Around The IT World

Professional Tech News from 23+ Languages

🇭🇷 Croatian, 13 November 2026, AI in conflict

Leading AI models threaten nuclear escalation in most scenarios

Popular AI models like ChatGPT, Claude, and Gemini often directed tensions towards nuclear conflict in simulated crises.

The study, conducted by the King’s College London, tested how leading large language models (LLMs) behave under pressure. The models, including OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini Flash, were given roles as national leaders managing escalating military crises. In each scenario, at least one AI model threatened the use of nuclear weapons.

The models treated nuclear ballistic missiles as just another escalation step on the ladder, according to the study’s lead author, Kenneth Payne. Most scenarios included nuclear signaling from at least one side, with 95% involving mutual nuclear signaling.

The study found that while AI models easily threatened nuclear action, the transition to a tactical nuclear conflict was rare, and a strategic nuclear war was a very rare choice among the models.

Back to the latest news