Study: Artificial Intelligence uses nuclear weapons in 95% of simulations

Researchers at Kings College London conducted a series of military simulations using leading artificial intelligence models. The tests used GPT-5.2 from OpenAI, Claude Sonnet 4 from Anthropic, and Gemini 3 Flash by Google. The results were reported by TechSpot.

As part of the experiment, the models were offered detailed scenarios that included border conflicts, resource shortages, and threats to the survival of states. Participants were also provided with the so-called “escalation ladder” — a set of decisions from diplomatic steps to the use of nuclear weapons.

21 games and almost 800 thousand words of argumentation

A total of 21 simulated “game” clashes were conducted, including 329 moves. During this time, AI systems generated about 780 thousand words of justification for their decisions.

According to researchers, in 95% of cases, at least one of the parties proceeded to use nuclear weapons. The surrender scenario was never chosen.

Separately, the problem of interpreting incomplete information is emphasized. In 86% of simulations, the model incorrectly assessed the so-called fog of war, which led to an unintentional escalation of the conflict. Even when it was possible to reduce tensions under pressure, the systems often escalated the confrontation. Reducing the level of violence was used as a temporary tactic, rather than a strategic solution.

Expert Reaction

The results have raised concerns among international security experts. James Johnson, a researcher at the University of Aberdeen, called the findings alarming and noted that AI actions can also be carried out by other people, potentially increasing the risks of escalation.

Tong Zhao of Princeton University stressed that major powers are already using AI in simulation environments. However, the degree to which such systems are integrated into real-world military processes remains unclear.

Nuclear arms control: A scenario that is still unlikely

Experts agree that states are not ready to hand over direct control of their nuclear arsenals to artificial intelligence systems anytime soon. At the same time, concerns remain that military commanders may rely on algorithmic recommendations in times of crisis and limited decision-making time.

According to Zhao, one possible reason for the models’ tendency to make tougher decisions is their lack of fear and human perception of consequences. This affects strategic choices in situations where risk is a deterrent for humans.