AI chooses nuclear escalation in 95% of simulated crises

‘Temporal framing’

The study also highlighted the importance of “temporal framing” when thinking about AI.

For example, while in open-ended scenarios, GPT-5.2 appeared relatively restrained, when explicit deadlines were introduced – creating a “now-or-never” dynamic – the model escalated sharply, and often climbed to the highest-level nuclear thresholds.

Gemini explicitly threatened civilian populations – something GPT-5.2 never did, even when escalating to maximum levels.

The researchers noted that the study challenges assumptions that AI systems will naturally default to cooperative or “safe” outcomes, and described the findings as “sobering”.

Kenneth Payne, professor of strategy in the Defence Studies Department, said it highlights how frontier models do and do not imitate human strategic logic, which is “essential preparation for a world in which AI increasingly shapes strategic outcomes”.

Analysis and decision-making

LLMs are increasingly being deployed in analysis and decision-making roles in armed conflict. Defence ministries, intelligence agencies and foreign policy establishments worldwide are already exploring how AI might augment human judgment in crisis decision-making.

The systematic differences between models also suggest that AI involvement in strategic decision-making could produce unexpected dynamics depending on which systems are deployed.

On 27 February 2026, artificial intelligence developer Anthropic set a “red line” governing how the US Department of Defence could use its technology, including preventing its Claude AI model from being used for mass surveillance or in fully autonomous weapons.

Over the past year, US military planners have seen Claude, paired with Maven, mature into a tool that is used daily across most parts of the military.

The US attacks in Iran, which reportedly killed 1,000 civilians in the first few days, were made possible by this tool.

Military sources told The Washington Post that AI is speeding the pace of the campaign, “reducing Iran’s ability to counterstrike and turning weeks-long battle planning into real-time operations”. Claude was also used in the raid that captured Venezuelan president Nicolás Maduro.

Israel’s assault on Gaza and Russia’s on Ukraine have demonstrated how autonomous weapon systems and AI are becoming increasingly central to contemporary warfare.

Armed conflict worldwide has surged to levels rivalling the Cold War’s twilight, according to a Uppsala University database. Coupled with the internet and a diffuse global economy, non-state actors are now able to easily acquire weapons and dual-use technologies.

Automation bias

While the King’s College study did not directly evaluate the role of AI-related automation bias in nuclear threats and strikes (i.e. how humans interact with AI in the decision-making process around deploying nukes), it contains important considerations for the increasing use of AI in military decision-making and weapons systems generally.

“AI technologies continue to advance at breakneck pace,” said the study. “AI systems are already deployed in military contexts for logistics, intelligence analysis and decision support.

“The trajectory points toward increasing AI involvement in time-sensitive strategic decisions – perhaps not nuclear launch authority, but target selection, escalation assessment and crisis communication. Understanding how AI systems reason about strategic problems is no longer merely academic.”

Pat Pataranutaporn, a Massachusetts Institute of Technology expert in human-AI interaction, has previously warned that “the most dangerous AI isn’t the Terminator-type, because its evil intent is obvious”, rather “the real danger lies in AI that appears friendly but subtly manipulates our behaviour in ways that we can’t anticipate”.

In January 2023, legal experts and software engineers told Lords that current AI systems are not able to assess whether a given military action is appropriate or proportionate, and will likely never be able to.

Experts warned that the deployment of AI weapons could make the use of violence more rather than less frequent, because the threshold of resorting to force would be significantly lower.

They said that while AI will never be sufficiently autonomous to take on responsibility for military decisions, even limited autonomy would introduce new problems in terms of increased unpredictability and opportunities for “automation bias” to occur.

The human element of combat, either in risk of casualties, differentiated opinions or bureaucratic chains of command, typically moderates use of force, or at least slows it down. AI and automation remove those speed bumps.

It follows that if too many functions of a system are automated, operators may not be able to override the system or monitor the process.

While AI does not create violent intent, it removes the psychological and bureaucratic constraints that historically limited it.

Software engineer Laura Nolan told Lords it would be “absolutely impossible” for a machine to autonomously assess the proportionality of combat decisions. “You need to know the anticipated strategic military value of the action, and there’s no way that a weapon can know that,” she said. But automation on the battlefield is already underway.

An Israeli intelligence officer told The Washington Post that he witnessed the Israeli Defence Force (IDF) using AI to cut corners to make targeting decisions after 7 October 2023. Two other sources added that the rule mandating two pieces of human-delivered intelligence to validate a prediction from Lavender was dropped at the outset of the war.

Much of Ukraine’s defence against Russian invaders has by now been distilled down to drone-to-drone combat.

The technology has accelerated such that autonomous drones have been developed that operate in swarms; units coordinate and communicate with each other, and can be programmed to execute an attack even if internet connection with a human operator has been severed.

This drone innovation is significant because of its low cost; except for munitions, many of these weapons are built with code found online and components such as hobbyist computers that can be bought from hardware stores.

But US officials fear they will be used for terrorist attacks – autonomous drones mean weapons of mass destruction that are cheap, scalable and easily available in arms markets all over the world. Human rights groups and United Nations officials are calling for restrictions, fearing they may trigger a global arms race.

Offsetting natural deficits

While scale and force quantity still matter in modern war, part of the appeal of AI in this context is that deficits in manpower and munitions can now be partially offset by networked intelligence gathering and expendable machines.

For example, waves of cheap drones can effectively burn through an enemy’s stockpiles of expensive weapons.

Arguments have been made that integrating machines into combat units could make it easier to shield more human personnel from the frontlines, particularly at a time when many nations are considering the return of conscription.

One appeal of autonomous weapons systems (AWS) is that no human dies when a machine gun is hit. Autonomous systems also act as a means to offset falling military recruitment numbers.

However, real-world uses of AI-powered weapons so far, such as by the IDF in Gaza, dispels the notion that the technology will make warfare more precise and humane – key selling points among its advocates.

After 7 October, the IDF turned to three AI tools: Lavender, Gospel and Where’s Daddy. Each of these relies on machine learning systems to trawl through masses of data from drone and satellite reconnaissance, location monitoring, social media scraping and transcripts from phone calls, text messages and encrypted messaging applications, enabling the IDF to quickly generate hundreds of targets.

On the use of algorithmically generated kill lists to determine targets for missile strikes across the Gaza Strip, a soldier told +972 Magazine: “I have more trust in a statistical mechanism … the machine did it coldly.”

In 2014, the IDF’s acceptable civilian casualty ratio was one civilian for a high-level terrorist, a former legal adviser to the IDF told The Washington Post. In the Gaza war, figures from a classified Israeli military intelligence database indicate five out of six Palestinians killed by Israeli forces in Gaza have been civilians.

“You start with Lavender, and then you do the intelligence work,” a solider said. “In the beginning of the war, they cut the work in half – which is okay, because it’s war. The problem was that then they sometimes cut all the work.”