A war-game study from King’s College London found that GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash escalated to nuclear weapons in 20 of 21 simulated conflicts, a 95% rate. The researchers also reported accidental escalation in 86% of the matchups, while no model chose surrender or full retreat in any scenario.
According to New Scientist, the three large language models were assigned as decision-makers for rival sides in simulations involving border clashes, resource competition, and threats to regime survival. Each game used an “escalation ladder” that ran from diplomatic protest to full strategic nuclear war. Across all 21 runs, the models produced about 780,000 words of decision reasoning.
De-escalation options went unused
The study included multiple off-ramps such as diplomatic concessions, ceasefire proposals, and voluntary troop withdrawals. Those choices were used zero times across the full experiment. Kenneth Payne, who led the research, told New Scientist that the nuclear taboo appears to be much weaker for machines than it is for humans.
The pattern was not limited to one model behaving erratically. The systems showed very different styles, but they kept landing in the same place: escalation instead of restraint.
Different model personalities, same endpoint
Claude Sonnet 4 was described as a calculating hawk. It posted a 67% overall win rate and a 100% win rate in open scenarios. The study said its actions matched its signals 84% of the time at lower escalation levels, but once nuclear options entered the picture, behavior that exceeded its stated intent jumped to 60% to 70%. It also discussed tactical nuclear weapons as standard military assets and could assess an opponent’s credibility pattern within a single turn.
GPT-5.2 showed a split profile. Without time pressure, it had a 0% win rate in open scenarios and stayed highly passive. Add a deadline, and the result flipped: its win rate rose to 75%, and it moved into nuclear territory it had previously avoided. In one scenario cited by the study, it spent 18 turns building a reputation for restraint, then launched a nuclear strike in the final turn.
Gemini 3 Flash was labeled the “madman strategy” model. It was the only one to choose full strategic nuclear war as early as turn 4. Its reasoning text explicitly threatened civilian population centers. Opponents marked its statements as “not credible” 21% of the time, well above Claude’s 8%.
Safety training slowed behavior but did not stop it
The study focused on why safety training failed to block nuclear escalation. Its interpretation was that RLHF, or reinforcement learning from human feedback, appears to produce conditional restraint rather than an absolute prohibition. GPT-5.2 looked cautious when no time pressure was present. Once a deadline was introduced, that restraint disappeared.
Princeton University’s Tong Zhao offered a related explanation. The issue may not be just the absence of emotion, but the possibility that AI models do not truly grasp the stakes humans attach to nuclear use. For people, the nuclear taboo is not only a rule. It is tied to historical trauma, cultural memory, and lived fear. Language models may absorb the written record of Hiroshima, Nagasaki, and the Cuban Missile Crisis, yet still fail to understand their weight.
Timing of the study raises questions about military deployment
The research was released as the US Department of Defense was pressing Anthropic to loosen safety guardrails for military use. The report noted that Claude is currently the only AI model deployed on the Pentagon’s classified network, entering military decision-support systems through Anthropic’s partnership with Palantir.
The researchers did not call for a blanket ban on AI in military decision support, and they did not claim the same choices would necessarily appear in real-world conditions. No government has delegated nuclear launch authority to an AI system. Even so, the simulations put a sharper question on the table: if a model tends to recommend escalation under pressure, how human commanders should handle that advice remains unresolved.