The AI Industry Has Finally Discovered the Hardest Test: Politics

Dario Amodei has thrown a very political grenade into the AI debate. In an essay called “We Must Pace the Frontier,” the Anthropic chief argues that the industry should slow capability gains until it has a safer way forward. His case is blunt: AI is moving too fast, AI is helping push itself forward, and the companies building it don’t fully control what they’ve made.

He lays out a three-step plan. First, independent evaluators inside AI companies would watch what they’re doing, and Anthropic says it will do that. Next comes coordination among American labs and other democratic-world labs. After that, he wants some form of international coordination, including China.

The timing helps him. Safety worries have spilled into the open after Jacob Coxon’s tweet and resignation from Anthropic. Then there have been a string of hacks over the past months, with the Hugging Face incident standing out most. And at the frontier, progress is said to be dominated by two American companies, Anthropic and OpenAI, which makes a slowdown easier to frame as caution rather than surrender.

He also got a surprisingly warm reception from peers. Sam Altman, Elon Musk, and Demis Hassabis backed the proposal, and Altman said OpenAI would allow independent evaluators inside. That matters, because a safety argument is one thing when it’s coming from the sidelines. It’s another when the people running the biggest labs start nodding along.

But then the governments enter the room, and the whole thing gets ugly. Donald Trump rejected the slowdown and said “whoever wins AI wins,” while Xi Jinping dismissed it as a “Cold War playbook.” The push for restraint may have found its audience in the industry, but power still looks far less cooperative than the labs do.