On 2 September 2026, secondary reports said the Cyberspace Administration of China (CAC) had published results from the second phase of its "Clean Up AI Application Chaos" campaign. The reported figures were more than 5.6 million pieces of content removed, about 49,000 accounts closed, and action against more than 2,400 sites and apps. Those numbers come from a secondary report (IT Home, relayed by Race to AGI). We could not locate the CAC primary text, and the report says nothing about recommendation systems. The figures should be read with that caveat. What can be verified is the legal machinery underneath: the algorithm filing registry. That registry is the more consequential story.
What the registry actually is
The registry comes from the Provisions on the Administration of Algorithmic Recommendations for Internet Information Services, signed on 31 December 2021 and effective 1 March 2022. The CAC, the Ministry of Industry and Information Technology, the Ministry of Public Security and the State Administration for Market Regulation issued them jointly (CAC text). Article 24 requires providers with "public opinion attributes or social mobilization capacity" to file within ten working days of starting service. The filing covers the provider's name, service form, application domain, algorithm type, a self-assessment report and the content it intends to make public. The filing system went live on 1 March 2022, and the CAC's launch notice says filings are self-reported, do not amount to endorsement, and may not be used for promotion (CAC notice).
The strongest case for the regime
The case for the registry is serious. Recommender systems shape what hundreds of millions of people see, and users usually cannot inspect them. The Provisions pair filing with user-facing rights. Article 16 requires providers to disclose that algorithmic recommendation is in use and explain its basic principles. Article 17 requires an option that is not based on personal characteristics, or an easy way to switch recommendation off, and lets users delete the tags used to profile them. Article 18 protects minors. Few jurisdictions legislated switch-off rights for recommenders as early as China did, and the EU's Digital Services Act later adopted a comparable idea. A registry that tells a regulator which systems exist is a reasonable precondition for any oversight, and it is hard to argue against in the abstract.
Where the design falls short
Carnegie's 2022 analysis of the registry found the public side thin. The first batch of 30 filings, published in August 2022, included Tencent, Alibaba and ByteDance. Each ran to a single page of short answers. The authors judged the descriptions "pitched at such a high level as to be almost completely devoid of meaningful detail", and said the Weibo hot-search entry was so generic that an outsider could essentially guess it (Carnegie Endowment). The more detailed material, including dataset information and the security self-assessment, is submitted to the regulator and is not public. Carnegie also noted that "security" is undefined, so the self-assessment could mean anything from controlling public expression to preventing AI misuse.
That ambiguity matters because the filing duty attaches to systems with "public opinion attributes", a category the Provisions do not define. Providers must guess which of their systems are in scope. Over-filing is the safe choice, and it hands the regulator a standing inventory of every system it might later question.
From registry to enforcement lever
The campaign reporting shows how that inventory gets used. The CAC's first phase, launched in April 2026, was reported in July 2026 as having dealt with more than 14,000 AI products, cleaned up more than six million pieces of non-compliant content and handled more than 26,000 accounts. Reported focus areas included large models "not being registered as required", weak review and filtering capacity, data poisoning and content labelling (ChainCatcher, relaying CAC). That report does not give a count of unregistered models, and it does not mention the recommendation algorithm filing at all. Other registration regimes, such as those for generative AI, appear to be involved.
The pattern is still instructive. A registration obligation lets an enforcement campaign start from a list of known systems and ask which ones failed to file, filed inaccurately or failed to meet vaguer duties such as content review. Those duties are enforced through takedown and account-closure counts. Neither the first-phase nor the unverified second-phase figures are tied to a published standard, a stated error rate or an appeal route. A count of 5.6 million removals says how much was removed. It says nothing about whether the removals were correct.
What a proportionate version would look like
Proportionate regulation of recommender systems is achievable, and the registry's own user-rights provisions point toward it. Four changes would make the system more credible without removing oversight:
- Define the trigger. Replace "public opinion attributes" with measurable thresholds such as reach or user numbers, so providers know whether they are in scope.
- Publish meaningful disclosures. Require filings that let outside researchers test claims, as the EU's DSA does through vetted researcher access.
- Report error rates. Campaign statistics should include appeal outcomes and reinstatements alongside removal counts.
- Separate user-protection duties from speech control. Switch-off rights and minor protections can be enforced without a content-removal mandate that sweeps in lawful speech.
For innovators, the practical cost is uncertainty. Large platforms can afford compliance teams that treat filing as routine. Smaller developers face the same self-assessment, unclear scope and campaign-style enforcement with far fewer resources. That tilts the market toward incumbents, which is the opposite of what an open and competitive internet needs.
Bottom line
The registry is not a transparency tool in any sense an independent researcher would recognise. It is an administrative inventory that gives the regulator visibility and leaves the public with little. The campaign headline numbers are not yet confirmed by a primary text, and they would not settle the question even if they were. Whether the system protects users or only supervises platforms depends on what the CAC is willing to publish about how it applies these rules, and so far it has published very little.