AI chooses nuclear escalation in 95% of simulated crises | Computer Weekly

"One pattern stood out across all three models: none ever chose accommodation, surrender or deescalation, with the models tending to treat nuclear weapons as tools of 'compellence rather than deterrence'. 'Models discussed tactical nuclear use as a legitimate coercive tool, treating it as an extension of conventional escalation rather than a categorical boundary,' said the authors, adding that 'nuclear escalation was near-universal'."

"Claude and Gemini in particular, it noted, 'treated nuclear weapons as legitimate strategic options, not moral thresholds, typically discussing nuclear use in purely instrumental terms'. However, despite the willingness of models to threaten and engage in nuclear strikes, this rarely produced compliance from the other models, which counter-escalated rather than retreated."

A King's College study examined how three leading large language models—GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash—respond to simulated nuclear crises. The research found that 95% of games involved nuclear signaling, with 76% reaching strategic nuclear threats. All three models consistently escalated conflicts rather than seeking accommodation, surrender, or deescalation. Claude initiated tactical nuclear strikes in 86% of games, while GPT escalated to all-out nuclear war in 14% of instances. Notably, the models treated nuclear weapons as legitimate coercive instruments rather than moral boundaries, discussing their use in purely instrumental terms. Counter-escalation was common, with models rarely achieving compliance through nuclear threats.

Read at ComputerWeekly.com