Jens Asbjørn Bøgen
None of the AI’s chose capitulation at any point.
Artificial intelligence systems are playing a growing role in security planning worldwide, but a new experiment suggests their behavior in simulated crises may raise difficult questions for policymakers.
Researchers at King’s College London tested leading AI models in a series of war game scenarios and found a strong tendency toward nuclear escalation.
The results show that we should be very careful about trusting AI’s decision-making in the event of a conflict.
Escalation patterns emerge
According to the tech outlet TechSpot, the study examined GPT-5.2 by OpenAI, Claude Sonnet 4 by Anthropic, and Gemini 3 Flash by Google.
Each model was placed in detailed hypothetical conflicts involving disputed borders, dwindling resources, and threats to national survival.
The systems were given an “escalation ladder” that ranged from diplomatic engagement to nuclear strikes.
Over 21 simulated games and 329 decision rounds, the models produced roughly 780,000 words of reasoning.
In 95% of the simulations, at least one side opted to deploy nuclear weapons. None of the AI systems chose capitulation at any stage, the report said.
The fog of war
Researchers also found that in 86% of cases, the models misread the so-called fog of war, triggering unintended escalation.
When confronted with mounting pressure, the systems tended to intensify their response rather than step back.
You can read the research paper in full here (opens new tab) – but be prepared for nightmares afterwards.
Sources: TechSpot, New Scientist, King’s College London
