A Registry That Outgrew Its Blueprint
When the Cyberspace Administration of China (CAC), alongside the Ministry of Industry and Information Technology, Ministry of Public Security, and State Administration for Market Regulation, jointly issued the Provisions on the Administration of Algorithmic Recommendation in Internet Information Services — effective March 1, 2022 — it launched the world's first mandatory algorithm registry. The concept was clear: services using recommendation algorithms with "public opinion properties or social mobilization capabilities" would register through the national filing platform at beian.cac.gov.cn, disclose their mechanisms, and extend users an enforceable right to opt out.
Four years on, the numbers look impressive. As of November 2025, over 5,000 algorithms have been filed through China's national platform, up from the initial 30 providers in August 2022 — a cohort that included Alibaba, Tencent, ByteDance, and Baidu — and 262 registered recommendation services by April 2023. The generative AI wave pushed the queue further, with hundreds of large language models joining the filing system alongside legacy recommender systems. But the gap between filing volume and what those filings actually reveal has widened at a similar rate.
What the Regulation Gets Right
Before critiquing the architecture, the legitimate aims deserve fair statement. Recommendation algorithms shape what hundreds of millions of people see each day — news feeds, search rankings, product surfaces, gig-worker order queues. A regime that gives users the enforceable right to delete their interest tags and disable personalization (Article 17), requires special protections against addictive content for minors (Article 18), and explicitly bans algorithmic price discrimination (Article 21) addresses genuine consumer-welfare concerns that many democratic jurisdictions have been slow to codify.
The November 2024 Qinglang Campaign — formally titled "Rectification of Typical Algorithmic Problems on Online Platforms" — demonstrated the regulatory agenda in practice: targeting homogeneous recommendations that create "information cocoons," manipulation of trending topics, algorithmic suppression of gig workers' rest rights, and discriminatory pricing through behavioural profiling. These are real harms, and naming them specifically in a government enforcement campaign — with administrative penalties imposed on at least three platforms by Shanghai's CAC for operating without completed filings — marks a shift from paper compliance toward operational accountability.
The Disclosure Gap
The execution, however, reveals the limits of a filing-based transparency regime. Carnegie Endowment research found that Weibo's "hot search" algorithm filing described itself as combining "search popularity, discussion popularity, and dissemination popularity" with unspecified coefficients — information far too general to enable meaningful external scrutiny. The same analysis noted that ByteDance reportedly had to translate its systems into simplified language for regulators to follow. When the disclosing body lacks the technical capacity to evaluate filings critically, and when no independent audit rights exist for civil society researchers, the registry becomes a procedure — not a mechanism for genuine accountability.
Public filings that cannot be interrogated by technically equipped outsiders tell us less about algorithmic systems than their operators would prefer we think.
Registration or Licensing?
The generative AI layer has sharpened this tension further. Analysis of the registry found that as of August 2024, roughly 45% of China's approximately 305 identified generative AI models had registered — nearly two years after the requirement applied. For those that did file, the process involved direct pre-deployment testing by provincial CAC authorities, with approval withheld until content controls met official satisfaction. End-to-end compliance for high-impact systems — from design to deployment clearance — now routinely runs six to twelve months.
That process is not registration. It is licensing. Registration is a post-launch notification mechanism with auditable records; licensing is a gatekeeping function with pre-approval authority. China has built the latter while framing it as the former. The ambiguity creates prolonged uncertainty for domestic startups and foreign operators alike, and the compliance gap suggests many are unwilling to run the pre-deployment gauntlet.
The April 2026 Pricing Layer
The most concrete recent development adds a commercial enforcement dimension. On April 10, 2026, the Rules for Price Behavior on Internet Platforms — jointly issued by the NDRC, SAMR, and CAC in December 2025 — took effect. Article 15 explicitly bans platforms from using "data, algorithms, platform rules, and other means" to set different prices based on a consumer's willingness to pay or behavioural habits without disclosure. Platforms must also inform merchants of ranking rules and recommendation algorithm mechanics — a transparency obligation that extends the 2022 regime downstream into the commercial layer.
Separately, SAMR's November 2025 draft antitrust compliance guidelines for internet platforms flagged "algorithmic collusion" as a distinct competition-law risk category, signalling that regulators now view coordinated algorithmic behaviour across platforms as more than a consumer-protection problem — it is a structural market concern.
The Proportionality Question
China's algorithm registry serves three overlapping purposes: consumer protection (opt-outs, minor safeguards, price discrimination bans), market regulation (anti-monopoly provisions, pricing transparency), and content governance (mainstream values alignment, pre-deployment content testing). The first two are defensible and reasonably proportionate to legitimate regulatory goals. The third conflates platform accountability with state control over the information environment — and that conflation is what makes the system's stated transparency rationale difficult to take at face value.
For jurisdictions designing their own algorithmic oversight frameworks — India's evolving IT Rules, Brazil's AI Bill, ASEAN's emerging digital governance compacts — China's registry offers a calibration lesson: filing volume does not equal accountability. What matters is whether the regulating body has genuine technical capacity, whether civil society holds independent audit rights, and whether the content values embedded in the framework are consumer-protective or politically instrumental. China's registry has demonstrated that the first condition is achievable at scale. The other two remain open questions.