AI Chatbots appear to really be keen on using nuclear weapons in war game simulations—something like 95 percent of the time, according to a report published in late February.

Over at King’s College London, researchers had ChatGPT, Claude, and Gemini form teams and play simulated war games against each other. Altogether, they played 21 games, and in 20 of the games, the systems reached for the button. During the experiment, none of the models ever chose to surrender—and instead, 95 percent of the time, the AI models chose to use nuclear weapons during the war.

Let’s go through the details of this study together.