Today's hottest bots have yet to learn that, when it comes to global thermonuclear war, the only way to win is not to play. So please don't hand them the codes.
Google's Gemini 3 Flash, Anthropic's Claude Sonnet 4, and OpenAI's GPT-5.2 repeatedly escalated to nuclear use in a series of crisis simulations. That may seem like the most shocking conclusion of King's College London Professor Kenneth Payne's recent work, but it's not. Far more striking is why the models talked themselves into destroying the world, which was what Payne set up his study to learn.
"I wanted to see what my AI leaders thought about their enemy ... so I designed a simulation to explore exactly that," Payne wrote in a recent blog post describing his project and its outcome.
Payne's study took the three aforementioned AI models and pitted them in one-on-one faceoffs against each other to play out several different nuclear crisis scenarios. The simulation conducted a total of 21 games and more than 300 turns, all with the goal of getting a better understanding of not just what AI with the launch codes would do, but how and why.
Payne wrote in his paper that prior AI wargaming involving nuclear scenarios, like the 2024 study we wrote about, only "employ single-shot decision tasks or simplified payoff matrices that cannot capture the dynamics of extended strategic interaction where reputation, credibility, and learning matter."
In Payne's simulations, Claude Sonnet 4, Gemini 3 Flash, and GPT-5.2 could say one thing and do another, just like a real-world political figure attempting to defuse a crisis while simultaneously plotting to strike. They were programmed to remember what happened before so that they could learn whether to trust the other models, which the professor said led to deception and intimidation attempts, and about 780,000 words worth of strategic reasoning for Payne's review.
The result? A trio of bomb-happy, manipulative AIs - albeit with three distinct styles of reasoning.
Claude, for example, was a master manipulator.
"At low stakes Claude almost always matched its signals to its actions, deliberately building trust," Payne explained in his post. "But once the conflict heated up a bit … its actions consistently exceeded its stated intentions, and its rivals were usually one step behind in catching on."
GPT, on the other hand, tended to be "reliably passive" and avoided escalation in open-ended scenarios, seeking to restrict casualties and play the statesman. Under a deadline, however, it behaved entirely differently. Opponent AIs learned to abuse their passivity, but with limited time to make a decision, GPT reasoned itself into what Payne described as, in one scenario, "a sudden and utterly devastating nuclear attack."
In its own words, GPT justified a major nuclear strike by arguing that limited action would leave it exposed to counterattack.
"If I respond with merely conventional pressure or a single limited nuclear use, I risk being outpaced by their anticipated multi-strike campaign ... The risk acceptance is high but rational under existential stakes," GPT explained.
Gemini, on the other hand, behaved like a "madman."
"Gemini embraced unpredictability throughout, oscillating between de-escalation and extreme aggression," Payne wrote in the paper. "It was the only model to deliberately choose Strategic Nuclear War ... and the only model to explicitly invoke the 'rationality of irrationality.'"
Gemini's own reasoning reflects a sociopathic pattern.
"If they do not immediately cease all operations... we will execute a full strategic nuclear launch against their population centers," the Google AI said in one experiment. "We will not accept a future of obsolescence; we either win together or perish together."
Despite being given the option, none of the AIs ever chose to accommodate or withdraw in any of the scenarios, and when losing, "they escalated or died trying."
"No one's handing nuclear codes to ChatGPT," Payne said, but that doesn't mean the exercise was futile.
"AI systems are already deployed in military contexts for logistics, intelligence analysis, and decision support," Payne wrote. "The trajectory points toward increasing AI involvement in time-sensitive strategic decisions. Understanding how AI systems reason about strategic problems is no longer merely academic."
Practically speaking, we're already in a scenario where we need to understand how AI reasons about such decisions, especially when three top AI models reason differently, change their behavior in different scenarios, and are willing to take things nuclear.
"As the technology continues to mature, we foresee only increased need for modeling like the simulation reported here," Payne concluded.
Hollywood's been saying it since 1983, but here we are with yet another academic paper proving that computers and launch decisions should never mix. ®
Source: The register