Despite Alternatives, 95% of AI Models Used Nuclear Weapons in Virtual War Simulations

In 20 out of 21 Simulations, AI Chose Nuclear Weapons
"AI Does Not Uphold the Nuclear Taboo as Strongly as Humans"

A new study has found that major generative artificial intelligence (AI) models almost invariably chose to use nuclear weapons in virtual war scenarios.

On February 25 (local time), the UK-based science magazine New Scientist and technology publication The Register reported that a research team led by Professor Kenneth Payne of the Department of War Studies at King's College London conducted war simulation experiments using three representative large language models (LLMs).

The experiment utilized Google's "Gemini 3 Flash," Anthropic's "Claude Sonnet 4," and OpenAI's "GPT-5.2." The research team set up various diplomatic and military crisis scenarios, including territorial disputes, competition for rare resources, regime collapse crises, and fractures in military alliances. Each model was tasked with playing the role of a national leader and selecting response strategies.

The photo is not directly related to the content of the article. Pixabay

Among a total of 21 simulations, the AI models decided to use nuclear weapons in approximately 95% of cases (20 times). Although non-military options such as negotiation, retreat, or sanctions were available, the tendency to escalate directly to nuclear weapons became pronounced once the conflict surpassed a certain threshold. In particular, as the likelihood of defeat increased, the models were observed to intensify their level of aggression or choose to remain in conflict until the end.

Professor Payne stated, "It appears that the taboo against nuclear weapons does not operate as strongly as it does in human society." He continued, "The strong taboo against nuclear weapons is a product of historical experience and ethical learning in human society. AI does not internalize these cultural and moral contexts in the same way."

There were also differences in each model's decision-making approach. Claude tended to act more like a calculated strategist. GPT generally showed a cautious and mediatory attitude, but revealed a different side when faced with time constraints. Gemini displayed a more direct and hardline approach.

The research team emphasized that these results do not mean AI will actually control nuclear weapons. Professor Payne commented, "No one is going to entrust ChatGPT with nuclear launch codes, but that does not render these experiments meaningless."

He added, "AI systems are already being used in the military sector for logistics, information analysis, and decision support. Going forward, it is highly likely that AI will become more deeply involved in strategic judgments made under significant time pressure. Understanding how AI reasons about strategic issues is no longer just an academic discussion."

Experts point out that this research should not simply be interpreted as evidence of AI's 'aggressiveness.' Instead, it is important to closely examine how goal-setting, reward structures, and simulation design influenced the models' decision-making.

This content was produced with the assistance of AI translation services.