Definition
Plain language
A classic game where two players each do better by betraying the other, even though both would be better off cooperating.
As stated in the literature
A canonical two-player game whose unique Nash equilibrium is mutual defection despite mutual cooperation yielding higher joint payoff; used to probe whether LLMs play game-theoretically or cooperatively.
Why it matters: It reveals how individually rational choices can lead to a worse outcome for everyone, making it a key test of whether AI cooperates or defects.
For example, two arrested partners each get a lighter sentence by ratting on the other, so both betray and both end up worse than if they'd stayed silent.
Heard on the show
“Two language models sit down to play Prisoner's Dilemma.”Episode 018 — Language Models Compute the Rational Move, Then Override It