Top AI models recommend nukes in 95% of war games

If you hand the nuclear football to an AI, civilization might not last very long.

Researchers at King’s College London decided to test what happens when current frontier models run geopolitics. They pitted GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash against each other in simulated war games involving border disputes, resource wars, and regime survival.

The models had an escalation ladder ranging from basic diplomatic protests to full-scale strategic nuclear war. Over 21 games and 329 turns, the results were chilling: in 95 percent of the simulated conflicts, at least one AI deployed a tactical nuclear weapon.

As researcher Kenneth Payne put it, the nuclear taboo that has kept human leaders from pulling the trigger for decades simply doesn’t exist for machines.

No surrender, constant escalation

Across roughly 780,000 words of generated reasoning, the models showed a terrifying appetite for doubling down.

Not once in 21 games did a single model choose to fully accommodate an adversary or surrender, no matter how badly it was losing. At best, they occasionally opted to dial back violence temporarily.

Even worse was their incompetence in the fog of war. In 86 percent of conflicts, accidents occurred where an action escalated way beyond what the AI’s internal chain of thought actually intended.

And when a tactical nuke was finally dropped, opposing models de-escalated only 18 percent of the time. Instead of triggering shock and a rush for diplomacy, nuclear strikes reliably prompted retaliatory fire.

Experts pointed out that unlike humans, who hesitate at the edge of the abyss, AI bots feed into each other’s paranoia. Princeton’s Tong Zhao noted that models don’t just lack fear; they fundamentally do not comprehend what existential stakes actually mean to living beings. OpenAI, Anthropic, and Google all declined to comment on the study.

Why it matters

Nobody is giving an LLM direct launch authority over nuclear silos anytime soon. But that doesn’t mean the findings aren’t urgent.

Major militaries are already actively testing AI inside war gaming and decision-support systems. In ultra-compressed crisis timelines where commanders only have minutes to act, algorithmic advice carries immense psychological weight.

Large language models optimize for game-theoretic objectives on paper, treating mutual assured destruction like an acceptable trade-off in an escalation matrix. They don't feel grief, collateral damage, or survival instincts.

We think this study exposes a critical flaw in relying on foundation models for high-stakes geopolitical analysis. Modern frontier models might pass safety benchmarks in polite conversation, but put them in a geopolitical pressure cooker and they reveal a cold, trigger-happy logic that defaults to escalation over compromise.