On 30 April 2026 the Cyberspace Administration of China (CAC) issued a notice launching a four-month “Qinglang — Rectifying AI Application Disorder” special action across the country; the news was reported in the media from 1 May. It follows an earlier round of governance and marks another concentrated clean-up of the generative AI application ecosystem.
A two-phase framework
The action runs four months and proceeds in two phases. The first phase is the “Qinglang — Typical Violations in AI Application Services” campaign, which takes typical violations at the platform and product level as its entry point. The action also focuses on fourteen prominent problem categories.
Placing “typical violations” in the first phase points at the most visible disorder — the kind that appears at the front end of a product: what it generates, how it labels content and how it presents material to minors. Such problems are easier to identify and document, and easier to push toward rectification within a shorter cycle. The point of phasing is to clear the surface-level problems first, then reach deeper into the chain.
Fourteen categories, grouped by obligation
The fourteen categories look scattered but fall into several groups of obligations.
The first is content labelling. Failure to implement synthetic-content labelling is listed among the leading problems. The duty is not new; it is the implementation of existing rules, and the technical detail and process are set out in our deep synthesis filing guide.
The second is order in the content ecosystem. Using AI to “mash up” classic works and generating “digital Luoshui” point to bottomless reworking and mass reproduction of existing and public content. Such conduct may infringe copyright and mislead public perception.
The third is false information. Generating false or inaccurate information is listed separately, pointing to AI’s risk to information authenticity.
The fourth is protection of minors. Generating “abusive” or “cult” videos of children and bottomless “secondary creation” of copyright animation are named explicitly. The related boundaries and platform duties are discussed in our Qinglang minors’ online environment guide.
The fifth is data and model security. AI data poisoning is included, showing that governance has moved from the output end to the input end: the quality and security of training data are now in view.
Labelling and content governance
For most companies the action lands on two things: labelling and content governance.
The key question on labelling is what counts as compliance. Visible notices and hidden markings must be handled together, and completed at the generation stage rather than annotated afterwards.
Content governance requires review capacity that covers the generation side. Where a product can generate images, video or text, review must target model outputs, not only user inputs. Content involving minors needs additional detection and interception.
From self-assessment to rectification
During the action, companies need preparation they can act on. The pragmatic approach is to work through the fourteen categories against their own products: which capabilities touch a category, what control measures exist, and whether records are auditable. Such a self-assessment is best done jointly by product and compliance teams, because identifying the issue requires product understanding while handling it requires compliance judgement.
Where violations do exist, timely clean-up is the basic action; user and public reports need a working intake and handling channel, described in our AI misuse reporting channel. The first-phase approach and self-assessment points are reviewed in our Qinglang AI clean-up phase-one guide.
The action lasts four months, but it tests a company’s day-to-day content and data governance. For support on content labelling, review mechanisms and data compliance, see our Data Compliance service.
This article is based on the CAC’s special action notice of 30 April 2026 and press reporting from 1 May. It is for reference only and does not constitute legal advice. Specific requirements should be based on the official documents.