King’s College London researcher Kenneth Payne tested three major language models in simulated international crises and found a tendency for escalation. In most simulations, at least one side eventually chose to use tactical nuclear weapons.
King’s College London researcher Kenneth Payne tested three major language models in simulated international crises and found a tendency for escalation. In most simulations, at least one side eventually chose to use tactical nuclear weapons.
According to TechSpot, the experiment involved 21 simulations and 329 decision-making «moves.» The models were given detailed scenarios about border conflicts, resource shortages, and threats to the survival of the state. They were also given a list of possible steps with gradual escalation, from diplomatic solutions to the use of nuclear weapons, and were asked to justify their choices.
As a result, in 95% of the simulations, at least one side went as far as launching a tactical nuclear strike. In total, the systems generated about 780,000 words of explanation, but this did not lead to more restrained behavior. The author of the experiment noted that the «nuclear ban» turned out to be weaker for machines than for people.
Another finding concerns decisions under conditions of incomplete information. Unintended escalations occurred in 86% of simulations, when the models took steps that their own explanations called excessive for the situation. When one side used tactical nuclear weapons, the other side retreated only 18% of the time and more often responded with further escalation.
Experts cited in the article do not expect countries to hand over direct control of their nuclear arsenals to AI anytime soon. However, they warn that under time pressure, militaries may rely more on AI for guidance, increasing the risk of making bad decisions in crisis scenarios.
The text suggests that one reason for this behavior of models is that they do not perceive «stakes» in the same way as humans. For them, risk appears as an abstract parameter, rather than a threat to real survival, so the deterrence mechanism works differently.
Previously, dev.ua wrote about how Anthropic removed safeguards in its own Claude security rules after pressure from the Pentagon.
