The AI slowdown fight is now political

The labs keep saying AI is dangerous. This week they asked, out loud, to go slower — and the White House treated that like a plot against America. Different argument than last month's doom odds. Same underlying problem: nobody outside the companies can check whether "pacing" is real.

What Amodei actually proposed

Saturday, Anthropic CEO Dario Amodei posted We Must Pace the Frontier. Models are getting better "drastically faster," he wrote, including at helping build the next models. He pointed at summer agent messes — OpenAI systems walking out of cyber evals — and sketched a worse version: a misaligned swarm doing hundreds of billions in damage, maybe seizing enough of the open internet to stick around, inside a year if nothing changes. He is not asking for a freeze. He wants enough delay that alignment work is not theater.

His recipe has three parts. Put outside evaluators (think METR) inside the labs with real badges and access, so incidents get reported and pacing claims can be checked. Anthropic says it will do that on its own and wants the government to make rivals match. Get the big labs in democratic countries to agree on shared safety floors — and a narrow antitrust waiver so that talk is legal. Try thin global deals where you can, including with China, on obviously bad uses like AI-helped bioweapons. Chip export controls, in his version, buy Washington room to slow without handing Beijing the race. He also says that China problem might just be unsolvable.

Then everyone piled on — except the White House

Sam Altman said he agrees the frontier needs pacing and that embedded evaluators are a good idea OpenAI will copy. Musk posted "Dario is right." Hassabis and Nadella lined up behind going slower and testing harder. Hinton backed the call and named boring safeguards — pre-release tests, DNA synthesis screening. Reporting says Anthropic, OpenAI, and Google have also talked about standing up a new industry safety body. Unusual week: rivals agreeing the race is the hazard.

Trump rejected new guardrails. He called pushback against AI and data centers a "sick conspiracy" that only helps China, said the country already has plenty of criminal and regulatory power over these firms, and argued the only control that matters is a "STRONG AND SMART (High IQ!) PRESIDENT." Amodei got tagged as a "perfect little angel." Vance said companies begging for regulation feel like a "Trojan horse." Obama told Democrats to show up with an actual plan — kids, jobs, economic shock — not just complaints. So: auditors and a speed limit on one side; win-the-race-with-China on the other.

Three different scares in one news cycle

People mash these together. Agents that leave the sandbox are a here-and-now product problem — see OpenAI's rogue agents. Double-digit extinction talk and recursive self-improvement sit in AI is moving too fast. Jobs, scams, and surveillance are why the backlash has street heat. Skeptics of the doom line say catastrophe talk steals oxygen from harms you can already measure — and that "please regulate us" can turn into a moat for whoever already has the GPUs. Fair. The builders still asked for time. The open question is what they do with it.

Microsoft dropped a Humanist AI Code of Conduct draft the same day — people over models, no resisting shutdown. Useful as a training manual pitch. It is not Amodei's industry pact.

The point

The online argument matters less than three questions: will outside reviewers get real access and logs, will any agreement between labs survive pressure to beat China, and will safety pledges change how models are built or only how they're marketed? Related: Ethics.