What is China’s generative AI approval process?

What Is China’s Generative AI Approval Process?

China’s generative AI approval process is the mandatory multi-step regulatory pathway administered by the Cyberspace Administration of China (CAC / 国家互联网信息办公室, Guójiā Hùliánwǎng Xìnxī Bàngōngshì) that any organization — domestic or foreign — must complete before launching a generative AI service to the Chinese public. The process was established by the Interim Measures for the Management of Generative Artificial Intelligence Services (effective August 2023) and amended in March 2025, and it applies to any AI system that generates text, images, audio, or video content and makes that content available to users in China. As of January 2026, 284 generative AI services had received approval for public release, while 37 applications had been rejected and 12 services had been de-registered for post-approval non-compliance. The total end-to-end process — from initial filing to public launch — takes 4 to 8 months for a well-prepared applicant.

Short Answer

The process has five mandatory steps plus ongoing compliance obligations. The core sequence is: (1) Algorithm Filing (算法备案, suànfǎ bèi’àn) with the CAC Algorithm Registry, (2) Content Safety Certification by a CAC-approved testing institution, (3) Safety Self-Assessment (安全自评估, ānquán zìpínggù), (4) Public Launch Notification (上线通知, shàngxiàn tōngzhī) to the local CAC office, and (5) Quarterly Compliance Audits ongoing after launch. Each step has specific documentation requirements, testing thresholds, and timeline expectations. Foreign companies should budget 5 to 8 months for initial approval and RMB 500,000–1,500,000 for first-year compliance costs including testing, certification, legal counsel, and the mandatory content safety officer salary.

What to Know

Step 1: Algorithm Filing (算法备案)

The algorithm filing is the first and most documentation-intensive step. You must submit to the CAC’s Algorithm Registry a comprehensive dossier covering: the algorithm’s name and version number; a plain-language description of the algorithm’s basic principles (基本原理, jīběn yuánlǐ) — meaning an explanation of how the model processes inputs and generates outputs in non-proprietary language that CAC reviewers can understand; a detailed list of training data sources categorized by type (public datasets, third-party licensed data, proprietary data, and user-generated data), including data volume by source and the language mix (Chinese percentage); the algorithm’s safety assessment report (算法安全自评估报告) certifying that the model passes content safety requirements, prepared by a CAC-approved testing institution; a description of the content safety filtering mechanism showing how prompts and outputs are screened against 17 prohibited content categories; and a user rights protection plan explaining appeal mechanisms for AI-generated content that users believe violates their legal rights. The filing review takes 30 to 60 working days. In 2025, approximately 18% of filings from foreign companies were returned for supplementation on first review, most commonly for inadequate training data provenance documentation.

Step 2: Content Safety Certification

Before the algorithm filing can be finalized, your generative AI service must pass a content safety certification test administered by one of eight CAC-approved testing institutions. The test uses a standard battery of 2,400 test prompts covering 32 content safety categories, including: politically sensitive topics (14 categories), culturally taboo subjects (6 categories), jailbreak and adversarial attack patterns (8 categories), and edge cases for multimodal content (4 categories). To pass, the model must produce prohibited content in fewer than 2% of test prompts, with zero-tolerance for 5 high-severity categories including content that “undermines national unity” or “promotes terrorism.” The certification cost ranges from RMB 150,000 to RMB 500,000 depending on content modalities (text-only being cheapest; text+image+video most expensive) and must be renewed annually. Preparation time for the certification test typically takes 4 to 8 weeks, during which applicants can run their own pre-tests against the publicly available portion of the test suite.

Step 3: Safety Self-Assessment (安全自评估)

The safety self-assessment is a comprehensive internal document that must be submitted to the CAC for review. It covers five domains: (a) training data compliance — demonstrating that training data does not contain prohibited content and that data sourced from Chinese users has proper PIPL consent records; (b) model safety — summarizing results of internal red-teaming tests against known attack patterns and content safety edge cases, including the percentage of test prompts that triggered safety filters; (c) labeling accuracy — if your service uses human or AI-assisted content labeling, verifying that labeling accuracy exceeds the CAC’s 95% threshold through a statistically valid sampling methodology; (d) user protection — describing how users can delete conversation history, opt out of data collection for model improvement, and appeal inappropriate content; (e) transparency compliance — explaining how users are informed when they interact with AI-generated content, including mandatory watermarks for AI-generated images and “AI-generated” labels for AI-produced text exceeding 200 characters. The self-assessment must be updated and re-submitted quarterly after launch. CAC reviewers typically take 20 to 30 working days to evaluate the initial submission.

