On 6 July 2026 the Cyberspace Administration of China (CAC) reported on the first phase of the “Qinglang — Rectifying AI Application Disorder” special action. Launched in April 2026, phase one identified four categories of problems.
Category one: failure to perform large-model filing and registration duties
This category ranks first and targets “should have filed but did not” and “should have registered but did not”. Generative AI services must complete filing before launch, and applications or features that call on already-filed model capabilities through APIs or otherwise must complete registration — a hard requirement at the access stage. For the process, see our Algorithm Filing service page.
The meaning for companies is direct: first check whether a product falls within the scope of filing or registration, then verify that the published information matches the actual situation. Failure to update promptly after a change of name, entity or service form is itself a problem.
Category two: inadequate platform security and content-filtering capability
This category targets platform-side governance capability. Content moderation is not a question of “having a keyword library” but of whether the mechanism covers input and output, whether it is traceable, and whether there are dedicated strategies for high-risk scenarios. With the arrival of new forms such as agents and anthropomorphic interaction, the moderation chain is longer than that of a traditional content platform.
Category three: AI data poisoning
Data poisoning is a relatively new category. It includes both training data maliciously contaminated and harmful or erroneous information injected into a model through user interaction or crawling. For companies providing training and fine-tuning services, this means source review and data cleaning are no longer optional. See our Data Compliance service page.
Category four: incomplete implementation of generative-content labelling
The labelling duty comes from the Measures for the Labelling of AI-Generated and Synthesised Content, which require explicit labels on generated and synthesised content and implicit labels in file metadata. At the implementation level, the easily overlooked part is the implicit label — it must be readable across the generation and dissemination chain, not merely a notice added to the interface.
The phase-one methodology: problem-oriented plus a reporting channel
The report also mentions opening a reporting channel for “AI application disorder”. This arrangement means that future enforcement leads will partly come from user reports; companies cannot simply wait for regulators to arrive but should proactively investigate. A later phase-two report set out seven further problem categories and enforcement priorities, tracked in our AI law hub.
Four things companies can do now
First, complete a filing and registration compliance inventory, confirming whether products fall within scope and whether published information is accurate. Second, turn content moderation into a verifiable process, retaining strategies and logs. Third, establish a source ledger for training and corpus data. Fourth, check whether both the explicit and implicit chains of generative-content labelling are in place.
For support on filing inventories, content-moderation mechanisms and data-compliance ledgers, see our Algorithm Filing and Data Compliance services.
This article is based on public information on the first phase of the “Qinglang — Rectifying AI Application Disorder” special action reported by the Cyberspace Administration of China on 6 July 2026. It is for reference only and does not constitute legal advice.