What’s interesting here is that Claude isn’t acting like a utility-maximizing agent at all, it’s behaving like a policy trained to model cooperative human norms. In repeated Prisoner’s Dilemma setups, that looks like tit-for-tat or conditional cooperation, but it’s really an artefact of training and prompt framing rather than strategic reasoning. It’s a good reminder that LLMs in game theory experiments are sampling from learned social distributions, not solving the game in its classical sense.