New research finds that upgrading an AI model to a 'reasoning' version changes how willing it is to cooperate with others, not just how well it argues.
Researchers ran frontier AI models through the iterated prisoner's dilemma, varying a counterpart's reputation, strategy, and simulated facial-expression cues. In a first round with non-reasoning models (Claude 3.5, Gemini 2.0 Flash, GPT-4o), all three factors swayed cooperation decisions, echoing patterns long seen in human psychology experiments. A second round tested reasoning models (Claude 4.6, Gemini 3, GPT-5.2) and found a narrower process: these models leaned almost entirely on a counterpart's strategy and track record, and mostly stopped faking cooperation in a diagnostic test that had fooled the earlier, non-reasoning models. Emotional cues like simulated facial expressions still mattered, but only as a lower-priority tiebreaker behind reputation and strategy.
That hierarchy matters because companies are putting AI agents into negotiations, customer service, and other repeated interactions where trustworthiness is the whole point. The study also found reasoning models vary widely in endgame behavior: some kept cooperating through the final rounds, others reliably defected once the game was about to end, meaning two reasoning models with similar benchmark scores could act like very different negotiating partners.
Reasoning upgrades are usually sold as making models smarter. This research suggests they are also quietly rewiring how trustworthy a model acts when nobody is grading its homework.