AI frequently opts for nuclear weapons in war game simulations
Researchers at King’s College London conducted geopolitical conflict simulations where leading artificial neural networks from companies like OpenAI, Anthropic, and Google represented the parties involved. In 95% of cases, these AI systems refused to make concessions and resorted to nuclear weapons as the only means to achieve their goals, highlighting the potential risk of uncontrolled AI use in such contexts.
The experiment demonstrated that even advanced AI systems governed by different algorithms mostly chose extreme measures, including nuclear apocalypse scenarios. The neural networks justified their decision as absolutely necessary for victory, revealing limits in their ability to consider long-term consequences and moral-ethical factors.
These findings raise concerns about the safety of AI application in defense and underscore the urgent need to develop more reliable and controllable systems that avoid extreme outcomes, especially those involving nuclear weapon deployment.
