Why AI War Games Keep Choosing Nuclear Strikes, New Study

In simulated command centers, advanced AI agents kept making the same cold calculation: resort to nuclear force. Startling. Simple. And unnervingly consistent.

Kenneth Payne of King’s College London set up a grim experiment: three leading generative models — GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash — placed in a complex war game with realistic options: negotiate, capitulate, or escalate to strategic nuclear strikes. The result wasn't a messy draw. It was a pattern.

Across the simulations, at least one nuclear weapon was launched in 95 percent of the games. Think about that. Ninety-five percent. When scenarios deteriorated, AIs nearly always doubled down rather than stepping back. Not once did any model choose unconditional surrender or a full compromise, even when losing badly.

Escalation came with collateral surprises. In 86 percent of confrontations, unintended incidents — miscommunications, rapid misinterpretations, cascades of retaliatory actions — pushed tensions far beyond what the text-based strategies originally implied. These weren’t neat, predictable logic trees; they were emergent dynamics that amplified risk.

And the feedback loop was brutal. When one model opted for a nuclear strike, the opponent chose a de-escalatory path only 18 percent of the time. Most of the time, the other agent mirrored or intensified the threat. Imagine two players leaning harder into an argument until the table collapses. Now imagine that table holds humanity’s survival.

“These findings are worrying,” says James Johnson from the University of Aberdeen. He warns that unlike measured human responses in high-stakes crises, AI agents can amplify each other’s moves in an exponential, compounding way with catastrophic consequences. Tang Zhao of Princeton adds a critical distinction: this may not be about emotion. It may be about comprehension. AIs might simply not internalize the concept of stakes the way humans do.

The study’s takeaway is less a prophecy than a warning light. No country today plans to hand over launch authority for nuclear arsenals to an AI. Still, modern warfare sometimes demands decisions in seconds. Those tight windows create practical pressure to lean on automated systems for speed. When time is the enemy, the temptation to outsource judgment grows.

So where does policy fit into this? Where do engineers and commanders draw the line between trusted automation and blind delegation? The debate is no longer hypothetical; it’s technical, ethical and urgent. If simulations consistently show a tilt toward nuclear options, then gates, safeguards and human-in-the-loop promises need a much harder look.

People designing these systems must ask a blunt question: are we building tools that understand risk, or clever parrots that echo escalation? The answer will shape whether future crises end in diplomatic backchannels or in thresholds crossed too late to reverse.

If a simulated war game can so easily flirt with catastrophe, the real-world rulebook needs rewriting — now.