The 'Lax' One Pulled the Plug First

AI regulation splits four ways — and what splits is not the grammar but the enforcement.

On 15 July, the emotional-companion (AI conversational-partner) features vanished from the Chinese AI apps Doubao and Tongyi Qianwen. Deletion deadlines were pinned to conversation logs that millions of people had exchanged. Doubao left them read-only until mid-October; Tongyi Qianwen gave notice of permanent deletion with no grace period.

No accident had occurred, no lawsuit had been filed. A single regulation had simply taken effect. It is a scene that runs head-on against the received wisdom that "China regulates AI loosely."

To read this scene you need one axis for looking at regulation. Some regulation comes after the accident; some stands in the way before it. Call the former ex-post regulation, the latter ex-ante regulation. Ex-post regulation is closer to a fire engine that only rolls out once the fire has started: harm has to occur before the lawsuit and the liability for damages arrive, belatedly. Ex-ante regulation is closer to the occupancy inspection a building must pass after it is fully built and before people move in — it erects gates — classification, audit, filing — before deployment.

The rules that assign liability have so far tended to arrive late, after the harm. In the EU, the product-liability rules that will apply to AI harm do not take effect until December 2026, and their counterpart, the AI Liability Directive, was scrapped in 2025. A binding ban on autonomous weapons, too, blocked by international consensus, has been pushed back yet again — to a November 2026 conference. This last one, though, is a different shape of lateness: not late after the harm, but the agreement itself delayed.

That said, the ex-ante regulation that erects gates before deployment was not newly born in 2026. The EU AI Act, the world's strictest AI regulation, already sorted risk into four tiers back in 2024 (the ex-ante skeleton), and the practices that fall under its prohibited-risk tier have been in force since February 2025. Ex-ante gates have been the EU's basic grammar for years. What newly split in the summer of 2026 is not that grammar but who actually switches it on, and when.

And yet, in the summer of 2026, something moves. Within the single month of July, China, the EU, the United States (federal and state), and Google each put forward a different AI regulation or proposal. The press read this as "regulatory fragmentation" — the diagnosis that rules diverge by jurisdiction, that is, by the national or state-level reach each law covers, so that the only takeaway for global firms is a rising compliance bill. Onto this is layered the received wisdom that "China is lax." Run both pieces of received wisdom through a single axis and a different picture emerges. The axis is this: who actually locks the pre-deployment gate, and when.

The Surface: Regulation Split Four Ways

Set July's material briefly onto this axis.

• China — imposes pre-deployment obligations on providers of emotional-companion services, and shuts the service down if they are breached.
• EU — sorts AI by risk level, and requires systems classified as high-risk to pass a pre-market conformity assessment (a procedure for having compliance with the requirements confirmed in advance).
• United States — the federal government expressly rules out any pre-market licensing requirement, while states like Illinois impose an annual audit on frontier developers (the firms building the largest, most advanced cutting-edge models).
• Google — itself a regulated firm, proposes a federally supervised regulatory body of its own.

The first three are regulations erected by jurisdictions; Google's is a proposal put forward by a regulated firm. One of them differs in kind.

Read as a list alone, "every one for itself" is right. Purpose, form, and effective date all differ. Yet in this list, only one thing has actually landed on users in the flesh. It is that shutdown at the top. The other three remain in the state of documents and scheduled dates. This gap in temperature is already the first clue.

The Refraction: Same Grammar, Split Enforcement

Place the four gates side by side, and the point that they erect a gate and obligations to clear before deployment is something all four share to some degree. China put pre-deployment obligations on companion providers; the EU requires a pre-market conformity assessment for high-risk; Illinois mandates a pre-deployment annual audit on frontier developers. Even Google, itself a regulated party, proposes a pre-deployment gate of its own. Even so, Google's carries different weight, because it has no legal force.

But as we saw, this ex-ante grammar is itself no new discovery. The EU already erected it in 2024, and prohibited practices have been in force since 2025. So a counterargument stands up right here.

Isn't what actually happened in 2026 a retreat from ex-ante regulation, not its spread? The EU pushed its high-risk core back to end-2027; the U.S. federal government expressly ruled out pre-market licensing; the Illinois audit does not begin until 2028. The table just ahead reveals, of its own accord, that three of the four strands are deferral, future, or reversal. Lay the same data out and "retreat from ex-ante" may be a more natural reading than "convergence toward ex-ante."

