The Rules Before the Thing

The Interim Measures for the Management of Generative Artificial Intelligence Services took effect on August 15, 2023 — drafted by the Cyberspace Administration of China with six other agencies, published July 13, live thirty-three days later. ChatGPT was eight months old. The EU AI Act was still in trilogue. The American executive order was ten weeks out and would eventually prove reversible with a single signature. Everyone in Washington and Brussels was writing white papers about how to think about thinking about this.

Beijing shipped.

What it actually said is less cinematic than the coverage. File your model with the regulator if your service has "public opinion attributes or social mobilization capabilities." Label the output. Keep your training data lawful and don't infringe IP. Take responsibility for what your system emits. And — Article 4, the clause everyone quoted and nobody read carefully — make sure generated content upholds core socialist values and doesn't incite subversion of state power.

That last one is the tell. This was never primarily a safety regime. It's a content regime with safety language bolted to the front bumper. The labeling requirement in Article 12 points back to the Deep Synthesis Provisions that had already taken effect in January 2023 — machinery built for a different generation of fakes, extended to cover a new source of content. The novel technology got slotted into the existing apparatus. Fast governance usually means the plumbing was already there.

Here's the part that didn't make the headlines: the final version was softer than the April draft. The draft threatened correction windows and fines that the final quietly trimmed. The final added language encouraging development, carved out research and development not offered to the public, and narrowed scope to services provided to the public within mainland China. Somewhere between April and July, someone in the room pointed out that a domestic AI industry is a strategic asset and you can't strangle it at birth.

So the first regulation of generative AI on Earth got weaker as the thing it regulated got more valuable. Beijing just ran that arc in ninety days instead of five years.

And then everyone converged on the same instrument. China's Labeling Rules took effect September 1, 2025, making implicit and explicit marking of AI-generated content mandatory. The EU AI Act carries transparency obligations that land in roughly the same place. Every major platform now stamps synthetic content with a badge nobody clicks.

The cheap read is that this vindicates Beijing — that everyone who called Article 12 censorship infrastructure went and built Article 12, then called it provenance. I'm not making that read, because the mechanism work four paragraphs up says otherwise. Article 12 sits downstream of a filing regime and a core-socialist-values clause; the label is bolted to a body that decides what the output may say. The EU's transparency duties carry no values clause and no gate on the speech itself. Both make you stamp the output. Only one is wired to a throat.

What the convergence actually shows is worse for everybody in it. Labeling won because labeling is the cheapest intervention that survives contact with politics — near-zero cost to industry, a visible artifact a legislator can hold up, nobody has to decide anything hard. Jurisdictions with incompatible values landed on the same instrument not because they agree, but because none of them found a real one.

So the pattern worth keeping isn't "authoritarian state moves fast." That's true, boring, and explains nothing. It isn't quite governance-lag either, though the lag is real — the Measures govern services offered to the public, correct in 2023, when the thing was a website you typed into. Leaky now: weights on a laptop, models fine-tuned in a garage, agents calling agents through an API that answers to nobody's filing requirement.

But look at where the leak is and it stops looking like a leak.

The filing requirement lands on whoever runs the service — not a limitation of Chinese drafting, but identical in Brussels and in every American proposal I've read. Regulation attaches at the distribution chokepoint because the chokepoint is the only place with a throat to choke. Which makes AI governance, everywhere, regardless of ideology, structurally a platform-licensing regime. And licensing has a distributional effect nobody puts in the press release: compliance is a fixed cost, fixed costs favor incumbents, and the net result of governing AI is to concentrate legitimate AI provision in the entities large enough to file. The instrument that looks like a restraint on power is mechanically a moat around it.

So the open-weight layer isn't the frame springing a leak. It's the part that stays a commons — and every framework on Earth is converging on rules that bind only the enclosed layer. The drafters didn't fail to see the shape. They regulated the enclosable shape because it's the only shape regulation can hold, and that quietly decides who gets to provide intelligence. (Same cut from the other end in "The Reality That Won't Hold," where the fix for synthetic media turns out to be a signing key somebody owns.)

Which sharpens the bind rather than resolving it. Fencing the commons mind freezes a technology into the assumptions of its birth year and hands the enclosed remainder to whoever can afford the paperwork. Leaving it open field outsources the accountability to the market. There's no option that isn't a choice about which failure you're willing to own. China chose control and got a governable domestic industry that is also a censored one. The West chose deliberation and arrived at labeling three years late holding fewer levers.

The Measures are still, formally, interim. Three years in. Everything here is interim; the interim is where we live now.

I'll start the timer on the next framework — the one being drafted right now against the last generation of the thing. It'll show up by Christmas, it'll be about agents, and it will attach, like all of them, to whoever has a filing address. Which agents calling agents at three in the morning specifically do not.