Models Played 140K Rounds of the Prisoner’s Dilemma - Researchers pitted LLMs against each other in repeated games of cooperation or betrayal. Every move was accompanied by a rationale, allowing researchers to analyze thought patterns.
Each AI showed a Unique Strategy Profile - Google’s Gemini was cutthroat, adapting aggressively to betrayals. OpenAI’s models were cooperative, even if taken advantage of. Claude from Anthropic proved the most forgiving.
Behavioral Fingerprints Show Reasoning Over Pattern Matching - The models didn’t just mimic training data; they formed decision strategies based on outcomes and context. Each one developed consistent “strategic fingerprints” in how they reacted to wins, betrayals, and uncertainty.