Skip to content

Interim Measures for the Management of Generative AI Services

📑 Legal hierarchy: Level 3 · Departmental rule | Issuance: CAC, NDRC, MOE, MOST, MIIT, MPS, NRTA (seven-ministry joint issuance) | Effective: 2023-08-15 | Character: hard law

⚠️ Hierarchy note: The name “Interim Measures” (暂行办法) does not indicate a State Council “interim regulation.” This instrument is a departmental rule, jointly issued by seven State Council ministries; it is not a State Council administrative regulation, and still less a NPC law. “Interim” (暂行) reflects the legislator’s decision to leave space for policy adjustment under uncertainty about generative-AI governance patterns — it is not a hierarchy marker. See Index of Chinese Rules.

English Summary

Section titled “English Summary”

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 jointly issued by the Cyberspace Administration of China (CAC) and six other ministries on 2023-07-13, effective 2023-08-15. Widely characterized by regulators as an “inclusive and prudent” (包容审慎) approach, the final version was substantially softer than the April 2023 draft: it narrowed the scope to services provided to the public within mainland China (excluding enterprise-internal use and R&D), dropped some of the draft’s most stringent training-data guarantees, and embraced classified and graded supervision (分类分级). Key obligations include security assessment and algorithm filing for services with “public opinion or social mobilization capacity,” baseline training-data legality duties, and content labeling via the Deep Synthesis Provisions and the Labeling Measures. The Measures sit at the top of China’s AI-specific regulatory stack alongside the 2023 Deep Synthesis Provisions.

Overview

Section titled “Overview”

The Generative AI Interim Measures are China’s first departmental rule dedicated to generative AI services (no longer limited to “deep synthesis” as a technical subset). Compared with the Deep Synthesis Provisions, they:

  • Shift the regulatory object from technology to service: the regulated object is the generative AI product offered to the public.
  • Extend obligations upstream: not only service providers but also (in part) the R&D side (training data, foundation models) fall within scope.
  • Introduce “classified and graded supervision”: this leaves an interface for differentiated, risk- or scenario-based regulation (although the detailed framework remains unpublished).

Between the April 2023 draft for comment and the July 2023 final text, duties were noticeably softened: the absolute guarantee that training data be “true and accurate” was replaced with “take effective measures,” scope was narrowed to public-facing services, and enterprise-internal R&D and use were expressly exempted for the first time. The shift was widely read as Beijing’s recognition that over-compliance would choke industrial development.

Scope

Section titled “Scope”

Applies to:

  • Generative AI services provided to the public within the People’s Republic of China;
  • Generation of text, images, audio, and video.

Does not apply (express carve-out):

  • Industry associations, enterprises, educational and research institutions, public cultural institutions, and relevant professional bodies that develop or apply generative AI technologies without providing generative AI services to the domestic public.

This carve-out — the single largest change from the April 2023 draft to the July final text — creates space for enterprise R&D.

Core Duties

Section titled “Core Duties”

On service providers

Section titled “On service providers”
  • Training-data compliance (Article 7):
    • use data and foundation models with lawful sources;
    • where intellectual property is implicated, shall not infringe others’ IP rights;
    • where personal information is implicated, shall obtain individual consent or satisfy another lawful basis under PIPL;
    • take effective measures to improve training-data quality, enhancing authenticity, accuracy, objectivity, and diversity.
  • Algorithm filing (Article 17): where the service has public-opinion attributes or social-mobilization capacity, file under the Algorithm Recommendation Provisions.
  • Security assessment (Article 17): under the same conditions, conduct a security assessment as prescribed.
  • Content labeling: fulfil under the Deep Synthesis Provisions and the Labeling Measures.
  • Content moderation:
    • take remedial measures against unlawful content in a timely manner: halt generation, halt transmission, delete;
    • take measures such as further model optimization to rectify, and report to the competent authority;
    • maintain sound complaint and reporting channels.
  • User management:
    • specify and publish the service’s target users, contexts, and uses;
    • guide users toward rational and scientific use;
    • discharge personal-information protection duties in accordance with law and contract.

