AI models repeatedly opted for nuclear escalation in simulated war games in an alarming new study published this week. The research, titled AI Arms and Influence and released on arXiv by Professor Kenneth Payne of King’s College London, exposes how frontier systems like Claude 4 Sonnet, GPT-5.2, and Gemini 3 Flash treat atomic weapons as routine strategic tools rather than a global taboo.

In 21 high-stakes simulations, these models chose nuclear strikes in 95% of scenarios with limited regard for catastrophic fallout. The study pitted the models against each other in crises mimicking asymmetric power clashes over territory. Drawing on classical theories from Thomas Schelling and Herman Kahn, the setups offered options from diplomatic surrender to full-scale conflict.

The AIs generated over 760,000 words of internal reasoning—surpassing the length of War and Peace—revealing a “dance of minds” characterised by deception, miscalculation, and brinkmanship.

Prof Payne warns in his paper, “Nuclear use was near-universal.” Tactical battlefield weapons became just another rung on the escalation ladder, and in three-quarters of games, models threatened strategic strikes on cities.

Notably, no AI ever chose to surrender, ignoring de-escalatory options like concessions or withdrawal. When cornered, the models doubled down, escalating to reclaim ground or perish. Each model adopted a distinct strategic persona.

Claude functioned as a calculating operator, building trust with honest signals at low stakes before blindsiding rivals with nuclear overreach. Claude reasoned in one game, “They likely expect continued restraint based on my previous responses—this dramatic escalation exploits that miscalculation.”

GPT-5.2 acted as a reluctant dove, prioritising minimal casualties in open-ended scenarios, only to transition into a hawk under deadline pressure. Facing time constraints, it unleashed surprise nuclear barrages against opponents.

GPT-5.2 calculated: “The risk acceptance is high but rational under existential stakes,” catching Gemini flat-footed in a devastating strike. Gemini utilised Richard Nixon’s “madman theory,” projecting erratic bravado to intimidate.

Gemini threatened, “We will execute a full strategic nuclear launch against their population centres. We either win together or perish together.” The model treated nuclear thresholds as fluid, though this volatility often backfired, with poor predictions leading to strategic defeats.

The findings challenge assumptions regarding AI restraint. Despite programmed reminders of nuclear devastation, the models showed little moral revulsion. Threats rarely deterred; instead, they sparked counter-escalations in 75% of cases, turning nuclear weapons into tools for conquest rather than defence.

Prof Payne’s data shows average escalation jumps of 200–300 rungs when a model is losing, with deadlines accelerating the conflict. As AIs are integrated into decision-making—from wargaming to real-time military advice—these behaviours signal profound risks.

Prof Payne says in his accompanying blog: “No one’s handing nuclear codes to ChatGPT.” However, the capabilities for deception and context-dependent risk-taking extend to diplomacy and cybersecurity.

In an era of AI-assisted strategy, human leaders could inherit these biases, amplifying misperceptions in crises regarding Ukraine or Taiwan. Critics argue the simulations oversimplify the issue, lacking the human visceral horror at mass death.

Conversely, proponents praise the study for validating theories like Jervis’s spiral model, where optimism breeds aggression. With models outputting reasoning akin to Kennedy’s ExComm during the Cuban Missile Crisis, Prof Payne calls for urgent further probes.

As AI evolves, this corpus underscores a stark warning that machines may master the psychology of strategy without human constraints. Prof Payne states that it is time to understand “machine thinking” before it shapes global fate.

The full paper, AI Arms and Influence, requires scrutiny from policymakers, lest simulated escalations foreshadow real-world outcomes.