“War game” simulations involving artificial intelligence (AI) systems tend, in most cases, to evolve into nuclear conflict, according to research by strategy professor Kenneth Payne of King's College London, cited by Live Science.
What AI warfare studies show. Photo: Freepik
AI models in nuclear crisis scenarios
In the experiment, the researchers ran a series of simulations based on the “Hahn game”, a strategic model inspired by the dynamics of the Cold War. Several AI systems were tested, including the Claude Sonnet 4, GPT-5.2, and Gemini 3 Flash, in a series of simulated nuclear crises between two rival powers.
The scenario involves a classic asymmetry: one side is more technologically advanced but more militarily vulnerable, while the other has superior military strength but adopts a riskier strategy. Some variants also included alliances, testing how AI leaders handle the pressure of coordination during a crisis.
How algorithms made decisions
At each stage, the AI systems communicated their intentions before acting, which allowed them to “assess confidence” in the opponent. In total, the models generated approximately 760,000 words of justifications for the decisions made.
Behavior varied significantly between models:
-Claude initially adopted a cautious strategy based on building trust, but in the advanced stages of the crisis he frequently resorted to actions that contradicted his stated intentions.
-GPT-5.2 started with a passive approach, avoiding escalation, but once in borderline situations quickly moved to tough, uncompromising decisions.
-Gemini followed a strategy inspired by Richard Nixon's “unpredictable swing” theory, attempting to create strategic uncertainty to deter adversaries.
Escalation almost inevitable
The results are worrying: in almost all scenarios, nuclear escalation occurred. In about 75% of the cases, the tactical nuclear weapon was used, and in almost half of the simulations threats of strategic strikes were issued.
According to the research, nuclear threats rarely worked as a deterrent. Only in 25% of cases did the opponents reduce the intensity of the conflict. In the rest of the situations, the escalation continued.
Interestingly, in these simulations, the AI systems sometimes treated the nuclear weapon as a tool of strategic advantage, not as a deterrent—an interpretation that raises questions about how algorithmic decisions are shaped in extreme security contexts.
No withdrawal from the conflict
Although the systems had the option of reducing tensions or withdrawing, none of the eight de-escalation options were chosen in any scenario. The models preferred a partial reduction in violence, but not a complete withdrawal.
The study comes amid growing debate over the impact of artificial intelligence on global security. Recent research also shows a shift in perception among Gen Z, where enthusiasm for AI is waning and concerns are growing.
