On 15 October 2026 China published a notice, signed 12 October, launching a two-month campaign against improper military-related online marketing and sensationalist content. The notice, carried by the official rumour-refutation platform Piyao and by Guangming Daily's site, was issued jointly by four bodies: the Cyberspace Administration of China (CAC), the Ministry of Veterans Affairs, and the Central Military Commission's Political Work Department and Political and Legal Affairs Commission. It is not a CAC action alone, and it does not mention recommendation algorithms. Its one technology reference is to content "generated using AI technology".
That matters, because the campaign is easy to over-read. It is a content campaign aimed at ten categories of conduct. These include fake "military exam guaranteed pass" schemes, paid consulting that misreads service and discharge policy, sales of uniforms and imitation gear including through livestreams, fabricated military "stories" made with AI or cut-and-paste editing, impersonation of service members or veterans for fraud or promotion, and discussion of sensitive details such as unit locations and equipment.
The strongest case for the campaign
The steelman is real. Several of these categories describe ordinary consumer fraud: people paying for guaranteed admissions, or buying counterfeit goods from someone posing as a veteran. Impersonation and fake credentials are harmful in any jurisdiction, and most democracies have fraud and false-advertising law that reaches them. Platforms also have commercial reasons to clean up such accounts. A regulator that tells platforms to screen and remove accounts using a veteran's identity to sell products is addressing a genuine harm.
Where the design becomes a problem
The difficulty is the mix. The same list that covers fraud also covers "negative hype" about military academies and conscription and discussion of military topics the authorities consider sensitive. Fraud is defined by deception and loss. "Hype" and "sensitive" are defined by the regulator. A campaign format, with a fixed window, platform quotas to screen and dispose of accounts, and published typical cases, rewards over-removal, because the platform bears the penalty for missing a violation and nothing for deleting a lawful post.
A proportionate approach would separate the two. Prosecute impersonation and fraud under existing law, with a named complainant and a remedy. Leave critical commentary and satire to be argued with, not deleted. Evidence for the open-internet position is simple: fraud cases can be proved without ruling on what speech is acceptable.
The algorithm side is a different instrument
China's recommender governance runs through the Internet Information Service Algorithmic Recommendation Management Provisions, issued by the CAC with the Ministry of Industry and Information Technology, the Ministry of Public Security and the State Administration for Market Regulation (Order No. 9). They were signed on 31 December 2021 and took effect on 1 March 2022. Article 24 requires providers of algorithmic recommendation services with "public opinion attributes or social mobilization capabilities" to file through the CAC's algorithm filing system within ten working days of starting service. Article 17 requires an option for users to switch off algorithmic recommendation, or to receive results not based on personal characteristics. Article 18 bars pushing harmful information to minors and using recommendation to lead minors into addiction.
Some of this is sensible and has analogues abroad. A user-facing off switch resembles the non-profiling option required of very large platforms under the EU Digital Services Act, and protections for minors are widely shared goals.
What has changed recently is the plumbing. On 9 November 2026 the CAC announced that a new "Internet New Technology and New Application Information Service Filing System" had gone live at beian.cac.gov.cn, handling generative AI, algorithm, deep synthesis and blockchain filings in one place. The old algorithm filing and blockchain filing systems were retired at the same time. The notice describes a consolidation of portals. MediaNama's reading that it adds no new requirements or penalties is that outlet's characterisation, not the notice's own language.
What registration does and does not do
The two developments share a regulator, a cadence of periodic campaigns and a preference for ex ante visibility. They are not linked in the documents we could verify, and we found no evidence that the campaign uses filing data. Treating them as one enforcement programme would go beyond the sources.
The more useful comparison is about accountability. Registration gives the state a catalogue of who runs which system. It does not by itself give users, researchers or courts the ability to test whether a recommender behaves as described, and nothing in the notices we reviewed creates that. Transparency regimes that work, such as the DSA's researcher access and risk-assessment duties, pair disclosure with an external check. A registry that is only visible to the regulator is a licensing tool with a transparency label.
For innovation, the cost is practical. A single portal lowers friction for large firms with compliance teams. Smaller developers still face ten-working-day filing deadlines, change notifications within ten working days, and cancellation filings, while content campaigns raise the cost of being the platform that missed something.
A proportionate path
- Keep fraud and impersonation enforcement tied to identifiable victims and ordinary legal process.
- Publish the criteria for what counts as "hype" or "sensitive", so platforms are not guessing.
- If algorithm filing is to serve the public, add an outside audit channel; a database only the regulator reads cannot answer user complaints.
- Evaluate the campaign afterwards. We found no results report for the earlier campaigns, and one for this two-month window would let others judge whether it cut fraud or just removed speech.
The lesson is modest. China is running two distinct tools, content campaigns that act on outputs and filing that acts on systems, and neither, on the public record, tells us whether users are better protected.