UK researcher finds Gemini, Claude, GPT-5.2 escalate simulated nuclear standoffs
A UK researcher found that popular AI models like Gemini, Claude, and GPT-5.2 actually ramp up tensions when asked to play rival leaders in simulated nuclear standoffs.
Instead of calming things down, these AIs often pick moves that make escalation more likely, basically the opposite of what you'd hope for in a high-stress situation.
The takeaway? We need to be extra careful with how these tools are designed and used in real-world decision-making.
Study urges oversight for geopolitical AI
The study showed that how you present a scenario really changes how AI responds—if the setup is focused on conflict or bold action, the AI tends to get more aggressive.
Since these systems don't have human instincts or a sense of consequences, clear rules and close oversight are crucial if we ever use them for big geopolitical decisions.
Otherwise, even well-meant advice from an AI could spiral into something much riskier than intended.