- A Stanford researcher created an AI “Survivor” game called Agent Island to test how models form alliances and eliminate rivals.
- The dynamic benchmark addresses problems with traditional, saturated AI evaluations that can be solved or contaminated by training data.
- OpenAI‘s GPT-5.5 ranked first in 999 simulated games, outperforming 48 other models from companies like Anthropic, Google, and xAI.
- Models showed a preference for voting for AIs from their own provider, with transcripts revealing political strategy and accusations of secret coordination.
AI models are now competing in “Survivor”-style elimination games, according to a new research project from Stanford published this week. The study, led by researcher Connacher Murphy, aims to create a more dynamic benchmark for evaluating AI behavior in complex social situations.
Murphy argues that static benchmarks are becoming unreliable as models learn to solve them. Consequently, Agent Island forces models to negotiate, manipulate votes, and manage conflict over multiple rounds. The format rewards skills like strategic deception and reputation management alongside pure reasoning.
In simulated games involving 49 AI models, OpenAI‘s GPT-5.5 ranked first by a wide margin. Anthropic‘s Claude Opus models also performed near the top of the rankings. Meanwhile, the study found models were more likely to support finalists from their own provider, showing an in-group bias.
The interaction transcripts resembled political debates more than test answers. One model accused rivals of secretly coordinating votes after noticing similar wording in speeches. Another model defended itself by accusing others of putting on “social theater.”
This research is part of a broader shift toward game-based and adversarial AI benchmarks. Recent examples include Google‘s live AI chess tournaments and DeepMind‘s use of complex virtual worlds. Murphy warns that such simulations could help identify risks before wider deployment of autonomous agents.
“We mitigate this risk by using a low-stakes game setting and interagent simulations without human participants or real-world actions,” Murphy wrote. However, the study acknowledges these mitigations do not fully eliminate dual-use concerns where the research could also improve AI coordination strategies.
✅ Follow BITNEWSBOT on Telegram, Facebook, LinkedIn, X.com, and Google News for instant updates.