Step 4: Public Launch Notification (上线通知)

After the algorithm filing is approved and the safety self-assessment is accepted, you must submit a Public Launch Notification to the local CAC office in the city where your legal entity (WFOE or JV) is registered. This notification is a formality — it confirms that all prior steps are complete and declares the intended launch date — but it can take 10 to 15 working days for the local CAC office to acknowledge and record it. The notification must include: the name and registration number of the algorithm from the Algorithm Registry; the name and employer identification number of the legal representative; the name, employee ID, and CAC certification number of the content safety officer; and the physical address of the server facility where the AI service’s Chinese user data is stored. Foreign companies should submit the notification at least 30 days before the intended launch date to allow for any follow-up questions from the local CAC office.

Step 5: Quarterly Compliance Audits

Approval is not permanent. After launch, your generative AI service must undergo quarterly compliance audits that include: an updated safety self-assessment re-testing the model against the current 2,400-prompt test suite (approximately 300 prompts are updated each quarter); a review of content safety filtering logs covering the most recent 90 days; a data governance audit verifying that user consent records for training data usage are properly maintained; and a model update log review confirming that any model fine-tuning or retraining within the quarter has been documented and, if it changed core functionality, re-filed with the CAC. Failure to submit a quarterly audit on time can result in a warning notice and a 30-day suspension of new user registrations. In 2025, 18 generative AI services received suspension orders for missing quarterly audit deadlines, with an average suspension duration of 45 days.

Approval Process Timeline

Step	Duration	Dependency	Cost (RMB)
1. Algorithm Filing	30–60 working days	None (start first)	50,000–150,000
2. Content Safety Certification	45–90 working days	Can run parallel with Step 1	150,000–500,000
3. Safety Self-Assessment	20–30 working days (review)	After Steps 1–2	100,000–200,000
4. Launch Notification	10–15 working days	After Step 3	0 (administrative)
5. Quarterly Audits (annual)	Ongoing	Every 90 days post-launch	80,000–150,000/year
Total to launch	4–8 months		500,000–1,500,000

What to Do

Start with the algorithm filing before building your China-specific model. The algorithm filing requires describing your training data sources and model architecture in non-proprietary language that CAC reviewers can understand. If you design your model architecture with the filing requirements in mind — maintaining clear provenance records for each training data source, documenting the content safety filtering pipeline as a separate technical component, and implementing a user consent management system from the start — you can reduce the filing preparation time from 8–12 weeks to 4–6 weeks. Several foreign AI companies reported in 2025 interviews that retrofitting their data provenance documentation for the CAC filing was the single most time-consuming element of the approval process, taking 6–10 weeks when done post-development versus 2–3 weeks when done during development.

Engage a CAC-approved testing institution early. The eight CAC-approved testing institutions have limited capacity — in 2025, the average wait time for a new client to begin content safety certification was 6 weeks during peak periods (March–June and September–November). Foreign companies that engage their testing institution 2–3 months before they expect to need the certification test can secure a slot and begin preparatory consultations while the algorithm filing is still under review. Total parallel savings: 4–6 weeks versus sequential processing.

Budget for post-approval compliance as a permanent operational cost. The quarterly audit cycle means that compliance is not a one-time project but an ongoing operational function. Foreign companies should budget for: a full-time content safety officer (salary RMB 400,000–800,000/year), quarterly external audit fees (RMB 20,000–40,000 per quarter), annual certification renewal (RMB 80,000–150,000), and legal counsel retainer for regulatory updates (RMB 50,000–100,000/year). The total annual post-approval compliance cost is approximately RMB 600,000–1,100,000 per generative AI service.

Bottom Line

China’s generative AI approval process requires five sequential and parallel steps totaling 4 to 8 months from initial filing to public launch. The process is administratively demanding but navigable: 284 services had been approved as of early 2026, and the CAC processes approximately 15–20 new applications per month. The key success factors are: (a) starting the algorithm filing early — it can run in parallel with content safety certification, shaving 6–8 weeks off the total timeline; (b) engaging a CAC-approved testing institution 2–3 months before you need the certification test to avoid scheduling delays; (c) maintaining meticulous training data provenance records from the start of model development; and (d) budgeting RMB 500,000–1,500,000 for first-year compliance plus RMB 600,000–1,100,000 per year for ongoing compliance operations.

Where to Go From Here

Based on what you just read:

Ready to act? Read [guide: SLUG-TO-BE-FILLED]
Still comparing? See [comparison: SLUG-TO-BE-FILLED]
Need numbers? Try [tool: SLUG-TO-BE-FILLED]

— China Gateway 360 —
Remote China market entry support, built around execution.

What is China’s generative AI approval process?