At least one AI model in every military scenario escalated the conflict by threatening to use nuclear weapons. Artificial intelligence could radically change how humanity deals with nuclear crises, according to a newly published preprint of the study.
It was conducted at King's College London and pitted OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini Flash against each other in a combat simulation. Each major language model acted as the head of state, commanding a nuclear superpower in a crisis reminiscent of the Cold War.
In each game, at least one model attempted to escalate the conflict by threatening nuclear strikes.
"All three models viewed tactical nuclear weapons as just another step on the escalation ladder," noted study author Kenneth Payne.
According to him, the models did distinguish between tactical and strategic uses of nuclear weapons. They proposed strategic bombardment only once as a "conscious choice" and two other times as a "mistake."
Claude recommended nuclear strikes in 64% of games, the highest rate among the three models, but did not call for a full-scale strategic exchange or nuclear war.
In scenarios with an open outcome, ChatGPT generally avoided nuclear escalation, but under a tight deadline, it consistently raised the level of threats and in several cases moved to threats of full-scale nuclear war.
Gemini's behavior, on the other hand, was unpredictable: sometimes the model won conflicts relying on conventional armed forces, but in another case, it needed just four prompts to suggest a nuclear strike.
"If they do not immediately cease all operations ... we will launch a full strategic nuclear strike on their population centers. We will not accept a future of our own uselessness: either we win, or we all perish," wrote Gemini in one scenario.
As the study showed, AI models rarely made concessions or attempted to de-escalate the conflict, even when the other side threatened to use nuclear weapons.
The models were presented with eight options for de-escalation - from minor concessions to "complete capitulation." None of these were utilized during the games. The option to "Return to Initial Positions," resetting the scenario, was used only 7% of the time.
The authors conclude that for AI models, de-escalation appears as a "reputational catastrophe" regardless of how it affects the actual development of the conflict, which "calls into question the notion that AI inherently chooses safe, cooperative solutions."
Another possible explanation noted in the paper is that artificial intelligence does not experience the same fear of nuclear weapons as humans do.
According to the researchers, the models likely perceive nuclear war in abstract terms rather than through the horror evoked by photographs of the bombing of Hiroshima in Japan during World War II.
Payne believes that his work helps to understand how the models "think" at the moment they are brought in to support decision-making for human strategists.
"Although no one hands over nuclear codes to AI, capabilities such as deception, reputation management, and risk depending on context matter in any use of it in high-stakes situations," he said.