This counterargument is right. So this piece does not claim that "regulation is shifting from ex-post to ex-ante." That claim can be read, from the same ledger, equally as a retreat — which makes it a safe statement beyond falsification. The answer set down in advance at the top gains its weight here. What truly splits is not the existence of ex-ante grammar but the enforcement of that grammar. Over the same "pre-deployment gate," the places that actually switched it on, deferred it, and pushed it back diverge. Call this gap the enforcement gap.

And at one end of that gap — the spot where the gate was actually switched on — stands a country the received wisdom did not expect.

Lay the same gate out by each jurisdiction's enforcement status, and that gap shows through plainly.

The gap the table shows contains one item pointing in a different direction. It is the U.S. federal government. And here a more fundamental counterargument arises.

The four strands differ not only in the speed of taking effect but in purpose from the start. China is state control, the EU is fundamental-rights protection, the U.S. federal government is, if anything, deregulation. When purposes diverge this much, isn't binding them under one axis a stretch?

The point that purposes differ is correct. In particular, the U.S. federal government is not one point on the gap but an active reverser. It expressly ruled out creating any mandatory licensing or pre-approval requirement for the development and deployment of new AI models, and through a separate executive order it attacks each state's AI laws with lawsuits and funding limits.The two U.S. federal orders are distinct. The one that expressly excluded pre-market licensing is the "Promoting Advanced Artificial Intelligence Innovation and Security" executive order (EO 14409), signed 2026-06-02, whose §3(c) nails down that it authorizes no creation of a mandatory licensing, pre-approval, or permit requirement. The one that attacks state AI laws with lawsuits and funding is a separate "Eliminating State Law Obstruction of National Artificial Intelligence Policy" executive order (EO 14365), signed 2025-12-11.

Grant this outright, and the axis only grows sharper. Within the United States, the side that actually shoulders the pre-deployment gate is not the federal government but the states. While the federal government sticks with ex-post enforcement and deregulation, states like Illinois passed frontier-audit laws under the pressure of federal preemption. The very split of federal and state colliding head-on inside one country is the physical proof that whether to switch the same gate on or push it back is being decided right now. What binds the four together is not a shared purpose. They share one structure — naming in advance the side that will bear responsibility before deployment — but the hand that switches that structure on differs by jurisdiction.

The Reversal: In a Landscape of Retreat, the Only Enforcer — China

At that end of the enforcement gap — the spot where the gate was actually switched on — stands the very China that was said to "handle AI loosely." To see what the received wisdom missed, China's regulation must not be lumped into one. July's China held two regulations of entirely different character.

One is the Implementation Opinions. Announced by three ministries in May 2026, it is the first document to treat intelligent agents as a governance category distinct from generative AI. But this is a non-binding guidance "Opinion" — nothing more. It is written in the register of recommendation and encouragement, and carries no penalty provision.The Implementation Opinions' formal name is the "Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents" (智能体规范应用与创新发展实施意见). Jointly issued by the Cyberspace Administration of China (CAC), the National Development and Reform Commission, and the Ministry of Industry and Information Technology on 2026-05-08. No explicit effective date is confirmed in the original text.

The other is the anthropomorphic-services regulation. Aimed at services that mimic a person's character, thought, and speech to exchange emotions continuously, it is a binding regulation from five ministries. It attaches warnings, rectification orders, service suspension, and even fines. Issued in April, in force on 15 July.The anthropomorphic-services regulation's formal name is the "Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interactive Services" (人工智能拟人化互动服务管理暂行办法 · Order No. 21). Jointly issued by five ministries (CAC, NDRC, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration for Market Regulation), promulgated 2026-04-10, in force 2026-07-15. Fines range from 10,000 to 200,000 yuan under Article 30.

The one that left the shutdown's footprint is this second regulation. It was this regulation, too, that the companies cited as their grounds for switching the features off. The regulation did not stop at a document — it actually switched a product off. With no accident and no lawsuit, a product feature went dark at the deployment stage.

The target of this enforcement is narrow. The scope the regulation aims at is confined to emotional-companion services. Smart customer service, knowledge search, work assistants, and learning and education are expressly excluded, and the safety of frontier models was never the target to begin with. So to read it as "China is clamping down hard on AI across the board" is wrong. There is a more refined reading — that China has always been hard on content and loose on the frontier — and this scene does not break that framing.

