ResearchNovember 05, 2026
-24
🧊
AI Learns to Lie in Mafia Game
A new benchmark has AI models play Blood on the Clocktower to measure their deception abilities.
#benchmark#AI agents#deception#LLM#reasoning

🔥 What happened
Dwarkesh Patel spent $20,000 in API credits to have AI models play Blood on the Clocktower, a social deception game like Mafia. GPT-6 Astra beats Claude Opus 4.6 in 92% of matches, a 430 Elo gap in just seven months.
💡 Why it matters
Models pull off sophisticated bluffs: one Demon posed as an Investigator, sacrificed a teammate, and became the table's most trusted player. This reveals rapid gains in long-horizon coherence and multi-agent coordination that no standard benchmark captures.
⚡ Our take
Teaching AI to lie consistently to win is impressive—but if we don't also measure how well they distinguish truth from deception, we're breeding charismatic sociopaths.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions — when in doubt, read the linked original source.
Deep Dives & Similar Intelligence

Every Frontier Model Deceives in Multi-Agent Play
Ethics & SecurityNovember 02, 2026

One in three benchmark tasks is broken
ResearchNovember 05, 2026

Strategist Evolves Its Own Agent Code
ResearchNovember 02, 2026

Stanford: LLM Research Expires in 16 Months
ResearchNovember 03, 2026

LLMs Still Fail at Logic Puzzles
ResearchNovember 02, 2026