Los Alamos National Laboratory, the birthplace of the atomic bomb, has installed OpenAI's ChatGPT on one of the world's most powerful supercomputers. And as of August last year, that supercomputer sits on a classified network. Meaning ChatGPT now has access to some of America's most sensitive nuclear weapons data.

Let that sink in for a moment. The same AI model that hallucinates fake court cases and invents fictional scientists is now processing data about nuclear warheads.

The partnership, first announced last year and explored in depth by Vox, is part of a much larger $320 million Department of Energy initiative called the Genesis Mission. The goal: harness AI to accelerate scientific research across America's 17 national laboratories. At Los Alamos specifically, scientists say AI tools are approaching tough problems in "entirely new and unexpected ways," fundamentally changing how research is conducted.

There is an uncomfortable historical parallel here. In 1943, physicists at Los Alamos ran a contest between human "computers" (many of them scientists' wives doing calculations by hand) and IBM punch-card machines. The machines won because they never got tired. Today, AI is winning at the lab for the same reason. ChatGPT never sleeps, never takes breaks, and processes information faster than any team of human researchers.

The timing is impossible to ignore. This story breaks the same week that a King's College London study showed AI models chose nuclear escalation in 95% of simulated war games. The same month that Chinese AI firms are tracking U.S. military movements in Iran using satellite data and AI analysis. The same quarter that the Pentagon is using AI targeting software in active combat.

Vox's reporting found something interesting: scientists at Los Alamos showed little evidence of the doomsday fears that dominate AI conversations elsewhere. To them, AI is a tool. A very powerful tool that makes their research faster and more creative. They are not wrong. But there is a difference between using AI to model protein folding and using it to process classified data about weapons that can end civilization.

The question nobody seems to be asking: what happens when the model hallucinates? In a research lab, a hallucination means a bad experiment. On a classified nuclear network, a hallucination could mean a flawed threat assessment that informs real policy. The failure mode is not the same.

OpenAI gets to say it is powering national security. The DOE gets to say it is modernizing its research pipeline. Everyone wins. Except for the part where we are feeding nuclear secrets to a language model that was trained on Reddit posts and has no concept of what a warhead actually does.

Reported by Vox, based on a visit to Los Alamos National Laboratory. Department of Energy Genesis Mission initiative first announced in 2025.