NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

China's Generative AI Measures Tie Public Launch to Filing and Safety Review

The seven-department interim measures moved from draft to effect in 2023, setting out scope, training-data duties, labelling and algorithm filing.

2026-11-22 · 1019 words · NeuroAI
China's Generative AI Measures Tie Public Launch to Filing and Safety Review

An app that answers questions in public occupies a different regulatory position from an internal research model. That line was drawn in a 2023 rulebook, and its distinctions still structure how generative AI (生成式人工智能) is offered in China.

Key takeaways

  • The Interim Measures for the Management of Generative AI Services (生成式人工智能服务管理暂行办法) were issued by seven departments under Order No. 15 (第15号) and took effect on 15 August 2023.
  • The text passed through four documented stages: a draft for public comment published on 11 April 2023, approval on 23 May 2023, publication on 13 July 2023, and effect on 15 August 2023.
  • Scope is defined by public availability inside China, not by the technology alone; research and development not offered to the domestic public falls outside it.
  • Article 17 links services with public-opinion attributes or social-mobilisation capability to a security assessment and algorithm filing (算法备案).

Four stages, precisely dated

The sequence is worth stating in full because the document is often confused with its draft.

Draft for comment (征求意见稿). On 11 April 2023 the Cyberspace Administration of China published the Draft Measures for the Management of Generative AI Services (生成式人工智能服务管理办法(征求意见稿)) and invited feedback through 10 May 2023. The draft used a different title and a different article structure.

Approval (审议通过). The final text was approved at the CAC's 12th chamber meeting of 2023 on 23 May 2023, with the agreement of the other six issuing departments.

Publication (公布). Order No. 15 was signed on 10 July 2023 and published on 13 July 2023 on the CAC and government websites. The order was issued jointly by the CAC, the National Development and Reform Commission, the Ministry of Education, the Ministry of Science and Technology, the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration.

Effect (施行). Article 24 sets 15 August 2023 as the implementation date.

The distinction between the draft and the published measure matters: the April text contained provisions that did not appear in the final version, so quoting the draft as current law is a common error.

What the text covers

Article 2 defines the regulated activity as using generative AI to provide generated text, images, audio and video to the public within China. Two boundaries follow.

One is geographic and public-facing. Organisations, companies, universities and research bodies that develop or apply generative AI without offering a service to the domestic public are outside the stated scope.

The other is sectoral. Activities in news publishing, film and television production, and literary creation are subject to separate rules where those exist. In other words, a generative service aimed at the general public and one embedded in a regulated content industry are not treated identically.

Article 22 defines the parties. A provider is an organisation or individual offering the service, including through a programmable interface. A user is an organisation or individual generating content with it. Writing API access into the definition closes off a route around the service label.

Training data and labelling

Articles 7 and 8 address inputs rather than outputs.

Article 7 sets out requirements on training-data handling: lawful sources and base models, respect for intellectual property, consent or another lawful basis where personal information is involved, and measures to improve data quality in terms of authenticity, accuracy, objectivity and diversity.

Article 8 addresses data labelling (数据标注). It calls for clear, specific and operable labelling rules, quality assessment with sampling checks, and training for labellers. This is the part of the measure that reaches the data-supply chain rather than only the model publisher.

Article 12 is short and consequential: providers label generated images and video in accordance with the earlier deep-synthesis provisions. That reference point was later built out in detail by the Measures for Labeling AI-Generated Synthetic Content and a companion mandatory standard, both effective on 1 September 2025.

Filing and security assessment

Article 17 is a provision frequently cited outside China. It addresses generative services with public-opinion attributes or social-mobilisation capability, linking them to a security assessment and to the algorithm filing, change-filing and cancellation-filing procedures under the algorithmic-recommendation rules.

The article does not say that every generative service undergoes that process; it identifies a category by capability and attaches the procedure to that category. That is a classification rule, not a blanket step.

Article 15 covers complaint mechanisms: providers set up an accessible entry point, publish a handling process and response time, and return results to complainants. Article 19 covers inspections and says providers supply explanations of training-data sources, scale and types, labelling rules and algorithm mechanisms.

Article 21 sets out consequences for breaches, drawing on the Cybersecurity Law, Data Security Law, Personal Information Protection Law and Science and Technology Progress Law. Where those laws contain no specific provision, the measure lists warnings, public criticism and an order to rectify within a time limit, with suspension of the service for refusal to correct or serious circumstances.

What followed

The measure created a vocabulary that later documents reuse: provider, user, public-facing service, algorithm filing, data labelling. The 2025 labeling measures explicitly build on it alongside the algorithmic-recommendation and deep-synthesis provisions.

For companies, the practical result has been a set of engineering and documentation tasks — filing records, data-source descriptions, labelling functions in export paths — rather than a single approval event.

Honest limitations

The account here follows the full text published by the CAC and mirrored on the government website, plus the April 2023 draft notice. I have summarised article functions rather than reproduced the text, and I have deliberately omitted sections of the measure that are not relevant to product and data compliance. The description of later labelling rules is drawn from the CAC announcement for those measures.

What readers can do now

  • Cite the published Order No. 15 text, not the April 2023 draft, when referring to current requirements.
  • Check whether a service you use has an algorithm-filing record; the number is a public registration, not a claim on a landing page.
  • When evaluating a Chinese generative product, separate three things: model capability, data documentation and the filing record — they are different questions.

Related coverage

Cite this article

NeuroAI (脑机网), "China's Generative AI Measures Tie Public Launch to Filing and Safety Review," 22 November 2026, https://neuroai.site/a/na-policy-genai-measures-filing

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Policy & Governance” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories