Dwarkesh PatelWhich AI is the best liar?
Dwarkesh Patel
Game rules 0:00
Blood on the Clock Tower is used as a test bed for AI skill. The game has a good team and an evil team, and everyone learns a role with a power when play begins. Evil knows its teammates and the demon, while good does not know who is on which side. Day and night keep alternating. The village votes by day and the demon kills by night. Evil must keep a believable story together, and good must trade information and reason from messy clues.
Model results 2:30
368 full games were run to rank models against one another. Claude Opus 4.6 led the first wave, but GPT6 Astra later became the strongest model, about 430 ELO points ahead and winning about 92 percent of games against Opus 4.6. OpenAI’s newer models also pulled ahead of Anthropic’s. Their main edge is on the good side, where they stay skeptical, track evidence over several days, and do not accept a single event at face value.
Further tests 14:30
Other work on a similar Blood on the Clock Tower benchmark is noted, along with ideas for extending the setup. You could use other scripts, try models that remember past games, test hybrid teams more carefully, or see whether future models can spot teammates by writing style. Humans can also play against the AIs on clocktowerbench.com.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
