Last reviewed: April 9, 2026
Jurisdictions covered: China (primary), EU (comparison)
Reading time: 15 minutes
Language note: Regulation text verified against the original Chinese on cac.gov.cn (read March 21, 2026). Article numbers reference the official Mandarin text.
China’s Generative AI Measures: What Providers Must Do
Baidu’s Ernie Bot launched to the Chinese public on August 31, 2023 — sixteen days after the Interim Measures for the Management of Generative AI Services took effect. Baidu was among the first batch of companies to complete the CAC’s mandatory security assessment and algorithm filing. Without that filing, the launch would have been illegal.
Two and a half years later, 796 generative AI services have filed at the national level and 481 more at the local level (CAC, March 17, 2026). Every one of them passed through the same gate Baidu did. Despite the “interim” label, these measures are the world’s first dedicated generative AI regulation and they are actively enforced — the February 2026 content labeling sweep penalized 13,421 accounts in a single action.
This article explains what the measures require, how they interact with China’s other AI rules, and what providers must do to comply. For the full regulatory picture, see our China AI Regulation 2026 guide.
Key Takeaways
- Every public-facing generative AI service in China must file with the CAC before launch. This is a binary gate: no filing means no legal operation. 796 services have filed nationally as of February 2026.
- Five training data requirements apply to all providers: legal sourcing, no IP infringement, personal data consent, quality assurance (truthfulness, accuracy, objectivity, diversity), and compliance with CSL/DSL/PIPL.
- Content must reflect “socialist core values” — this is not optional guidance, it is a legal requirement with 14 prohibited content subcategories.
- R&D is exempt. Art. 2 explicitly excludes organizations using generative AI for internal research, education, or non-public purposes.
- API providers are covered. Art. 22 explicitly includes providers offering services via programmable interfaces.
The Filing Process: Pre-Launch Gate
Before any generative AI service can go live, it must complete the CAC filing process. This is the most consequential compliance obligation — everything else follows from it.
The process:
1. Provider self-assessment. The company evaluates its service against the measures’ requirements.
2. Provincial CAC review. The local CAC office reviews documentation and conducts a security assessment.
3. Central CAC review. For services with national reach or “public opinion influence,” the central CAC reviews and may consult MIIT and MPS on national security matters.
4. Filing number issued. Once approved, the service receives a filing number that must be displayed on the product page.
Timeline: 2-5+ months from submission to approval. Services using pre-approved third-party models (API-based) go through a shorter local registration process (~2-3 months).
Filing statistics (as of February 28, 2026):
| Level | Filings | Growth |
|---|---|---|
| National (full filing) | 796 | +48 in Jan-Feb 2026 |
| Local (API-based registration) | 481 | +46 in Jan-Feb 2026 |
| Total | 1,277 | ~38 new national filings/month (mid-2025) |
No publicly disclosed cases of applications being formally denied exist. The system operates as a gate: services that do not meet requirements simply do not receive approval. The absence of public denials likely reflects pre-filing consultation rather than a rubber-stamp process.
It is technically illegal for a public-facing AI service in China to use APIs from overseas developers (Oxford China Policy Lab, January 2026). This means foreign model providers cannot serve the Chinese market directly through API — they need a domestic partner or a local entity.
Content Requirements: Socialist Core Values and 14 Prohibitions
This is where the GenAI Measures diverge most from Western AI regulation. Art. 4 establishes content compliance requirements that have no equivalent in the EU AI Act or any other jurisdiction.
Generative AI services must comply with five content categories:
Category 1: Socialist core values alignment. Content must not undermine state authority. 14 subcategories of prohibited content include: subversion of state power, separatism, terrorism advocacy, ethnic hatred, violence, pornography, and false information.
Category 2: Anti-discrimination. Content must not discriminate based on ethnicity, belief, nationality, region, gender, age, or occupation.
Category 3: IP and commercial rights. Output must respect intellectual property and commercial reputation.
Category 4: Personal rights. Output must not infringe on others’ image, reputation, privacy, or other personal rights.
Category 5: Transparency and accuracy. Providers must take measures to improve training data quality and accuracy of generated content.
Providers must implement a content review system that catches prohibited content before it reaches users. The February 2026 enforcement sweep — 13,421 accounts penalized for AI content labeling violations — demonstrates that these are not paper requirements.
Our view: The content requirements create the most significant dual-compliance challenge for companies operating in both China and the EU. A generative AI model trained and filtered for Chinese content requirements cannot simultaneously serve EU users without a separate content governance pipeline. This is an architectural requirement, not a policy preference. For practical guidance on managing this, see the dual-compliance section in our China vs EU comparison.
China vs EU: Generative AI Rules Compared
| Dimension | China (Interim Measures) | EU (AI Act Chapter V + Code of Practice) |
|---|---|---|
| Effective | August 15, 2023 | GPAI obligations: August 2, 2025 |
| Pre-market gate | CAC filing mandatory (796 filed) | No pre-market filing. Conformity self-assessment for high-risk downstream uses |
| Content control | 14 prohibited categories + socialist core values | No state content requirements |
| Training data | 5 explicit requirements (Art. 7): legal, no IP violation, personal data consent, quality, law compliance | Art. 53: training data summary, copyright policy. Less prescriptive |
| Annotation workers | Regulated (Art. 8): clear rules, quality assessment, sample verification | Not specifically regulated |
| Labeling | Mandatory visible + invisible per GB 45438-2025. Technical specs published | Art. 50: “clearly and distinguishably” mark. No technical spec published |
| R&D exempt? | Yes (Art. 2) | Yes (Art. 2) |
| API providers | Explicitly covered (Art. 22) | Covered under provider/deployer framework |
| Compute threshold | Drafted but not published | 10^25 FLOP for systemic risk classification |
| Enforcement | 13,421 accounts penalized (Feb 2026); filing gate prevents non-compliant launch | Zero enforcement actions (Apr 2026) |
| Code of practice | No equivalent (regulation is binding from day one) | GPAI Code of Practice (voluntary, published July 2025) |
Disclaimer: This content is for informational purposes only and does not constitute legal advice. Chinese generative AI regulation is enforced by multiple agencies and changes frequently. Organizations should engage qualified legal counsel with China-specific expertise. Reg Intel is not a law firm and does not provide legal services.
Last verified: April 9, 2026
Compare: EU vs China
For the global keystone comparison across twelve dimensions — algorithm filing vs conformity assessment, content moderation conflicts, asymmetric extraterritoriality, enforcement philosophy, and a five-step dual-market compliance baseline — see EU vs China AI Regulation: Two Systems, Two Philosophies (2026).