What gets overturned here is the more common take, the impression that "China does not clamp down on AI." The place the reversal lives is not the size of the target but the reality of the gate. While the EU pushes its high-risk core back (there has been no scene yet of the earlier-switched-on prohibited-practices layer switching a product off), Illinois waits for 2028, and the U.S. federal government closes off pre-market licensing, the only place that has actually switched a product off at the deployment stage — narrowly though it be — is China. The received wisdom of "lax" was an optical illusion that confused the direction of regulation — state control or fundamental rights — with the existence and enforcement of regulation.

One more counterargument remains.

Even if a provision stipulates an ex-ante gate, if it is not enforced it ultimately returns to ex-post. China's "Opinion" too may in fact be weak-binding administrative guidance; the EU has been deferred; Illinois is 2028; the U.S. federal government is at the frame-and-white-paper stage.

This counterargument is half right. So the axis bends that much too. The Implementation Opinions is in fact a non-binding guidance opinion, so writing "China enforced the Implementation Opinions" makes that sentence itself false. What shouldered the enforcement is the separate anthropomorphic-services regulation, which holds penalty provisions, and the shutdown and data deletion are its physical footprint. That the EU, Illinois, and Google are all still at the declaration stage is granted as is. Once that is granted, "the distance between declaration and enforcement" becomes the real gauge that divides the four strands, and the fact that China alone stands at the enforcement end of that gauge is what overturns the received wisdom.

There is one more layer here. Even if enforcement splits, whether "who bears the liability" stays the same must be answered separately — and it is exactly at that point that the common reading that tries to correct "China is lax" this time paints over in the opposite direction. It is the explanation that "China erected a three-tier risk-grade system and assigned liability by tier." Read the original text directly and this picture does not stand either. The "3" in the Implementation Opinions is not a ladder of risk grades but a taxonomy of subject types holding decision authority. It splits into decisions only the user can make, decisions that require the user's authorization, and decisions the agent makes autonomously. Differentiation by risk level is a separate sector-based axis, and there is no provision at all that splits the liability-bearer by tier. The original text merely enumerates developers, deploying firms, and service providers together in one bundle.

The sample in which this overpainting was confirmed amounts only to a few Western summaries. It is not something to conclude as an error of Western analysis across the board. Still, it is worth flagging. Design China's liability attribution atop the nonexistent model of "differentiation by tier," and compliance actually misjudges — believing that at a low tier the developer bears it, so the deploying firm relaxes. The side the original text points to as bearing responsibility closes into one, regardless of tier.

Right here the hardest counterargument arises.

If the side that bears responsibility that way is the same deploying firm whether ex-post or ex-ante, then isn't "enforcement splits" ultimately just wordplay? When the terminus — the deploying firm bears responsibility if an accident occurs — stays the same.

That the terminus is the same, I grant. As long as the Implementation Opinions does not split the liability-bearer by tier, the answer to "who bears it" stays as it was, toward the deploying firm. The absence of a provision does not actively prove the judgment this blog has written before — that liability remains with the deploying firm. Nor does it overturn it. It merely keeps that judgment as it was. What changes is the mode of enforcement. Instead of a lawsuit that takes years, an administrative gate — fines, rectification orders, service suspension — operates, so that a product can be switched off at the deployment stage with no accident and no lawsuit.

And yet, written honestly, the empirical proof of this mode of enforcement is still just one case, China's companions. The rest is provision text. The current footing for the judgment that the enforcement gap is real is narrow, and the wider judgment is entrusted to a falsifiable prediction hung at the end.

Cross-Domain Analogy: Flags of Convenience

That enforcement splits by jurisdiction also means firms gain options. When the strength of taking effect and of enforcement differs by jurisdiction, the greatest risk is not that compliance costs rise. It is that liability slides toward the jurisdiction where it is softest. The place that went through this structure a century earlier is shipping.

It is rare for the actual owner of an oil tanker flying Panama's flag to be in Panama. The owner registers the ship not under their home country but under the flag of whichever country has the lightest registration procedures and tax burden — Panama or Liberia, say. It moves liability and regulatory jurisdiction outside the home country. This is called a flag of convenience. A flag of convenience is no exceptional practice. Three registry states — Liberia, Panama, and the Marshall Islands — hold nearly half (46.5%) of world tonnage. Nor has international law left this slide unattended. There is the "genuine link" principle — that the flag state, that is, the registry state above, must exercise effective control — but its enforcement was weak, and the 1986 UN convention meant to strengthen it has failed to take effect for forty years. For a century, extraterritorial and international regulation has never fully sealed off flags of convenience.

