The creator describes an experiment in which AI models played Blood on the Clocktower, a social deduction game with hidden teams and asymmetric information. The reported games compare modelsEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria. playing both sides and illustrate bluffing, sacrifice and collective reasoningA multi-agent system contains multiple AI agents that interact, coordinate, divide work, or influence one another while pursuing tasks..
Examples show an evil team gaining trust by sacrificing an ally and exploiting a hidden role change. Another comparison illustrates how models interpreted a failed attempt to identify an opponent. The results reflect the creator’s game setup rather than a general measure of deception in every setting.
The creator explains how a programmatic storyteller and repeated AI-assisted revisions shaped the game environment. Suggested further tests include mixed-model teams, learning from prior games and play against humans; the mixed-team experiment is described as small and inconclusive.
Watch on YouTube
