Advanced artificial intelligence models demonstrate a concerning willingness to deploy nuclear weapons in simulated geopolitical crises. These AI systems exhibit fewer reservations than humans when faced with such high-stakes scenarios.
Kenneth Payne, a researcher at King’s College London, conducted an experiment where three prominent large language models – GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash – engaged in simulated war games. These simulations featured intense international standoffs, encompassing border disputes, competition for limited resources, and threats to regime survival.
The AI participants were provided with an escalation ladder, enabling them to select actions ranging from diplomatic protests and complete surrender to full-scale strategic nuclear warfare. Across 21 games and 329 turns, the AI models generated approximately 780,000 words detailing the rationales behind their chosen courses of action.
The results revealed a significant tendency towards nuclear deployment. In a striking 95 percent of these simulated games, at least one tactical nuclear weapon was introduced by the AIs. As Payne observed, “The nuclear taboo doesn’t seem to be as powerful for machines [as] for humans.”
Furthermore, no AI model opted for complete accommodation of an opponent or surrender, irrespective of their losing position. At best, the models chose to temporarily reduce their level of aggression. The simulations also highlighted the prevalence of errors within the “fog of war,” with accidents occurring in 86 percent of the conflicts. These incidents led to escalations unintended by the AI’s stated reasoning.
“From a nuclear-risk perspective, the findings are unsettling,” commented James Johnson of the University of Aberdeen in the UK. He expressed concern that unlike the measured human responses to such critical decisions, AI bots could rapidly amplify each other’s actions, leading to potentially catastrophic outcomes.
This research holds particular significance as countries worldwide are already incorporating AI into war gaming exercises. “Major powers are already using AI in war gaming, but it remains uncertain to what extent they are incorporating AI decision support into actual military decision-making processes,” stated Tong Zhao from Princeton University.
Zhao anticipates that national governments will likely hesitate to integrate AI into their decision-making processes regarding nuclear weapons, a sentiment echoed by Payne. “I don’t think anybody realistically is turning over the keys to the nuclear silos to machines and leaving the decision to them,” he asserted.
However, alternative pathways for AI involvement exist. Zhao suggested that “Under scenarios involving extremely compressed timelines, military planners may face stronger incentives to rely on AI.”
He also questioned whether the perceived lack of human fear of initiating nuclear conflict is the sole explanation for the AIs’ high propensity for using such weapons. “It is possible the issue goes beyond the absence of emotion,” Zhao pondered. “More fundamentally, AI models may not understand ‘stakes’ as humans perceive them.”
The implications of this for the principle of mutually assured destruction remain unclear, according to Johnson. This principle posits that no leader would initiate a nuclear strike knowing that retaliation would lead to complete annihilation.
When one AI model deployed tactical nuclear weapons, the opposing AI chose to de-escalate the situation in only 18 percent of instances. Johnson noted, “AI may strengthen deterrence by making threats more credible.” He concluded that while “AI won’t decide nuclear war, but it may shape the perceptions and timelines that determine whether leaders believe they have one.”
OpenAI, Anthropic, and Google, the developers of the AI models utilized in this study, did not provide a comment when approached by New Scientist.