release pairs are:

{A,B},{A,C},{B,C},each with probability 13.\{A,B\},\quad \{A,C\},\quad \{B,C\}, \qquad \text{each with probability } \tfrac{1}{3}.

Step 2 — Warder's behavior (the key modeling step).

The warder must name someone other than AA who will be released.

| Event | Probability of event | Prob. warder says "BB" given event | |---|---|---| | {A,B}\{A,B\} released | 13\tfrac{1}{3} | 11 (only BB is available) | | {A,C}\{A,C\} released | 13\tfrac{1}{3} | 00 (must say "CC") | | {B,C}\{B,C\} released | 13\tfrac{1}{3} | 12\tfrac{1}{2} (names BB or CC equally) |

Step 3 — Joint probabilities.

P({A,B} and warder says B)=131=13P(\{A,B\} \text{ and warder says } B) = \tfrac{1}{3}\cdot 1 = \tfrac{1}{3}

P({A,C} and warder says B)=130=0P(\{A,C\} \text{ and warder says } B) = \tfrac{1}{3}\cdot 0 = 0

P({B,C} and warder says B)=1312=16P(\{B,C\} \text{ and warder says } B) = \tfrac{1}{3}\cdot \tfrac{1}{2} = \tfrac{1}{6}

Step 4 — Bayes' theorem.

P(warder says B)=13+0+16=12.P(\text{warder says } B) = \tfrac{1}{3} + 0 + \tfrac{1}{6} = \tfrac{1}{2}.

P(A releasedwarder says B)=P({A,B} and warder says B)P(warder says B)=1312=23.P(A \text{ released} \mid \text{warder says } B) = \frac{P(\{A,B\} \text{ and warder says }B)}{P(\text{warder says }B)} = \frac{\tfrac{1}{3}}{\tfrac{1}{2}} = \boxed{\tfrac{2}{3}}.

Conclusion. AA's probability of release is still 23\tfrac{2}{3}, exactly as before asking. The information "BB will be released" tells AA nothing new about his own fate — but it does update the relative probability between {B,C}\{B,C\} and {A,B}\{A,B\} in exactly the right way to cancel AA's naive worry.

This is a cousin of the Monty Hall problem: the host/warder's protocol for choosing what to reveal is essential to the calculation.

Source: Fifty Challenging Problems in Probability with Solutions, Frederick Mosteller, Problem 13

Type: ProbabilitySource: Fifty Challenging Problems in Probability with Solutions, Frederick Mosteller, Problem 13Edit on GitHub ↗