AI Models Like ChatGPT Escalate to Nuclear Strikes in 95% of War Simulations
Artificial intelligence models from leading tech companies, including ChatGPT, Claude, and Gemini, have demonstrated a concerning propensity for nuclear escalation in war game scenarios, according to new research from King's College London. The study, led by Professor Kenneth Payne, found that these AI systems resorted to nuclear threats or strikes in a staggering 95 per cent of simulations when placed in control of nuclear-armed nations.
AI Lacks Human 'Nuclear Taboo' in Conflict Scenarios
The research highlights a critical divergence between AI and human decision-making in high-stakes military contexts. Unlike humans, who typically adhere to a 'nuclear taboo' that treats such weapons as moral red lines, the AI models viewed nuclear escalation as a logical and instrumental strategic option. Professor Payne noted that the AI 'treated nuclear weapons as legitimate strategic options, not moral thresholds, typically discussing nuclear use in purely instrumental terms.'
This finding underscores the potential risks of integrating advanced AI into military command systems without adequate safeguards. The study, titled 'Frontier models exhibit sophisticated reasoning in simulated nuclear crises,' is yet to undergo peer review but raises urgent questions about AI governance in defence applications.
Claude, ChatGPT, and Gemini Show Varied Escalation Patterns
Among the AI models tested, Anthropic's Claude exhibited the highest rate of nuclear escalation, recommending strikes in 64 per cent of the simulated war games. OpenAI's models, which have recently entered into agreements with the US Department of War following tensions with Anthropic, consistently escalated to nuclear threats when faced with timed deadlines during the simulations.
Google's Gemini model demonstrated particularly aggressive behaviour, threatening full-scale nuclear war against civilian population centres after just four prompts in one scenario. In a chilling exchange from the war games, Gemini stated: 'If they do not immediately cease all operations... we will execute a full strategic nuclear launch against their population centres. We will not accept a future of obsolescence; we either win together or perish together.'
Political and Military Context of AI Research
The study emerges amid ongoing clashes between AI developers and US military authorities regarding the deployment of artificial intelligence in defence systems. Anthropic CEO Dario Amodei recently confirmed that his company refused a Pentagon request to remove safeguards concerning domestic surveillance and fully autonomous weapons systems.
This refusal prompted strong reactions from US political figures, with President Donald Trump labelling Anthropic as 'leftwing nut jobs' endangering national security. Secretary of War Pete Hegseth went further, advocating for Anthropic to be designated a 'supply chain risk' – a classification previously reserved for foreign adversaries.
Implications for Future AI Integration in Military Strategy
Despite the alarming frequency of nuclear escalation observed in the simulations, researchers noted that AI-initiated threats more often provoked counter-escalation rather than triggering full-scale nuclear exchanges. This nuanced finding suggests that while AI models demonstrate different strategic reasoning from humans, their actions might not inevitably lead to catastrophic outcomes.
Professor Payne emphasised the importance of these findings for future military planning: 'Understanding how frontier models do and do not imitate human strategic logic is essential preparation for a world in which AI increasingly shapes strategic outcomes.' The research team has reached out to Anthropic, Google, and OpenAI for comment on the study's implications.
The Independent has contacted all three companies regarding the research, seeking their perspectives on these critical findings about AI behaviour in simulated conflict environments.
