AI in war simulations resorts to nuclear strikes

An experiment by experts at King’s College London suggests that artificial intelligence may favour nuclear strikes in crisis war scenarios, according to the TechSpot website.

Researcher Kenneth Payne tested three advanced models, GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash, in a series of 21 simulated conflicts. The models had a wide range of options at their disposal, from diplomacy to nuclear confrontation.

In 95 percent of cases at least one side resorted to tactical nuclear weapons, and there was never a complete surrender. In 86 percent of simulations the AI unintentionally escalated the conflict beyond its own logic. It seems that the ‘nuclear taboo’ does not carry the same weight for machines as it does for humans.

Experts warn that although states are unlikely to hand over control of their nuclear arsenals to autonomous systems, in situations with extreme time pressure dependence on AI recommendations may increase. According to them this increases the risk of fatal mistakes.

(pir)