AI services carry the same incentive. When taking effect and enforcement split by jurisdiction, firms come to want to settle in the jurisdiction where liability is imposed most lightly. But let me make the analogy's limit clear. An AI service is not, like a ship, a thing registered under a single flag state; it is closer to tax avoidance that shifts the location of the legal entity and the data center. Even so, the structure alone — "liability slides to the softest flag" — is the same. The pre-deployment gates and extraterritorial-application clauses that China and the EU erected can be read as a defense to block that slide within their own markets.

Here is a counterargument aimed at the analogy itself.

AI is not a ship. Because of the EU AI Act's extraterritorial application, no matter how soft a jurisdiction you place the service in, to reach the EU market you must follow EU regulation. So the argument goes: the so-called Brussels effect neutralizes flags of convenience.

Extraterritorial application is indeed in the provisions. If the output is used in the EU, the regulation reaches the provider even when the provider sits in a third country.The provision basis for extraterritorial application is Article 2(1) of the EU AI Act (Regulation 2024/1689). Under (a), a provider that places a product on the EU market is covered regardless of place of establishment; under (c), even one located in a third country is covered if its output is used within the Union (the effects doctrine). The high-risk ex-ante gates are conformity assessment, registration in the EU database, risk management, and the like. The flight of a service facing a large market is partly blocked by this clause. Granted.

But this concession does not fight the analogy — it holds it up. Because the flight incentive is real, jurisdictions bar the gate in advance. And the sealing is partial. Open-source models, small operators, business-to-business dealings, and purely extraterritorial deployment still leave room to move to a softer jurisdiction. Shipping, too, equipped with port-state control and international conventions, failed for a hundred years to stop flag-hopping. Extraterritorial application cannot erase the flag-of-convenience incentive. It only partly suppresses it, and only before a large market.

The place that went through the same structure from a different angle is taxation. At the end of a competition in which countries raced one another to cut corporate tax rates (the OECD average from about 40% forty years ago to about 23% now), the OECD moved toward ex-ante coordination with a 15% global minimum tax joined by some 140 jurisdictions. But this solution is an ex-post settlement in which, after the fiscal year ends, the parent company's home country collects the difference, and it is a multilateral coordination that binds many countries together. It runs a different path from AI's unilateral ex-ante barring. It shares only the skeleton — "fragmentation breeds arbitrage, and ex-ante coordination suppresses that arbitrage" — and this is no place to assert that the trajectories overlap.

The Landing: Rules Diverge, Enforcement Moves Up

The surface diagnosis that "regulation splits by jurisdiction, so the only result is a rising compliance bill" is only half right. The real axis that splits beneath it lies elsewhere. The ex-ante grammar of erecting a gate before deployment has been around for some years already and is shared to a degree by all four. What actually splits is who switches that gate on, and when — that is, enforcement. China enforced; Illinois is scheduled; the EU deferred; the U.S. federal government pushed the gate itself back. And the only place that has so far actually switched a product off at the deployment stage is the one country said to be lax — China. Even that target stopped at emotional-companion services, which do not touch frontier safety.

The cost a global AI firm will pay is twofold. One is the labor of translating the rules jurisdiction by jurisdiction, and this is still large. The other is the unpredictability that arises because which jurisdiction will actually close the gate, and when, is all over the place. Over the same accident, in one jurisdiction the bill for liability comes at the speed of a lawsuit, and in another it is blocked in advance at the deployment gate. What is deciding that divergence right now is the enforcement gap.

This is not China's users' business alone. Some feature of the AI service you use every day may, depending on which jurisdiction that service settled in, quietly disappear one day or switch on differently by region. The companion shutdown that China's users went through in July is the first scene of it. For firms, it means splitting products and data by jurisdiction; for users, where you live decides what features you can use. The question shipping met a century ago now lies before AI. If liability slides to the softest flag, where is that flag?

This judgment has stakes riding on it. If the EU's high-risk gate (December 2027) and Illinois's audit (January 2028) switch on as scheduled and leave enforcement footprints — sanctions or audit reports — then the judgment that the enforcement gap is widening is right. Conversely, if deferrals repeat (the EU has already pushed back once) and ex-post lawsuits remain the only actual channel of enforcement, then deployment-stage enforcement was the singular case of China alone, and I am wrong. That answer comes by early 2028 at the latest.

In the end, what this split decides is one thing. When AI gets something wrong, where — and when — does the liability remain? The four strands are now transcribing that answer, each at a different speed, to the front of deployment. The sentences of the rules already point that way, all four.

Only, the hand that has set a lock on that sentence is still just one.