Dwarkesh Patel builds Blood on the Clocktower benchmark for AI social deception
Article timeline
The articles in this story. Times are ET.
Dwarkesh PatelXDwarkesh Patel benchmarks AI models playing Blood on the Clocktower social deceptionAIDwarkesh Patel spent $20,000 in API credits having AI models play each other at Blood on the Clocktower, a social deception game. He says the models improved over the year at improvised bluffing and team reasoning to expose conspiracies, and that OpenAI's newer models outperform Anthropic's at the game. The full benchmark and a playable version against the AIs are available at clocktowerbench.com.
- Dwarkesh PodcastBlogDwarkesh Patel builds a Blood on the Clocktower benchmark testing AI deception
AIDwarkesh Patel says he built a benchmark around the social deception game Blood on the Clocktower, spending about $20,000 in API credits to run AI models against each other. He reports that GPT-6 Astra beats Claude Opus 4.6 by about 430 Elo points, winning roughly 92% of their games. He also says OpenAI's models have improved more at playing the good team, being more skeptical of evidence than Anthropic's models.