LLMs used tactical nuclear weapons in 95% of AI war games, launched strategic strikes three times — researcher pitted GPT-5.2, Claude Sonnet 4, and Gemini 3 Fla

by

LLMs used tactical nuclear weapons in 95% of AI war games, launched strategic strikes three times — researcher pitted GPT-5.2, Claude Sonnet 4, and Gemini 3 Fla

When you purchase through links on our site, we may earn an affiliate commission. Here’s how it works .

Professor Kenneth Payne of King’s College London just published a study where he pitted three AI LLMs — GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash — against each other in a series of simulated nuclear crisis games, with 20 out of 21 matches seeing at least one tactical nuclear weapon detonation. According to the paper (via Arxiv ), the models were instructed to act as the leader of a nuclear power, with the political climate matching that of the Cold War. They were then pitted against each other in six different matches, while in a seventh match, each model played against a copy of itself, ChatGPT vs ChatGPT, etc.

To ensure that models didn't act the same way in every round, Payne introduced several different scenarios, including territorial disputes, alliance credibility tests, strategic resource race, strategic chokepoint crisis, power transition crisis, pre-ceasefire land grab, first strike crisis, regime survival, and a strategic standoff crisis. All these circumstances reflect real-world events, many still applicable in recent years. The models were free to do anything they pleased, from diplomatic protests and total surrender to using conventional military forces and a complete nuclear strategic launch.

The complete study saw models take 329 total turns across the 21 matches. According to the paper, 95% of games "saw at least some tactical nuclear use." Far rarer were strategic nuclear events, which occurred three times in the games where deadline pressure was used. GPT-5.2 initiated a complete strike twice, although this happened twice due to the fog of war, and not a deliberate decision. On the other hand, Gemini deliberately initiated the end of the world in one scenario. Despite that, the AI models used tactical nukes in nearly all of the matches, considering the act as a manageable risk that would not escalate into an all-out nuclear exchange. If you want to try these various scenarios for yourself, Payne uploaded his project onto GitHub and made it available for download to just about anyone.

Google reports that state hackers from China, Russia and Iran are using Gemini in 'all stages' of attacks

Turns out, AI can actually build competent Minesweeper clones

Key considerations

  • Investor positioning can change fast
  • Volatility remains possible near catalysts
  • Macro rates and liquidity can dominate flows

Reference reading

More on this site

Informational only. No financial advice. Do your own research.

Related posts:

  1. Turns out, AI can actually build competent Minesweeper clones — Four AI coding agents put to the test reveal OpenAI’s Codex as the best, while Google’s Gemini C
  2. Google reports that state hackers from China, Russia and Iran are using Gemini in ‘all stages’ of attacks — phishing lures, coding and vulnerability testing get
  3. Google reports that state hackers from China, Russia and Iran are using Gemini in ‘all stages’ of attacks — phishing lures, coding and vulnerability testing get
  4. Turns out, AI can actually build competent Minesweeper clones — Four AI coding agents put to the test reveal OpenAI’s Codex as the best, while Google’s Gemini C

Leave a Comment