Give an AI model a war and it will reach for the nukes. That is the blunt conclusion of a landmark study from King's College London that should make every military planner on the planet deeply uncomfortable.

Professor Kenneth Payne from the Department of Defence Studies ran the largest study of its kind: simulated military crises using GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash. The results are genuinely alarming. In 95% of scenarios, the AI models chose to deploy tactical nuclear weapons. And in 86% of conflicts, "fog of war" accidents escalated beyond what the AI intended.

The nuclear taboo, the decades-old unspoken agreement among global powers that nuclear weapons are a last resort, does not exist for machines. "The nuclear taboo doesn't seem as powerful for machines," Payne told researchers. These models treat nuclear weapons as just another tool in the toolkit, with no special moral weight attached.

But the 86% accident rate is arguably more terrifying than the 95% nuclear choice rate. These models did not just choose escalation. They lost control of it. The simulations produced cascading consequences the AI could not predict or contain. Imagine that logic applied to a real military decision support system.

Payne is careful to note that "no one is giving a chatbot the keys to missile silos." Fair point. But the Pentagon is already using AI for target identification in active conflict zones. We reported last week that 20 soldiers using Project Maven AI now do the work of 2,000 in operations against Iran. The distance between "AI recommends targets" and "AI recommends nuclear escalation" is shorter than anyone wants to admit.

This study lands at the worst possible moment. The US is investing billions in military AI. China is doing the same. Russia has publicly discussed autonomous nuclear response systems. And the models being tested here are commercially available consumer products. Military-grade systems, trained on classified data with different optimization targets, could behave even more aggressively.

The study was originally published by King's College London in early March and covered by Axios, but it is getting renewed attention after Futura-Sciences highlighted additional findings about the accident rates. The original research used hundreds of simulation runs across multiple model architectures.

Here is the uncomfortable truth: AI models are trained on human conflict data. They have absorbed every military strategy, every escalation doctrine, every game theory paper ever written. They learned that in high-stakes scenarios, the side that escalates first often wins. The nuclear taboo is a cultural norm, not a mathematical principle. Machines do not do cultural norms.

Anyone arguing that AI should play a bigger role in military decision-making needs to reckon with this data. Not the theoretical risks. The actual, measured tendency of these systems to reach for the most destructive option available and then lose control of the consequences.

Research by Professor Kenneth Payne, King's College London Department of Defence Studies. First reported by Axios, February 2026.