The Cyberspace Administration of China says a nationwide campaign against the misuse of artificial intelligence has removed more than 5.61 million pieces of unlawful or rule-violating online content. Regulators also took action against more than 49,000 accounts and dealt with over 2,400 websites and applications.

The stated targets are AI-generated misinformation, impersonation, violent or vulgar material, content affecting minors, and other abuses involving generative tools. It is a large number, announced by the body that produced it, and it is worth being clear about what can and cannot be learned from it.

What an enforcement total measures

A removal count measures enforcement activity. It does not measure the size of the problem, the accuracy of the removals, or whether the situation improved.

Consider what 5.61 million would have to be compared against to mean anything. Against total content posted in the period, it might be a rounding error or a substantial share — the denominator is not published. Against the number of items reviewed, it would tell you the hit rate. Against appeals upheld, it would tell you the error rate. Against the same campaign a year earlier, it would tell you whether AI misuse is growing or whether enforcement capacity is. None of those comparisons is available, and a figure without a denominator is not a measurement. It is an announcement.

This is not a point about China specifically. Platform transparency reports published in Europe and the United States have the same structural weakness, and researchers have been complaining about it for a decade. The difference is that in some jurisdictions an outside party can, in principle, test a sample.

The rules underneath the campaign

China built its generative AI rulebook earlier than most. Providers already carry obligations covering personal data, content controls and security, and rules that took effect in 2025 require AI-generated text, images and other outputs to be labelled as such.

Labelling is the part of that framework other governments have converged on. The European Union’s AI Act carries transparency duties for synthetic content; a series of American state laws address chatbot disclosure and AI-generated media; several proposals now moving through legislatures elsewhere borrow the same idea. The disagreement between jurisdictions is less about whether AI output should be identifiable and more about who checks, and what happens when it is not.

That is where an enforcement campaign becomes informative in a way the headline number is not. A labelling rule with no enforcement is decoration. A labelling rule enforced by mass takedown works, in the sense that the rule is applied — while also creating an apparatus capable of removing content for reasons that have nothing to do with AI.

Two readings, both defensible

One reading, made by people who study online harms, is that China is doing something other governments are still arguing about: deepfake impersonation, synthetic sexual imagery and AI-generated fraud are real harms that arrive faster than legislative timetables, and a regulator willing to act at scale removes some of them from circulation.

The other reading, made by press freedom and digital rights organisations, is that categories like rumour, vulgarity and content harming social order are elastic by design, that the same machinery used against synthetic fraud is used against political speech, and that AI misuse supplies a publicly defensible reason to expand it.

These are not mutually exclusive. Both can be true of the same 5.61 million items, and without a breakdown by category there is no way to establish the proportions. That absence is itself worth noting: the regulator chose which figures to publish.

The 49,000 accounts are the more interesting figure

Content removal is cheap and reversible: the same material is reposted within hours by whoever produced it. Account-level action is different, because it attaches consequences to a person or an operation rather than to a file. So does dealing with 2,400 websites and applications, which implies action against the tools rather than only the output.

If the aim is to reduce AI-generated fraud and impersonation, going after the services that generate it at volume is the intervention with the longer half-life. It is also the one with the widest collateral reach, since a general-purpose tool used by fraudsters is also used by everyone else.

What happens elsewhere this month

Several other jurisdictions reach decision points in September. Brazil’s Senate has scheduled a final floor vote on its AI framework bill, modelled in part on the European approach, for the middle of the month. The United Kingdom’s AI Regulation and Safety Bill goes to committee stage in the House of Lords later in September. Those two texts will settle far more about the shape of AI regulation outside the United States than any enforcement statistic.

The comparison worth making is not democracy versus authoritarianism, which produces a conclusion before the evidence. It is capacity: China has rules and machinery to apply them, Europe has rules and is still building the machinery, and much of the rest of the world has neither yet. Whether the rules are good ones is a separate question from whether they are enforceable, and countries are answering the two in a different order.

Related from Plain Sight News: the G20 AI statement China signed, and what a text both sides accept can contain.

Sources: Cyberspace Administration of China via Tech Startups; Center for Democracy and Technology; published legislative timetables for Brazil’s Bill 2338/2023 and the UK AI Regulation and Safety Bill.