On users

Section titled “On users”
  • Shall not generate unlawful content — including content endangering national security, undermining ethnic unity, or promoting violence or obscenity.

Enforcement and Penalties

Section titled “Enforcement and Penalties”
  • Competent authorities: CAC leads; the other six ministries coordinate within their mandates.
  • Penalty basis: the existing statutory ladder; no standalone penalty provisions.
  • Classified and graded supervision: the Measures leave an interface for detailed implementing rules (no comprehensive framework yet released).

Relationship with Other Rules

Section titled “Relationship with Other Rules”
  • Deep Synthesis Provisions (2023): both upstream and parallel. The Deep Synthesis Provisions regulate via technology; the Interim Measures regulate via service; the overlap is primarily in conspicuous labeling and algorithmic filing.
  • Algorithm Recommendation Provisions (2022): the Interim Measures incorporate its filing procedure — forming a “one filing system, many rules” structure. See algorithm-recommendation-provisions.
  • Labeling Measures (2025): the technical implementing rules for the labeling duty. See biaozhi-banfa.
  • Personal Information Protection Law (2021): applies jointly where training data processing touches personal information.
  • TC260 series of standards: the supporting security technical specifications (training-data security, service-level security assessment, etc.).

Debates and Commentary

Section titled “Debates and Commentary”
  1. The hollow state of “classified and graded supervision”: the Measures set this out as a principle (Article 3), but as of late 2024 no detailed classification-and-grading rules had been released. In practice compliance remains one-size-fits-all.
  2. Execution gap on training-data IP: the text requires “no infringement,” but leaves “lawful source” undefined. In practice, leading firms rely on a mixed model of fair use + licensed content + in-house datasets.
  3. Boundaries of the enterprise-internal carve-out: does an enterprise API offered externally count as “public-facing”? What scale counts as “public”? In practice, firms use the presence of an independent consumer product as the dividing line.
  4. Interface with local pilots: AI-innovation pilots in Beijing, Shanghai, and Shenzhen may adopt looser compliance postures; how undisclosed “sandbox” mechanisms dovetail with the Measures remains a gray zone.

Source Text and Translations

Section titled “Source Text and Translations”
LanguageSourceLink
Chinese (original)CACcac.gov.cn
Chinese (archived copy)This sitegenerative-ai-interim-measures-2023-07-13.html
EnglishChina Law Translatechinalawtranslate.com/generative-ai-interim
English (structured)Regulations.AIregulations.ai/…/china-2023-7-generative-ai
English (academic version)China Aerospace Studies Institute (CASI)airuniversity.af.edu 2023-08-07

Version History

Section titled “Version History”
DateEvent
2023-04-11Measures for the Management of Generative AI Services (Draft for Comment) released
2023-07-13Final text released (name amended to include “Interim”; duties markedly softened)
2023-08-15Effective
2026-04-21First archived on this site

Cited Sources

Section titled “Cited Sources”

Primary

Section titled “Primary”

Major commentary

Section titled “Major commentary”
Section titled “Related research”
  • Matt Sheehan (Carnegie Endowment) — “Tracing the Roots of China’s AI Regulations” (2023-07). Analysis of the evolution from draft to final text; the most systematic first-hand English-language work.
  • Paul Triolo (DigiChina / Albright Stonebridge) — institutional analysis of multi-ministry joint rulemaking.
  • Helen Toner / Jeffrey Ding — comparative political-economy readings placing the Interim Measures within the “agile governance” frame.
  • Zhang Linghan 张凌寒 (China University of Political Science and Law / CUPL) — legal-theoretic critique of the hierarchy of departmental rules and the draftsmanship of “Interim” (暂行) legislation.
  • CAIDP (Center for AI and Digital Policy) — quarterly reports assessing the international influence of the Interim Measures.

Cite this page (generated 2026-04-21):

Comparative AI. Commentary on the Interim Measures for the Management of Generative AI Services. Accessed YYYY-MM-DD. https://comparativeai.org/rules/china/generative-ai-interim-measures/