On 30 September 2026 the Cyberspace Administration of China (CAC) launched a nationwide, two-month 'Qinglang' campaign against fake personas in short videos. The notice says fake-persona videos may not be pushed in rankings or recommendation slots. It tells platforms to self-inspect across content creation, traffic distribution and review. That is the notable part. China's algorithm rules, in force since March 2022, have mostly been described as a filing and transparency regime. This campaign uses them to police what gets distributed.
The case for the campaign
The strongest argument for it is consumer protection, and it deserves a fair hearing. The CAC lists five problem areas. They are people posing as vulnerable groups, fabricated professional credentials, AI or impersonated real figures, staged displays of wealth and superstition, and invented conflict storylines. Several of these are plain fraud. The notice cites AI-generated audio or video of public figures used for fake endorsements, and people posing as doctors or lawyers to sell paid consultations. It also singles out fake personas aimed at older users, such as 'AI CEO' or 'wellness master' characters. Anyone who has watched a relative lose money to a synthetic celebrity endorsement will not find the goal controversial. Most democracies already treat deceptive endorsements and impersonation as unlawful.
What the legal machinery actually does
The Algorithmic Recommendation Management Provisions were issued by the CAC, the Ministry of Industry and Information Technology, the Ministry of Public Security and the State Administration for Market Regulation, and took effect on 1 March 2022. They require services with public-opinion attributes or social mobilisation capacity to file through the algorithm filing system within ten working days of launch (Article 24). Regulators then issue a filing number, which the provider must display (Articles 25-26). Article 17 gives users a way to switch off personalised recommendation. Article 14 bars manipulating rankings or search ordering. Article 11 requires 'positive mainstream' content to be placed on key slots such as homepages, hot lists and rankings.
The last provision matters most for this campaign. The same law that obliges platforms to promote approved content now lets the regulator order them to withhold distribution from disfavoured content. The registry is the enforcement hook. Because every ranking system is on file, the CAC can address a named system and a named company, not merely individual posts.
Labels as a compliance trap
The campaign builds on earlier steps. China's AI-labelling rules took effect on 1 September 2025. As Loeb & Loeb summarised, they require visible labels and embedded metadata on AI-generated content. In May 2026 the CAC reported that platforms must give uploaders six mandatory labels, including 'fictional or dramatized content', 'AI-generated content' and 'personal opinion'. It also reported that platforms had cleared more than 520,000 violating short videos since January and penalised over 68,000 accounts.
The new notice closes a loophole in that system. Platforms may not let a label such as 'fictional performance', 'AI-generated' or 'personal opinion', or a no-harm disclaimer, excuse a fake-persona video from the rules. In effect, a label no longer protects a video from enforcement. It is also a fact the regulator can cite against an account. For creators this makes the line between a satirical character and a deceptive persona the platform's call, made under the threat of a campaign.
Where proportionality breaks down
Our objection is to the design, not the aim. Three problems stand out.
First, several of the categories are defined by intent and effect, not by a verifiable act. Impersonating a public figure with synthetic video is checkable. 'Staging conflict storylines' or 'demeaning particular regions or groups' is not. A rule that cannot be predicted will be over-complied with, and a platform facing a two-month campaign will remove borderline comedy, role-play and fiction along with the fraud. The notice itself says accounts of an egregious nature will be 'resolutely banned', and gives no appeal path or penalty scale.
Second, distribution controls are less visible than takedowns. A video that is removed leaves a record. A video quietly excluded from recommendation slots does not, and the creator may never learn why reach fell. Because the 2022 rules already give the regulator a filed description of each system, the missing piece is not information but notice to users. A proportionate regime would pair demotion with a stated reason and a route to contest it.
Third, a campaign format substitutes for rulemaking. Two-month drives produce sharp spikes in enforcement and then lapse, which makes compliance planning hard for smaller platforms and independent creators. Durable rules with defined offences would cut fraud as well, and would impose less collateral cost on lawful expression and on the creative industries that depend on it.
What other jurisdictions can take from this
The useful lesson is not that fraud should be tolerated. It is that the same tool can be built for narrow or broad ends. The EU's Digital Services Act and the UK's Online Safety Act also push platforms toward systemic risk management of ranking systems, but both pair that with published reports, user redress and judicial review. Those procedural safeguards are what separate an anti-fraud measure from a general-purpose lever over reach. China's campaign has the lever without the safeguards.
For builders, the practical point is that recommender systems in China are now compliance surfaces. Ranking logic, label handling and persona verification are all subject to inspection. Companies operating there should document why each demotion happened, because the regulator may ask.
Bottom line
Targeting fraudulent personas is defensible, and the AI-impersonation category in particular addresses real harm. But steering what recommender systems may promote, through a time-limited campaign with broad categories and no stated remedies, trades predictability for speed. The better path is narrow offence definitions, stated reasons for demotion and a right to contest it. Those would protect users from fraud while leaving honest creators room to work.