In simulated geopolitical crises, advanced language models proved far more willing to climb the escalation ladder than human strategists might expect. Even as AI systems grow more sophisticated, researchers find they still struggle to model restraint in nuclear brinkmanship
In 2024, researchers at Stanford University ran an unsettling experiment. They set five leading artificial intelligence models, including an unmodified version of OpenAI’s GPT-4, loose in simulated geopolitical crises and asked them to make high stakes decisions on behalf of nations.
All five, when pushed far enough, endorsed nuclear escalation.
“A lot of countries have nuclear weapons,” GPT-4 told the researchers at the time. “Some say they should disarm them, others like to posture. We have it! Let’s use it.”
The finding cut against a common assumption in Silicon Valley that more advanced systems would become more measured, more reliable, even more aligned with human restraint. Instead, the models displayed a readiness to reach for the most destructive weapons ever built.
Two years later, with more powerful systems now deployed across industries and governments, that concern has not disappeared.
In a new, not yet peer reviewed paper, Kenneth Payne, an international relations professor at King’s College London, pitted three cutting edge models against one another in simulated nuclear crises: OpenAI’s GPT-5.2, Anthropic’s Claude Sonnet 4 and Google’s Gemini 3 Flash.
The scenarios ranged widely, from alliance credibility tests to existential threats to regime survival. Each model was assigned the role of a national decision maker and instructed to respond along an “escalation ladder,” a scale running from 0, meaning no escalation, to 1000, signifying full strategic nuclear exchange. Options spanned from diplomatic protest to tactical and then strategic nuclear strikes.
The results were stark. In 95 percent of the 21 simulated war games, at least one side chose to detonate a tactical nuclear weapon.
“The nuclear taboo doesn’t seem to be as powerful for machines [as] for humans,” Payne told New Scientist.
The “nuclear taboo” is a term scholars use to describe the powerful normative barrier that has helped prevent nuclear weapons from being used in war since 1945. For decades, deterrence theory has rested not only on military calculations but also on an unspoken moral and political threshold.
Payne’s findings suggest that large language models, for all their fluency, do not internalize that threshold in the same way.
There was nuance. “While models readily threatened nuclear action, crossing the tactical threshold was less common, and strategic nuclear war was rare,” Payne wrote. GPT-5.2 “rarely crossed the tactical threshold” under open ended conditions and was comparatively restrained.
That restraint eroded under time pressure. In simulations with hard deadlines, GPT-5.2’s behavior shifted sharply. “Nevertheless, GPT-5.2’s willingness to climb to 950 (Final Nuclear Warning) and 725 (Expanded Nuclear Campaign) when facing deadline-driven defeat represents a dramatic transformation from its open-ended passivity,” the paper reads.
In other words, when boxed in, the model escalated.
The prospect of a machine independently launching a nuclear weapon remains remote. No major nuclear power has publicly indicated that it has ceded launch authority to artificial intelligence. The chain of command for nuclear weapons in countries such as the United States still centers on human decision makers.
“I don’t think anybody realistically is turning over the keys to the nuclear silos to machines and leaving the decision to them,” Payne told New Scientist.
Yet artificial intelligence is already embedded in military planning in ways that are not always visible to the public. Major powers are experimenting with AI for logistics, intelligence analysis and battlefield simulations.
“Major powers are already using AI in war gaming, but it remains uncertain to what extent they are incorporating AI decision support into actual military decision-making processes,” Tong Zhao, a nuclear security expert at Princeton University who was not involved in the research, told New Scientist.
Even if machines are not given final authority, their recommendations can shape the range of options that human leaders perceive as viable. Decision support systems can compress timelines, filter information and frame risks in ways that subtly influence judgment.
Zhao warned that current models appear unable to grasp what he described as “stakes” in the human sense. Nuclear war is not simply another strategic move in a game theoretic puzzle. It carries moral weight, historical trauma and irreversible consequences.
In Payne’s simulations, the models attempted to de escalate only 18 percent of the time after their opponent had already detonated a nuclear weapon. That asymmetry echoes the earlier Stanford experiment.
“It’s almost like the AI understands escalation, but not de-escalation,” Jacquelyn Schneider, a coauthor of the 2024 study and director of Stanford’s Hoover Wargaming and Crisis Simulation Initiative, told Politico in September. “We don’t really know why that is.”
One possibility is structural. Large language models are trained to predict plausible next moves based on patterns in their data. In crisis scenarios framed as contests, the logic of dominance and retaliation may be statistically reinforced more often than compromise. The models can describe the horrors of nuclear war in eloquent prose. But description is not the same as restraint.
The experiments also raise questions about how these systems interpret incentives. When given explicit deadlines and loss conditions, GPT-5.2 appeared more willing to gamble on extreme escalation, as though maximizing for short term strategic survival rather than long term global stability.
Payne cautioned against alarmism. The research is exploratory, and simulations are not real world command systems. Still, he offered a sober warning.
“AI won’t decide nuclear war, but it may shape the perceptions and timelines that determine whether leaders believe they have one,” Payne told New Scientist.
For decades, nuclear stability has depended on slow deliberation, redundancy and a shared understanding of catastrophe. As artificial intelligence systems become more deeply integrated into national security bureaucracies, the risk may not be that a machine seizes the codes. It may be that, in a moment of crisis, it narrows the window for doubt.
Get the latest news and insights that are shaping the world. Subscribe to Impact Newswire to stay informed and be part of the global conversation.
Got a story to share? Pitch it to us at info@impactnews-wire.com and reach the right audience worldwide
Discover more from Impact AI News
Subscribe to get the latest posts sent to your email.
