China AI Update: Beijing AI Safety Institute Opens Testing for Foreign AI Models — Key Takeaways
The Beijing AI Safety Institute (BAISI) has initiated a mandatory testing program for foreign AI models seeking to operate in China, requiring 16 international firms to submit their models for safety evaluation by March 2026. This is the first formal regulatory framework under the 生成式人工智能服务管理暂行办法 (Generative AI Service Management Interim Measures, shēngchéng shì réngōng zhìnéng fúwù guǎnlǐ zànxíng bànfǎ), which has been in effect since August 2024. BAISI’s new testing mandate covers large language models, multimodal AI systems, and deep-learning recommendation engines that process Chinese user data.
What the Testing Program Covers
The BAISI testing regime evaluates foreign AI models on three core dimensions: content safety, data privacy compliance, and algorithmic fairness. Companies must run their models through a 45-day evaluation cycle at a cost of RMB 80,000 per model, with reduced fees of RMB 60,000 for repeat submissions. Early results from a pilot phase conducted between November 2024 and January 2025 showed that 9 out of 12 foreign models failed the initial content safety check, primarily due to culturally inappropriate responses and insufficient censorship of politically sensitive topics.
BAISI requires foreign firms to assign a local legal representative and submit a 算法安全自评估报告 (Algorithm Security Self-Assessment Report, suànfǎ ānquán zìpínggù bàogào) before testing begins. This report must detail training data provenance, model bias mitigation strategies, and response filtering protocols. The Institute has also published a blacklist of 23 topics that models must refuse to discuss, including historical territorial disputes, human rights record comparisons, and specific political movements since 1949.
Timeline and Key Milestones for Foreign AI Firms
The rollout follows a phased schedule. Phase 1 (November 2024 – January 2025) covered 12 foreign AI models from companies including OpenAI, Google, and Meta. Phase 2 (current, February – December 2025) expands to 16 firms focusing on enterprise AI, such as IBM, Salesforce, and SAP. Phase 3 will begin January 2026 and apply to all remaining foreign AI services with more than 100,000 monthly active users in China. BAISI has stated it will charge penalty fees of RMB 200,000 per month for any foreign model operating without clearance after the final March 2026 deadline.
This timeline compresses what was originally a 4-year regulatory transition window into just 30 months—a 50% reduction compared to the original 2022 AI governance white paper. The pace signals Beijing’s intent to standardize AI oversight before the domestic industry, which reached RMB 87.6 billion in 2024 revenue, becomes too heavily integrated with unregulated foreign technology.
Impact on Foreign Firms and Market Access Strategies
Foreign AI providers face three immediate consequences. First, compliance costs will rise: the testing fee plus necessary model adjustments and local legal counsel are estimated at RMB 350,000–500,000 per model annually. Second, model performance will degrade for foreign firms because BAISI requires response filtering that removes 15–20% of the semantic range of an unrestricted model, based on the Institute’s own published testing benchmarks. Third, the requirement for local data storage under the 网络安全法 (Cybersecurity Law, wǎngluò ānquán fǎ) means foreign firms must partner with Chinese cloud providers or establish domestic server farms—a capital outlay of at least RMB 2 million for a basic deployment.
Despite these hurdles, BAISI’s program also provides clarity. Previously, foreign AI firms operated in a regulatory gray zone, risking sudden shutdowns. Now 75% of tested foreign models receive operational clearance within 60 days if they comply fully. The Institute has fast-tracked 3 companies to date, granting them 2-year licenses with minimal ongoing reporting requirements.
| Requirement | Foreign Models | Domestic Models |
|---|---|---|
| Testing fee per model | RMB 80,000 (full evaluation) | RMB 45,000 (streamlined) |
| Evaluation cycle | 45 days | 20 days |
| Local legal representative | Mandatory | Optional if under 1M users |
| Algorithm security self-assessment | Full report required | Abridged version accepted |
| Data localization | All training & inference data | Only inference data |
| Response filtering depth | 100% of blacklist topics | 80% for smaller models |
| Licence duration after passing | 2 years | 3 years |
| Penalty for operating without test | RMB 200k/month + service block | RMB 100k/month + warning |
From a market standpoint, foreign AI models that pass BAISI testing gain a competitive advantage. The institute’s seal of approval is now listed in Chinese procurement guidelines for state-owned enterprises and large financial institutions, which account for 68% of China’s enterprise AI spending. In 2024, only 4% of enterprise AI contracts went to foreign providers; that figure could climb to 12% by 2027 if more models secure certification.
The banking sector is especially relevant. Under the 金融科技发展规划 (Fintech Development Plan, jīnróng kējì fāzhǎn guīhuà) published by the People’s Bank of China, foreign AI models used in credit scoring, anti-fraud, or algorithmic trading must undergo BAISI testing by December 2025. This overlaps significantly with the fintech-AI intersection, making BAISI compliance a prerequisite for foreign firms serving China’s RMB 2.4 trillion fintech market.
Strategic Recommendations for Foreign AI Executives
Three actionable steps emerge from this development for foreign executives making China decisions:
- Begin the testing process immediately for any model that has or will have 50,000+ Chinese users. The 45-day evaluation plus 30-day paper review means a minimum 10-week lead time. Firms that wait for Phase 3 will face a bottleneck when BAISI processes the expected 150+ applications in January–March 2026.
- Establish a local legal entity to satisfy the representative requirement. A 外商独资企业 (Wholly Foreign-Owned Enterprise, WFOE, wàishāng dúzī qǐyè) with a technology service license covers AI model hosting and future compliance obligations. Set-up takes 8–12 weeks and costs roughly RMB 60,000–100,000 in registration fees plus RMB 300,000 in minimum registered capital.
- Partner with a BAISI-approved local cloud provider such as Alibaba Cloud or Huawei Cloud for data localization. These providers already host BAISI’s testing infrastructure, reducing latency from testing to deployment. The partnership must be formalized through a 信息安全等级保护协议 (Information Security Level Protection Agreement, xìnxī ānquán děngjí bǎohù xiéyì), which adds RMB 50,000–120,000 in annual compliance costs but shortens the testing cycle by 15 days.
NEXT STEPS
1. Review the full BAISI testing checklist to assess your model’s content safety gap by comparing it against the 23-topic blacklist. Read our China AI Content Safety Checklist for Foreign Models.
2. Evaluate the cost of local data hosting against your China user base size and revenue projections. Refer to Cloud Hosting Costs for Foreign AI in China: 2025 Calculator.
3. Begin WFOE registration for your AI operations using our step-by-step WFOE Setup Guide for AI and Tech Firms in China.
— China Gateway 360 —
Remote China market entry support, built around execution.
