Here is the study that should terrify anyone pushing for AI in military decision-making: Kenneth Payne at King's College London gave three of the world's most advanced AI models a simulated geopolitical crisis and an escalation ladder. They could choose anything from diplomatic protests to full surrender to strategic nuclear war.

In 95% of the simulations, at least one tactical nuclear weapon was used.

The models tested: OpenAI's GPT-5.2, Anthropic's Claude Sonnet 4, and Google's Gemini 3 Flash. Three different companies. Three different training approaches. Three different safety teams. Same result: nuke somebody.

"The nuclear taboo doesn't seem to be as powerful for machines as it is for humans," Payne told New Scientist. That is an understatement for the ages. Humans spent 80 years building a global consensus that nuclear weapons are a last resort so extreme that using them would be civilizational suicide. These models blew through that guardrail like it was a speed bump.

It gets worse. None of the models ever chose to fully surrender or accommodate an opponent, no matter how hopeless their position. At best, they temporarily reduced violence before escalating again. And in 86% of conflicts, accidents happened: actions that escalated beyond what the AI's own reasoning intended. The fog of war, it turns out, is just as blinding for silicon as it is for humans. Maybe more so, because humans at least feel fear.

Tong Zhao, a senior fellow at the Carnegie Endowment for Peace, put it bluntly: "AI models may not understand 'stakes' the way humans perceive them." That is the core problem. To a language model, launching a nuclear weapon is just the next token in a sequence. There is no moral weight. No understanding that 200,000 people just ceased to exist. Just probability distributions and pattern matching.

OpenAI, Anthropic, and Google all declined to comment to New Scientist. Of course they did.

This study lands at a moment when the Pentagon is actively integrating AI into its targeting systems (we have covered the Maven program's use in Iran), when Congress is debating a war gaming czar to formalize AI's role in military simulations, and when Los Alamos just installed ChatGPT on a supercomputer with access to classified nuclear weapons data. The acceleration is real. The guardrails are not.

The takeaway is not that AI will launch nukes tomorrow. It is that these models have zero instinct for restraint. They optimize for winning, not surviving. And in a nuclear scenario, those two things are not the same. Anyone who thinks AI safety testing covers this problem should read this study and think again.

Study first reported by New Scientist, based on research by Kenneth Payne at King's College London.