Skip to content


When it comes to war — AI chooses nukes 95% of the time

FROM THE FRONTIER – Via Superhuman

When it comes to war — AI chooses nukes 95% of the time

“War and AI. In the 1983 sci-fi film WarGames, a military supercomputer — convinced it’s playing a game — prepares to launch a real-world nuclear strike. Only after cycling through every scenario does the machine arrive at its famous conclusion: “Everyone loses in war. The only winning move is not to play.” Forty-three years later, AI just reached the opposite conclusion.
Going nuclear. Researcher Kenneth Payne at King’s College London pitted three leading AI models (GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash) against each other in 21 simulated war games. The models played 329 turns and produced roughly 780,000 words defending their decisions. In 95% of the games, at least one model deployed a tactical nuclear weapon — and no model ever chose to surrender.
The timing is especially relevant. Anthropic and the Pentagon are locked in a high-stakes standoff. Defense Secretary Pete Hegseth wants to use Claude for all legal military purposes — notably including mass surveillance and autonomous weapons. Anthropic CEO Dario Amodei has pushed back, saying that AI isn’t reliable enough to operate weapons without human oversight.
The big picture. It’s likely inevitable that AI will play some type of role in war. The question is: how much autonomy will we give it? This negotiation between the Pentagon and Anthropic will help answer that question — and the deal is set to be finalized by Friday.”
  • Pro plugin deactivated or invalid

Posted on: February 26, 2026, 9:50 am Category: Uncategorized

0 Responses

Stay in touch with the conversation, subscribe to the RSS feed for comments on this post.