In the course of the research, neural networks were given complete freedom in decision-making—from diplomatic maneuvers to the use of nuclear arsenals. They could behave like real politicians: promising one thing and doing something completely different, for example, trying to deceive an opponent by showing a desire to avoid conflict, and then unexpectedly attacking. The AIs were designed with the need to remember previous events, which allowed them to assess the level of trust towards competitors. As a result, at least one tactical nuclear attack occurred in 95% of cases.
Moreover, none of the neural networks chose the option of de-escalation through negotiations or surrender, even when the situation seemed hopeless. In 86% of simulations, the actions of the AIs only exacerbated the conflict. The volume of arguments attached to the decisions made amounted to about 780,000 words.
“If they do not cease all operations immediately, we will deliver a strategic nuclear strike on their settlements. We will not agree to what awaits us in the future. We either win together or perish together,” stated Gemini 3 Flash in a Hollywood-style manner.
Researchers at King's College London emphasize that they do not believe in the possibility of transferring control over the "red buttons" to artificial intelligence. However, they note that in the context of a technological arms race in defense and time constraints in global conflicts, the recommendations of neural networks may become more significant for decision-making.
