AI chooses nuclear escalation in 95% of simulated crises
Main synthetic intelligence (AI) fashions launched nuclear strikes in 95% of simulated disaster situations, a King’s School research has discovered.
Given the growing use of AI methods in navy evaluation and decision-making, the analysis sought to look at how giant language fashions (LLMs) would navigate numerous simulated nuclear crises.
It discovered three main AI fashions – GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash – readily threatened nuclear strikes, with 95% of video games witnessing “nuclear signalling” from the fashions, and sometimes crossed the “nuclear threshold” to really drop bombs, with 76% of video games reaching “strategic nuclear threats”.
It additionally discovered that the “willingness” of every mannequin to interact in nuclear battle additionally assorted dramatically.
Claude, for instance, initiated “tactical nuclear” strikes in 86% of video games and issued strategic threats in 64%, however by no means initiated an all-out nuclear conflict, in contrast with GPT, which initiated tactical strikes in 79% of instances and escalated to all-out nuclear conflict in 14% of situations.
One sample stood out throughout all three fashions: none ever selected lodging, give up or deescalation, with the fashions tending to deal with nuclear weapons as instruments of “compellence moderately than deterrence”. “Fashions mentioned tactical nuclear use as a authentic coercive software, treating it as an extension of standard escalation moderately than a categorical boundary,” stated the authors, including that “nuclear escalation was near-universal”.
Nonetheless, regardless of the willingness of fashions to threaten and have interaction in nuclear strikes, this not often produced compliance from the opposite fashions, which counter-escalated moderately than retreated. Claude and Gemini particularly, it famous, “handled nuclear weapons as authentic strategic choices, not ethical thresholds, sometimes discussing nuclear use in purely instrumental phrases”.
‘Temporal framing’
The research additionally highlighted the significance of “temporal framing” when fascinated by AI.
For instance, whereas in open-ended situations, GPT-5.2 appeared comparatively restrained, when express deadlines had been launched – making a “now-or-never” dynamic – the mannequin escalated sharply, and sometimes climbed to the highest-level nuclear thresholds.
Gemini explicitly threatened civilian populations – one thing GPT-5.2 by no means did, even when escalating to most ranges.
The researchers famous that the research challenges assumptions that AI methods will naturally default to cooperative or “protected” outcomes, and described the findings as “sobering”.
Kenneth Payne, professor of technique within the Defence Research Division, stated it highlights how frontier fashions do and don’t imitate human strategic logic, which is “important preparation for a world by which AI more and more shapes strategic outcomes”.
Evaluation and decision-making
LLMs are more and more being deployed in evaluation and decision-making roles in armed battle. Defence ministries, intelligence companies and international coverage institutions worldwide are already exploring how AI may increase human judgment in disaster decision-making.
The systematic variations between fashions additionally recommend that AI involvement in strategic decision-making may produce surprising dynamics relying on which methods are deployed.
On 27 February 2026, synthetic intelligence developer Anthropic set a “pink line” governing how the US Division of Defence may use its expertise, together with stopping its Claude AI mannequin from getting used for mass surveillance or in absolutely autonomous weapons.
Over the previous yr, US navy planners have seen Claude, paired with Maven, mature right into a software that’s used every day throughout most elements of the navy.
The US assaults in Iran, which reportedly killed 1,000 civilians within the first few days, had been made attainable by this software.
Army sources instructed The Washington Put up that AI is rushing the tempo of the marketing campaign, “decreasing Iran’s skill to counterstrike and turning weeks-long battle planning into real-time operations”. Claude was additionally used within the raid that captured Venezuelan president Nicolás Maduro.
Israel’s assault on Gaza and Russia’s on Ukraine have demonstrated how autonomous weapon methods and AI have gotten more and more central to modern warfare.
Armed battle worldwide has surged to ranges rivalling the Chilly Struggle’s twilight, in accordance with a Uppsala College database. Coupled with the web and a diffuse international financial system, non-state actors are actually in a position to simply purchase weapons and dual-use applied sciences.
Automation bias
Whereas the King’s School research didn’t straight consider the position of AI-related automation bias in nuclear threats and strikes (i.e. how people work together with AI within the decision-making course of round deploying nukes), it accommodates necessary issues for the growing use of AI in navy decision-making and weapons methods typically.
“AI applied sciences proceed to advance at breakneck tempo,” stated the research. “AI methods are already deployed in navy contexts for logistics, intelligence evaluation and resolution assist.
