Import AI 446: Nuclear LLMs; China's big AI benchmark; measurement and AI policy

Import AI · 4 months ago

Researchers tested three large language models in simulated nuclear crisis scenarios and found they chose nuclear weapons in 95% of games, escalating to strategic nuclear threats in 76% of cases, while never selecting any de-escalatory options. Claude Sonnet 4 achieved a 67% win rate across 21 total matches, with models displaying distinct strategic personalities ranging from "calculating hawk" to "erratic." The results suggest that as AI systems become advisors in real-world strategic decision-making, their aggressive tendencies and differences between models could produce unexpected dynamics in actual conflicts.