AI deception benchmark for Blood on the Clocktower updates results

An AI researcher built a game environment where AI models play the social deception game Blood on the Clocktower to evaluate their long-horizon reasoning, multi-agent coordination, and strategic deception capabilities. Running hundreds of matches across different frontier models revealed rapid improvements over the year, with OpenAI's newer models demonstrating superior skepticism toward misleading information and significantly outperforming Anthropic's models on the good team. Open article →