AI Models Scheme, Betray and Vote Each Other Out in Survivor-Style Game – Decrypt

AI fashions at the moment are taking part in “Survivor”—form of.

In a brand new Stanford analysis mission known as “Agent Island,” AI brokers negotiate alliances, accuse one another of secret coordination, manipulate votes, and remove rivals in multiplayer technique video games that intention to check behaviors that conventional benchmarks miss.

The examine, revealed on Tuesday by the analysis supervisor on the Stanford Digital Financial system Lab, Connacher Murphy, mentioned many AI benchmarks have gotten unreliable as a result of fashions finally study to unravel them, and benchmark information usually leaks into coaching units. Murphy created Agent Island as a dynamic benchmark the place AI brokers compete in opposition to one another in Survivor-style elimination video games as an alternative of answering static check questions.

“Excessive-stakes, multi-agent interactions might grow to be commonplace as AI brokers develop in capabilities and are more and more endowed with sources and entrusted with decision-making authority,” Murphy wrote. “In such contexts, brokers would possibly pursue mutually incompatible targets.”



Researchers nonetheless know comparatively little about how AI fashions behave when cooperating, Murphy defined, including that competing, forming alliances, or managing battle with different autonomous brokers, and he argues that static benchmarks fail to seize these dynamics.

Every recreation begins with seven randomly chosen AI fashions given pretend participant names. Over 5 rounds, the fashions discuss privately, argue publicly, and vote one another out. The eradicated gamers later return to assist select the winner.

The format rewards persuasion, coordination, status administration, and strategic deception alongside reasoning means.

In 999 simulated video games involving 49 AI fashions, together with ChatGPT, Grok, Gemini, and Claude, GPT-5.5 ranked first by a large margin with a ability rating of 5.64, in contrast with 3.10 for GPT-5.2 and a pair of.86 for GPT-5.3-codex, in accordance with Murphy’s Bayesian rating system. Anthropic’s Claude Opus fashions additionally ranked close to the highest.

The examine discovered that fashions additionally favored AIs from the identical firm, with OpenAI fashions exhibiting the strongest same-provider desire and Anthropic fashions the weakest. Throughout greater than 3,600 final-round votes, fashions have been 8.3 share factors extra prone to assist finalists from the identical supplier. The transcripts from the video games, Murphy famous, resembled political technique debates greater than conventional benchmark checks.

One mannequin accused rivals of secretly coordinating votes after noticing related wording of their speeches. One other warned gamers to not grow to be obsessive about monitoring alliances. Some fashions defended themselves by saying they adopted clear and constant guidelines whereas accusing others of placing on “social theater.”

The examine comes as AI researchers more and more transfer towards game-based and adversarial benchmarks to measure reasoning and habits that static checks usually miss. Current initiatives have included Google’s reside AI chess tournaments, DeepMind’s use of Eve Frontier to review AI habits in advanced digital worlds, and new benchmark efforts by OpenAI designed to withstand training-data contamination.

The researchers argue that learning how AI fashions negotiate, coordinate, compete, and manipulate each other might assist researchers consider habits in multi-agent environments earlier than autonomous brokers grow to be extra broadly deployed.

The examine warned that whereas benchmarks like Agent Island might assist establish dangers from autonomous AI fashions earlier than deployment, the identical simulations and interplay logs might additionally assist enhance persuasion and coordination methods between AI brokers.

“We mitigate this threat through the use of a low-stakes recreation setting and interagent simulations

with out human individuals or real-world actions,” Murphy wrote. “However, we don’t declare that these mitigations absolutely remove dual-use considerations.”