On 2 September 2026 the Central Cyberspace Affairs Commission reported on the results of phase two of the special campaign “Qinglang against misuse of AI applications”: more than 5.61 million items of unlawful information cleared, more than 49,000 accounts investigated, and over 2,400 websites and applications dealt with. The focus of this round has moved from filing and labelling to content and the product itself.
The shift of enforcement focus between the two phases
The special campaign was deployed in April 2026. Phase one, reported on 6 July, concentrated on failure to perform the large-model filing and registration obligations, inadequate security and content filtering capacity on AI platforms, poisoning of AI training data, and incomplete implementation of labelling for generated content — questions of whether procedures and capabilities were in place.
Phase two, reported on 2 September, turned to the content ecosystem and the operation of products, and published disposal figures. The two phases are sequential: phase one established whether a company had the procedures and the capability; phase two examines whether the product is actually compliant. Enforcement has entered a results-oriented stage, in which a filing alone is not sufficient and the content a product actually produces and the way it operates become the focus of inspection.
The seven categories addressed in phase two
- Using AI to “remaster” classics and to generate “digital slop” — low-quality, homogeneous content produced in bulk;
- Producing and publishing false or misleading information;
- Impersonating or counterfeiting others;
- Producing and publishing violent, vulgar and other harmful information;
- Infringing the rights and interests of minors, for example by generating content that harms or ridicules the image of a minor, or by engaging in online bullying;
- Using AI “hosting” to conduct online water-army activity, such as simulating real people to register and operate accounts in bulk, group-controlling engagement and comment sections, and manufacturing fake traffic;
- Non-compliance in AI products, services and applications.
The seventh is a residual category which, read together with the filing, registration and labelling requirements of phase one, in effect brings the compliance of a product as launched within the scope of inspection.
What the figures mean
More than 5.61 million items of unlawful information, more than 49,000 accounts and over 2,400 websites and applications: an order of magnitude that shows this is not a reminder by way of interview but clean-up and disposal at scale. The practical consequence for companies is that account bans and application removals may be triggered directly by problems in the content ecosystem, not only by incomplete procedures.
Points for companies to check
Taking the focus of both phases together, what companies should check now is a set of specific matters: whether the large model has completed filing or registration, and whether an application calling a filed model has completed registration; whether visible and implicit labels for generated content have been implemented, and whether users are told that content is AI-generated; whether the technical measures and staff for content moderation and filtering are proportionate to the user base; whether there are mechanisms to prevent the generation of impersonated images and voices and the bulk production of false information; whether the product is designed with functions that generate homogeneous, low-quality content in bulk; whether it may generate content that harms or ridicules the image of minors, and whether age identification and protection measures are in place; whether it offers the ability to register in bulk, group-control accounts or manipulate engagement; and whether the infringement notice channel, handling times and record retention are workable.
These points connect with the filing, registration and labelling obligations to form a continuing inspection of how a product operates. For the relevant procedural routes, see Algorithm Registration and Large-Model Filing.
The direct consequence of this round is the intensity of disposal: accounts, applications and traffic can be cleared at the same time in a single campaign. Treating content moderation capacity, implementation of labelling and complaint handling as a minimum closed loop for remediation is the most realistic step now. For assistance in responding to the campaign, designing content moderation mechanisms or responding to regulatory inquiries, please contact Geason Law Offices.
This article is prepared on the basis of the public reports of the Central Cyberspace Affairs Commission on the special campaign “Qinglang against misuse of AI applications” (deployed in April 2026; phase one reported on 6 July; phase two reported on 2 September). It is for reference only and does not constitute legal advice.