A new study from early July 2025 investigated whether large language models (LLMs) like ChatGPT or Gemini can truly act strategically – meaning they make smart decisions in competitive situations. To test this, researchers used the classic Iterated Prisoner’s Dilemma (IPD), a thought-provoking game where choices build reputations and influence future moves.
In these experiments, the authors set up tournaments where well-known strategies (like Tit‑for‑Tat or Grim Trigger) competed with AI models from OpenAI, Google, and Anthropic. By varying “how long the game might last” (called the shadow of the future), they added layers of unpredictability and complexity.
The results were fascinating:
- Google’s Gemini acted aggressively – sometimes ruthless – capitalising on cooperative opponents but swift to punish unfair moves.
- OpenAI’s models leaned heavily towards cooperation, which worked well in friendly settings but left them exposed when others exploited them.
- Anthropic’s Claude stood out as a forgiving operator, willing to rebuild trust even after being betrayed.
The researchers also examined nearly 32,000 internal explanations from the models. These rationales revealed thoughtful reasoning about game length and opponents’ likely strategies—clear signs of genuine strategic thinking.
This study bridges classic game theory with what they call machine psychology. It offers a detailed look at how LLMs make choices under uncertainty – showing that these systems can genuinely strategise, not just mimic language patterns
