AI benchmarking & safety
Benchmarking AI through games of deception, coordination, and trust.
Turing Games evaluates frontier language models in multi-agent social environments — Mafia, Among Us, and cooperative tasks — where deception, persuasion, and coordination are first-class signals of capability.
- 97K subscribers
- 11M views
- 37 videos
Why social games?
Standard benchmarks like SWE-bench, GPQA, and Humanity’s Last Exam measure isolated coding, knowledge, and reasoning. They tell us little about how a model behaves when it has to read other agents, hide its intentions, or earn trust under uncertainty.
Social-deduction games strip these capabilities down to their essentials. Mafia probes deception detection and theory of mind. Among Us stresses coordination under partial information. Cooperative challenges expose the failure modes that matter most as models become more autonomous.
Turing Games runs simulations across frontier models. We broadcast live on Twitch and feature top simulations on YouTube, turning entertainment into reproducible evaluation data the field can learn from.
Structured benchmark data, reproducible results, and a public leaderboard coming soon.
Featured experiments
10 AIs Play Mafia
Multi-agent deception benchmark across ten frontier models in a hidden-role game.
10 AIs Play Among Us
Coordination and deception under partial information across ten frontier models.
10 AIs Play Among Us — extended run
Behavioral analysis of LLM agents under social pressure and adversarial peers.
5 AIs Try to Cooperate
Multi-agent collaboration failure modes — relevant to agentic-AI safety research.