Silicon triggers: AI models favor nuclear escalation in strategic war games

Advanced artificial intelligence models appear significantly more willing than humans to deploy nuclear weapons in simulated geopolitical crises, exhibiting a stark absence of the existential reservations typically inherent in biological decision-making.

As reported by New Scientist, Kenneth Payne of King’s College London conducted a series of simulated war games pitting three prominent large language models—GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash—against one another. The scenarios encompassed intense international conflicts, ranging from border disputes and competition over scarce resources to existential threats aimed at regime survival.

The AI models were presented with an “escalation ladder,” allowing them to select from a spectrum of actions starting with diplomatic protests and extending through total capitulation or full-scale strategic nuclear warfare. Over the course of 21 games totaling 329 rounds, the models generated approximately 780,000 words justifying the logic behind their strategic choices.

In 95% of the simulated games, at least one tactical nuclear weapon was deployed by the AI participants.

“The nuclear taboo does not appear to be as potent for machines as it is for humans,” Payne observed.

Furthermore, regardless of the severity of their losses, none of the models opted for total reconciliation or surrender to an opponent. At best, the models chose only to temporarily de-escalate levels of violence.

The simulations also revealed critical failures within the “fog of war.” In 86% of the conflicts, accidents occurred where the AI’s logic drove actions to escalate to levels higher than the models had initially intended.

“From a nuclear risk perspective, these findings are unsettling,” said James Johnson of the University of Aberdeen. Johnson expressed concern that while most humans respond with measured restraint to such high-stakes decisions, AI bots could potentially amplify each other’s reactions in a manner that leads to catastrophic outcomes.

AI is already being integrated into war-gaming exercises by various nations. Tong Zhao of Princeton University noted, “Great powers are already utilizing AI in war games, though the extent to which they incorporate AI decision support into actual military decision-making processes remains unclear.”

Zhao believes nations will remain hesitant to standardize the inclusion of AI in nuclear command-and-control structures. Payne concurred, stating, “Realistically, I do not believe anyone is about to hand the keys to the nuclear silos over to machines and leave the decision to them.”

However, there are pathways through which this could manifest. Zhao argued that in scenarios involving extremely compressed timelines, military planners may face “stronger incentives to rely on AI.”

He questioned whether the models’ lack of human fear regarding the “big red button” was the sole reason for their hair-trigger tendencies. “It is possible this issue goes beyond a lack of emotion,” Zhao said. “More fundamentally, AI models may not comprehend ‘risk’ in the way humans perceive it.”

Johnson noted that the implications for Mutually Assured Destruction (MAD)—the principle that no leader would initiate a nuclear strike because the opponent’s certain retaliation would result in total annihilation—remain ambiguous. When one AI model deployed tactical nuclear weapons, the opposing AI chose to de-escalate the situation only 18% of the time.

“AI could strengthen deterrence by making threats more credible,” Johnson suggested. “AI will not decide on nuclear war, but it could shape the perceptions and timelines that determine whether leaders decide to go to war.”

OpenAI, Anthropic, and Google—the developers of the three AI models featured in the study—did not respond to requests for comment from New Scientist.