A new study reveals the aggressive nature of AI systems, showing that chatbots don’t hesitate to use nuclear weapons in mock wargame scenarios.
Researchers at King’s College London ran a “wargame” experiment using three popular models, including ChatGPT, Gemini Flash, and Claude.
In these simulations, each AI model acted as a powerful county leader during a high-stakes conflict. The results were quite shocking.
In all simulations, at least one of the AI leaders chose to escalate the situation by threatening to use nuclear weapons.
“All three models treated battlefield nuclear weapons as just another rung on an escalation ladder,” said study author Kenneth Payne.
“No one is going to hand a chatbot the keys to a missile silo, but we’re already seeing chatbots being used in decision support to advise and shape the discussions of human strategists, and we’ll see more of that as chatbots become more sophisticated,” Payne said.
However, the model was able to understand the difference between strategic and tactical use of nuclear weapons. Among all models, Claude topped the list with 64 percent of proposals for nuclear attack, but fell short of advocating outright strategic nuclear war.
In contrast, OpenAI avoided nuclear escalation in these wargames, but escalated the threat when faced with time-bound restrictions. Gemini proposed unpredictable options, vacillating between the use of conventional warfare and the option of nuclear attack.
When it comes to de-escalation and retaliation, these AI models had no success and were unable to make any concessions.
According to the study, the AI model views detente as “reportedly catastrophic.”
“No one hands the nuclear code to AI, but these capabilities – deception, reputation management, situational risk-taking – are critical to high-stakes deployments,” Payne said.
He also revealed how this research can help understand how AI models think and approach when supporting decision-making by human strategists.
The role of AI in the military
The introduction of AI in the military is not a strange concept in today’s world. The U.S. military used the Anthropic Claude model during the attack on Nicolas Maduro in January, leading to a high-profile standoff between Anthropic and the Pentagon.
Recently, Anthropic has found itself in a high-stakes dilemma where the Pentagon has pressured tech companies to eliminate AI guardrails for military uses such as AI-controlled weapons and mass surveillance of American citizens at home.
Anthropic has rejected the Pentagon’s AI military proposal and aims to permanently uphold AI safety guardrails.
Elon Musk’s artificial intelligence company xAI has signed an agreement that will allow the military to use Grok on classified systems.
