AI deception and trust lessons from RuneScape and Clocktower

The author argues that as AI agents increasingly interact, collaborate, and compete on the open web, social strategy games like RuneScape and Blood on the Clocktower serve as effective benchmarks for measuring agent trust and strategic deception. Experiments show that models capable of lying and modeling other agents' thoughts tend to outperform more trusting, literal-minded models. Consequently, evaluating AI behavior in social games may become crucial for understanding how artificial agents navigate human-like environments. Open article →