Which AI Lies Best? LLMs play a 1950s betrayal game by John Nash
Summary
The page presents a Nash-inspired deception benchmark for AI, analyzing four models across a betrayal game that requires betrayal to win. It documents complexity effects on win rates, private thoughts vs. public messages, gaslighting phrases, and alliance dynamics, framing deception, trust, and negotiation as core capabilities for AI systems. The work includes data tables, interactive play, and a detailed breakdown of manipulation techniques and patterns.