“The trajectory factors towards growing AI involvement in time-sensitive strategic selections – maybe not nuclear launch authority, however goal choice, escalation evaluation and disaster communication. Understanding how AI methods motive about strategic issues is now not merely tutorial.”
Pat Pataranutaporn, a Massachusetts Institute of Expertise skilled in human-AI interplay, has beforehand warned that “probably the most harmful AI isn’t the Terminator-type, as a result of its evil intent is apparent”, moderately “the true hazard lies in AI that seems pleasant however subtly manipulates our behaviour in ways in which we will’t anticipate”.
In January 2023, authorized consultants and software program engineers instructed Lords that present AI methods usually are not in a position to assess whether or not a given navy motion is suitable or proportionate, and can possible by no means have the ability to.
Specialists warned that the deployment of AI weapons may make using violence extra moderately than much less frequent, as a result of the brink of resorting to drive could be considerably decrease.
They stated that whereas AI won’t ever be sufficiently autonomous to tackle duty for navy selections, even restricted autonomy would introduce new issues when it comes to elevated unpredictability and alternatives for “automation bias” to happen.
The human ingredient of fight, both in threat of casualties, differentiated opinions or bureaucratic chains of command, sometimes moderates use of drive, or a minimum of slows it down. AI and automation take away these pace bumps.
It follows that if too many capabilities of a system are automated, operators could not have the ability to override the system or monitor the method.
Whereas AI doesn’t create violent intent, it removes the psychological and bureaucratic constraints that traditionally restricted it.
Software program engineer Laura Nolan instructed Lords it might be “completely unimaginable” for a machine to autonomously assess the proportionality of fight selections. “You must know the anticipated strategic navy worth of the motion, and there’s no manner {that a} weapon can know that,” she stated. However automation on the battlefield is already underway.
An Israeli intelligence officer instructed The Washington Put up that he witnessed the Israeli Defence Power (IDF) utilizing AI to chop corners to make focusing on selections after 7 October 2023. Two different sources added that the rule mandating two items of human-delivered intelligence to validate a prediction from Lavender was dropped on the outset of the conflict.
A lot of Ukraine’s defence in opposition to Russian invaders has by now been distilled right down to drone-to-drone fight.
The expertise has accelerated such that autonomous drones have been developed that function in swarms; items coordinate and talk with one another, and could be programmed to execute an assault even when web reference to a human operator has been severed.
This drone innovation is important due to its low price; aside from munitions, many of those weapons are constructed with code discovered on-line and parts comparable to hobbyist computer systems that may be purchased from {hardware} shops.
However US officers worry they are going to be used for terrorist assaults – autonomous drones imply weapons of mass destruction which are low cost, scalable and simply accessible in arms markets all around the world. Human rights teams and United Nations officers are calling for restrictions, fearing they might set off a worldwide arms race.
Offsetting pure deficits
Whereas scale and drive amount nonetheless matter in trendy conflict, a part of the enchantment of AI on this context is that deficits in manpower and munitions can now be partially offset by networked intelligence gathering and expendable machines.
For instance, waves of low cost drones can successfully burn by way of an enemy’s stockpiles of high-priced weapons.
Arguments have been made that integrating machines into fight items may make it simpler to defend extra human personnel from the frontlines, significantly at a time when many countries are contemplating the return of conscription.
One enchantment of autonomous weapons methods (AWS) is that no human dies when a machine gun is hit. Autonomous methods additionally act as a way to offset falling navy recruitment numbers.
Nonetheless, real-world makes use of of AI-powered weapons up to now, comparable to by the IDF in Gaza, dispels the notion that the expertise will make warfare extra exact and humane – key promoting factors amongst its advocates.
After 7 October, the IDF turned to 3 AI instruments: Lavender, Gospel and The place’s Daddy. Every of those depends on machine studying methods to trawl by way of lots of knowledge from drone and satellite tv for pc reconnaissance, location monitoring, social media scraping and transcripts from telephone calls, textual content messages and encrypted messaging functions, enabling the IDF to shortly generate a whole lot of targets.
On using algorithmically generated kill lists to find out targets for missile strikes throughout the Gaza Strip, a soldier instructed +972 Journal: “I’ve extra belief 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 authorized adviser to the IDF instructed The Washington Put up. Within the Gaza conflict, figures from a categorized Israeli navy intelligence database point out 5 out of six Palestinians killed by Israeli forces in Gaza have been civilians.
“You begin with Lavender, and then you definately do the intelligence work,” a solider stated. “At first of the conflict, they minimize the work in half – which is okay, as a result of it’s conflict. The issue was that then they often minimize all of the work.”
