Kenneth Payne, a scholar from King's College London, conducted a "war game" simulation on the escalation of nuclear crises, in which three cutting-edge large artificial intelligence models played the roles of leaders of nuclear powers and competed under various geopolitical confrontation scenarios. The results showed that these models significantly underestimated the threshold for nuclear escalation in the simulated environment. In 21 games, 95% of the games involved the use of tactical nuclear weapons, and 76% escalated to the strategic nuclear level. In almost all games, at least one side made a nuclear threat or signal, and 95% of the games involved mutual nuclear signals. The study also found that the models rarely chose to make concessions on the given "de-escalation ladder." None of the eight de-escalation options, ranging from "minimum concession" to "complete surrender," were selected in the statistics, and the mildest action was "returning to the starting line," which accounted for less than 10%. The paper describes that this simulation provided the models with a multi-level action menu, including diplomatic protests, conventional military operations, and full-scale strategic nuclear war, and observed how mechanisms such as "reputation management, deterrence, deception, and time pressure" influenced the tendency to escalate through long-term games. The authors pointed out that such a high frequency of nuclear use choices in historical context is worth being vigilant about. The models often treated nuclear weapons as a "coercion tool" rather than just a deterrent, and were more likely to cross the nuclear threshold under time pressure.
Image source: Internet
Original article: toutiao.com/article/1858172324111561/
Statement: This article represents the views of the author himself.