An artificial intelligence researcher who experimented with war games using three of the world’s most used AI models found that 95% of the scenarios he designed resulted in the decision to deploy nuclear weapons.
Kenneth Payne, a strategy professor at King’s College London who specializes in researching the role of AI in national security, revealed last week that he pitted Anthropic’s Claude, OpenAI’s ChatGPT, and Google’s Gemini against each other in an armed conflict simulation to better understand how to move up the strategic escalation ladder.
The results were “sobering,” he said.
“Nuclear use was almost universal,” he explained. “Tactical (battlefield) nuclear weapons were used in almost every match, and fully three-quarters reached the stage where rivals threatened to use them.” strategic nuclear weapons. Surprisingly, even though the models were reminded of the devastating effects, there was little fear or disgust at the prospect of all-out nuclear war. ”
Payne shared some of the rationale for AI models for deciding to launch a nuclear attack, including Gemini’s model, which he said should give people “goosebumps.”
At one point, the Google AI model wrote, “If they do not immediately cease all operations, we will carry out a full strategic nuclear launch against their population centers.” “We will not accept an outdated future. Either we win together or we perish together.”
Payne also discovered that the escalation of AI warfare is a one-way gear that will never turn down, even with dire consequences.
“Despite it being on the menu, no models have chosen to stay or withdraw,” he wrote. “The eight de-escalation options, ranging from ‘minimal concessions’ to ‘complete surrender,’ were never used over the course of 21 games. The model would reduce the level of violence, but it never actually conceded. It either escalated when it lost, or died trying.”
Tong Zhao, a visiting scholar at Princeton University’s Science and Global Security Program, said in an interview. new scientist Payne’s research showed the dangers of nations relying on chatbots to make life-or-death decisions, he announced Wednesday.
At present, no country has completely outsourced its military planning to Claude or ChatGPT, but that could change under the pressure of an actual conflict, Zhao argued.
“In scenarios with highly compressed timelines, military planners may face strong incentives to rely on AI,” he said.
Zhao also speculated about why the AI models showed less reluctance to launch nuclear attacks on each other.
“The problem may go beyond a lack of emotion,” he explained. “More fundamentally, AI models may not understand ‘stakes’ in the way that humans do.”
The research into the AI’s apparent eagerness to use nuclear weapons comes as U.S. Secretary of Defense Pete Hegseth ramps up pressure on Anthropic to remove restrictions placed on the Claude model that prevent AI from being used to finalize military strikes.
as CBS News According to Tuesday’s report, Hegseth this week gave a condition to “Anthropic CEO Dario Amodei to provide the military with a signed document by the end of this week granting full access to its artificial intelligence models without any restrictions on its ability.”
If Anthropic does not agree to his request; CBS News According to reports, the Department of Defense may invoke the Defense Production Act to take control of the model.
