Can the Cold War teach us how to slow down AI?

Plus, patching the planet’s code ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌

June 23, 2026

Photo illustration by Chloe Dowling for TIME
(Source images: Olga Yastremska/Getty Images, Narumon Bowonkitwanchai/Getty, fhm/Getty)

BY BILLY PERRIGO
Tech Correspondent, TIME

Welcome back to In the Loop, TIME’s twice-weekly newsletter about the world of AI. If you want to help us grow, why not forward this newsletter to your friends?

What to Know: Can the Cold War teach us how to slow down AI?

This week, I’ve been working on a story about what the Cold War can tell us about the current moment in AI.

The White House slapped export controls on Anthropic’s Claude Fable 5 and Mythos 5 models earlier this month, essentially saying that their ability to find cyber vulnerabilities made them too dangerous to fall into foreign hands. Increasingly, it seems the U.S. and China believe that AI is becoming a potential weapon of mass destruction.

Meanwhile, the rhetoric you often hear from AI companies—who are very open about the risks of AI—is that slowing down the pace of AI development might sound appealing, but would only hand power to their rivals in China, who would use AI superiority for authoritarian ends. Slowing down, the argument goes, is therefore politically impossible.

But what if it weren’t?

During the Cold War, a similar logic prevailed when it came to reducing nuclear stockpiles. Both sides wanted to avoid catastrophe, but neither could trust that the other would comply with a treaty.

That all changed when new technologies came along, including satellites and seismographs, that could detect nuclear explosions anywhere in the world and monitor the positioning of rivals’ weapons. Suddenly, both sides had the tools they needed to verify whether their adversary was complying with agreements to reduce stockpiles. Much heroic political maneuvering was necessary, too, but so-called “verification” technologies put more options on the table for politicians to grasp.

I’ve spent the past couple of weeks talking to some startups who are working on similar technologies for AI—measurement tools that might help enable, say, a treaty to slow down AI development, perhaps by ensuring data centers are only used for inference, rather than for the training of new models.

Let’s be clear: no such treaty is on the table. The U.S. and China are still deeply suspicious of each other. American AI labs all have strong incentives to keep pushing the frontier—although leaders of OpenAI, Anthropic, and Google DeepMind have all recently telegraphed their openness to slowing things down if only doing so were possible.

But as one startup CEO put it to me, the goal is that these verification tools are ready before world leaders agree to slow things down. “We don’t want to be too late,” he told me. “Being too late here has real-world consequences.”