China's Algorithm Filing System — Mandatory Disclosure to Regulators Without Public Transparency

China's Provisions on the Administration of Algorithmic Recommendation for Internet Information Services, issued by the Cyberspace Administration of China (CAC) and effective 1 March 2022, require platforms whose algorithms have 'public opinion properties or social mobilisation capacity' to file for a registration number with CAC or its provincial branches. A parallel regime under the 2023 Interim Measures for Generative AI Services extended filing obligations to generative-AI models and applications.

The filing mechanism requires platforms to submit disclosures before deployment; non-public sections of the filings, such as training datasets and security self-assessments, are visible only to regulators.

By August 2024, CAC had confirmed that 190 generative-AI models had filed successfully under the regime, and independent tracking around the same period found that only a portion of submitted models (estimated near half) cleared filing at any given time — evidence the mechanism screens submissions rather than acting as a rubber stamp, and that it now operates at national scale across thousands of platforms' recommendation algorithms as well.

An independent review by the Carnegie Endowment for International Peace of the disclosures made public found them 'pitched at such a high level as to be almost completely devoid of meaningful detail': Weibo's 'hot search' ranking algorithm, for example, is described only as combining 'search popularity, discussion popularity, and dissemination popularity' multiplied by an 'interaction rate coefficient' — a description Carnegie says an outside observer with no prior knowledge of the algorithm 'could essentially guess.' Carnegie concludes that, based on what is public, the filings 'provide no meaningful insights into the algorithms, how they were trained, or how they might perform,' even though non-public sections may contain more detail visible only to regulators. This is included as an honest cautionary case: a compliance mechanism that has scaled institutionally and legally but delivers little of the public accountability that algorithm registers elsewhere aim to provide.