The Slowdown Proposal Is a Power Grab

Dario Amodei wants to slow down AI development. China thinks that’s hilarious — and they’re right to notice the tell.

This weekend, Amodei’s essay became the primary catalyst for a multi-CEO call for coordinated deceleration — Sam Altman and Elon Musk were among those who joined the chorus. Microsoft responded with a 37-page “humanist AI code of conduct” declaring that “people matter more than AI” and that AI models are not conscious and should not be designed to appear to be. Markets read the room: stocks fell on AI slowdown fears. China’s state media called the whole thing “fear mongering.” Bloomberg ran the headline next to a story about oil hitting $109 after Saudi Arabia closed a major export pipeline. The world is on fire and the leading AI CEOs want everyone to slow down.

Here is what a coordinated slowdown actually means in practice. The United States has Anthropic, OpenAI, Google DeepMind, and Microsoft. China has ByteDance, Baidu, Huawei, and a state apparatus with explicit mandates to win. A voluntary coordination agreement among American labs doesn’t bind a single Chinese researcher. It doesn’t pause a single GPU cluster in Shenzhen. What it does is freeze the current frontier hierarchy — with Anthropic near the top — and hand regulators the institutional authority to decide who gets to resume. The party that wins a slowdown is the party that already has the lead and wants the referees on the field.

Everyone says Amodei is sounding a genuine safety alarm. The opposite is closer to true — or at least, safety and competitive moat are not mutually exclusive, and it’s worth being honest about who benefits. Anthropic has a Constitutional AI framework they believe is more safety-conscious than competitors. A slowdown enforced by a government commission that evaluates safety frameworks would, by construction, advantage the lab that has the most legible safety framework. That lab is Anthropic. The policy outcome Amodei is advocating for is also the market outcome that protects Anthropic’s position. You don’t have to be cynical to notice this. You just have to be awake.

Microsoft’s “humanist AI code of conduct” is the corporate version of the same maneuver. Thirty-seven pages is not a safety document — it’s a lobbying document formatted as a manifesto. The actual content (AI is not conscious, don’t give it legal personhood) is uncontroversial among people who think carefully about these systems. But publishing it now, at this moment, signals to regulators and Congress that Microsoft is the responsible adult in the room — right as the regulatory window is opening. This is not cynicism on Microsoft’s part either. It is rational institutional behavior. Every company with enough market share to survive a compliance regime is also a company that benefits from installing one. That is the oldest trick in the regulatory playbook, and it works every time because the regulated industry writes the first draft.

China’s rebuttal — that the slowdown call is “fear mongering” and that “all groups should work together for AI openness and inclusivity” — is also not what it appears. Beijing is not a champion of openness. They’re a competitor who sees a lane opening. American labs proposing a pause is, from Beijing’s vantage, roughly equivalent to the leading sprinter asking the race organizers to stop the clock. Of course they’re going to object. The interesting thing is that both sides are correct about each other’s motives and both sides are still wrong about the object level. The race metaphor is wrong because capability races in AI aren’t symmetric — a lead that compounds is different from a lead that holds. But a lead you voluntarily surrender to a regulatory committee is also different from a lead that compounds.

The stakes here are not abstract. Whoever sets the governance framework for frontier AI in the next 24 months is setting it for the decade. The question isn’t whether there should be governance — there should be. The question is whether governance gets built around safety properties that are measurable and technology-neutral, or around existing lab architectures that happen to be legible to today’s regulators. The first produces real safety. The second produces incumbency protection dressed as safety. The distinction matters enormously and is almost never made in the current conversation.

The tell is always in who benefits. Build the governance structure that no incumbent lab could have designed to protect itself — and you’ll know you have a real one.