A chatbot that writes, draws or voices anything on request is now a regulated service in China, not just a piece of software. The baseline rulebook arrived as seven ministries lined up behind a single document promising both encouragement and guardrails. For builders of large models (大模型), the trade is clear: innovate fast, but account for your data and your outputs.
The document at a glance
The Interim Measures for the Management of Generative Artificial Intelligence Services (生成式人工智能服务管理暂行办法) were issued by seven departments — the CAC, the National Development and Reform Commission, the Ministry of Education, the Ministry of Science and Technology, MIIT, the Ministry of Public Security and the National Radio and Television Administration — as Order No. 15, and took effect on 15 August 2023.
They apply to services that provide generated text, images, audio or video to the public inside China. Crucially, internal research and development that is not offered to the public sits outside the scope, which keeps labs and academia freer to experiment.
Development and security, weighed together
The measures open with a guiding principle: development and security are equally important, innovation and governance go together, and services get "inclusive, prudent, classified and graded" supervision (分类分级监管). Rather than a blunt license, the state says it will encourage application across industries, support self-innovation in algorithms, frameworks and chips, and push for public training-data platforms and shared compute.
What providers must handle in training data
Articles 7 and 8 put real obligations on how models are built:
- Use data and base models from lawful sources.
- Respect intellectual property and trade secrets.
- Where personal information is involved, obtain consent or meet another legal basis.
- Improve training-data quality — truthfulness, accuracy, objectivity and diversity.
- Write clear, operable annotation rules, assess annotation quality by sampling, and train annotators.
Service-side duties
Once a service is live, providers wear the hat of both content producer and personal-information processor. They must publish who the service suits and warn against minor over-reliance or addiction (Article 10), protect user inputs and records without hoarding unnecessary data (Article 11), and label generated images and videos under the Deep Synthesis Provisions (Article 12). They also must run complaint channels and act fast on illegal content — stopping generation, retraining, and reporting.
Security assessment and algorithm filing
Mirroring the deep-synthesis and algorithm-recommendation rules, services with "opinion attributes or social-mobilization capability" must complete a security assessment and file their algorithm (Article 17). Regulators can demand disclosure of training-data sources, scale, types, labeling rules and algorithm mechanisms during inspections (Article 19). Providers that fail can be warned, ordered to correct, or — if serious — have service suspended.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the measures read as a permission-to-innovate framework where the cheapest compliance path is clean data and honest labeling from the start.
Why seven ministries, and why "classified and graded"
The seven-department authorship is not ceremony. Generative AI touches content (CAC, NRTA), industry (MIIT), education and science (their ministries), and macroeconomic planning (NDRC) at once, so a single-agency rule would have left gaps. The "classified and graded" (分类分级) principle is the mechanism that keeps the regime from being a blunt instrument: an internal enterprise copilot and a public-facing chatbot carry different duties, and a model with opinion-shaping reach is treated more strictly than a narrow tool. The practical effect is that compliance cost scales with risk and audience, which is what lets the state promise encouragement while still drawing hard lines.
What the internal-R&D carve-out protects
Excluding non-public research and development is one of the most consequential choices in the text. It means a lab can train, evaluate, and even open-weight an experimental model without triggering the full service regime, as long as it is not offered to the public. That buffer is what keeps Chinese frontier labs able to compete in open research and benchmark against global models without clearing a security assessment for every experiment. The boundary is the moment of public release, which is why the duties concentrate on the service, not the science.
How it stacks with the other rules
The measures are deliberately layered, not standalone. They point back to the Deep Synthesis Provisions for labeling and to the Algorithm Recommendation Provisions for filing, so a generative service that also edits faces inherits both rulebooks. For builders this means compliance is cumulative: clean training data and honest labeling are the cheapest path because they satisfy the most overlapping obligations at once. Read against overseas regimes such as the EU AI Act, China's framework is earlier, more sector-blended, and more focused on who is accountable for outputs than on model-capability thresholds. The open question is whether the innovation incentives survive the gradual tightening of graded rules.
Honest limitations
This article summarizes the enacted 2023 Interim Measures and is not legal advice. "Classified and graded" implementation details and sector-specific rules are still being rolled out by individual authorities, and enforcement tone shifts over time. The measures interact with the Deep Synthesis Provisions and the Algorithm Recommendation Provisions, so they should be read as part of a stack rather than in isolation. Foreign-invested providers must also meet applicable foreign-investment rules.
What readers can do now
- If you plan to launch a public generative-AI service, bake data-lawfulness and annotation governance into the training pipeline before any release.
- If you use such services, know your inputs may be processed as personal information and use the complaint channels if outputs are unlawful.
- If you follow the field, track the sector-specific classified-and-graded rules, since they define the real day-to-day obligations.