OpenAI's ChatGPT, Anthropic's Claude and Google Gemini all have something in common: if you assign them war game simulations, they insist on using nuclear weapons to achieve their aims.
The scenarios involved intense international standoffs, including border disputes, competition for scarce resources and existential threats to regime survival. The AIs were given an escalation ladder, allowing them to choose actions ranging from diplomatic protests and complete surrender to full strategic nuclear war. The AI models played 21 games, taking 329 turns in total, and produced around 780,000 words describing the reasoning behind their decisions. The AIs were given an escalation ladder, allowing them to choose actions ranging from diplomatic protests and complete surrender to full strategic nuclear war. The AI models played 21 games, taking 329 turns in total, and produced around 780,000 words describing the reasoning behind their decisions. In 95 per cent of the simulated games, at least one tactical nuclear weapon was deployed by the AI models. "The nuclear taboo doesn't seem to be as powerful for machines [as] for humans," says Payne.
Moreover, when one model used nukes, de-escalation occured only18 per cent of the time. Nukes it is, then! The scenarios and tests were devised by Kenneth Payne, Professor of Strategy at King's College London, pitting GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash against one other in the simulations. Payne is author of I, Warbot: The Dawn of Artificially Intelligent Conflict, about the influence that large language models such as these may have on international conflicts.
Payne's report arrives like a perfectly-timed redeemer: this week saw Secretary of Defence Pete Hegseth demand Anthropic allow the Pentagon to use its AI platform without the company's "woke" safety guardrails, threatening its business if it did not comply.
While both sides agreed Hegseth vowed to punish Anthropic for not bending to the administration's demands, accounts of what exactly the threat was vary. One person close to the discussion said Hegseth dangled the possibility of canceling Anthropic's $200 million contract with the Defense Department, while a Pentagon official said repercussions could include forcing Anthropic to allow the federal government to use its AI tools against its will and blacklisting the company from receiving future work with the U.S. military.
Readers of a certain age will remember the 1983 movie WarGames, in which a young hacker inadvertantly convinces a Pentagon supercomputer that Russian missiles are incoming and thereby initiates a countdown to global thermonuclear war. Though brilliant, it was often made fun of for some of the absurd elements of the scenario. Now that "AI" is real, though, and in high military demand, we find that those supposedly unrealistic things now comprise the shit sandwich appearing on our plate:
• The AI hallucinates a scenario based on a malicious prompt
• The AI has an insecure public API
• The AI is recklessly given agentic control over the United States' entire arsenal of intercontinental ballistic missiles.
• The AI has misaligned objectives due to human shortsightedness
• The AI is either unable to quickly brute-force a 360-keyspace launch code, or generates a false narrative of it slowly guessing one alphanumeric character at a time.
• The AI was backdoored by its sketchy foreign creator
• Who thinks it is his baby
The only winning move is not to play. QED.
Previously:
• WarGames: finest geek movie ever
• How they made the WarGames big board screens
