China Generative AI Registration (备案 / bei'an) is the regulatory regime governing market access for generative AI services offered to the Chinese public. Any firm offering such services must complete formal registration with the Cyberspace Administration of China (CAC) (CAC). The Incentive Architecture Export Controls Cannot Reach (Sun, Lawfare, May 6 2026) describes the regime as the structural anchor of the Chinese AI incentive architecture documented in PEAT — Proactive Elite Alignment Theory.
Regulatory basis
The regime rests on two instruments. The 2023 Interim Measures for the Management of Generative AI Services, issued by the CAC together with six ministries, established the formal registration (备案 / bei'an) requirement (Source: http://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm). The revised Cybersecurity Law, effective January 1, 2026, for the first time incorporated AI governance provisions into China's statutory legal framework, expanding the regulatory backbone of the registration regime and giving AI governance violations statutory backing (Source: http://www.npc.gov.cn/npc/c1773/c1848/c21114/wlaqfxz/wlaqfxz002/202511/t20251103_449242.html).
Scale
As of end-2025, 748 generative AI services had completed formal registration (备案 / bei'an) with the CAC, and a separate 435 generative AI applications or features drawing on registered models had completed a distinct filing process (登记 / dengji) (Source: https://www.cac.gov.cn/2026-01/09/c_1769688009588554.htm).
Registration requirements
Applicants must submit a security self-assessment report exceeding 100 pages covering training data provenance, content safety benchmarks, and emergency response protocols. The required documentation includes training data provenance records; content safety benchmarking against 31 enumerated safety risks across five categories, per CSET analysis (Source: https://cset.georgetown.edu/publication/china-safety-requirements-for-generative-ai-final/); and emergency response protocols governing how the model behaves when it generates noncompliant output.
Enforcement
Models that fail or skip registration face removal from app stores, fines, and operational suspension. Under the revised Cybersecurity Law, AI governance violations have statutory backing (Source: http://www.npc.gov.cn/npc/c1773/c1848/c21114/wlaqfxz/wlaqfxz002/202511/t20251103_449242.html).
Sun's analysis: a survival condition, not a tier of privilege
The Incentive Architecture Export Controls Cannot Reach (Sun, Lawfare, May 6 2026) argues that the registration system is not a benefit-conferring mechanism that rewards the compliant but a filter that eliminates everyone else: unregistered models cannot legally operate in the Chinese market, so the firms inside the gate are the only firms that exist. In this account the filter is upstream of the subsidy structure — every firm that participates in the compute voucher programs, competes for government procurement, or builds applications on domestic AI infrastructure has already passed through the registration gate and demonstrated compliance with state content and security standards.
Sun draws several consequences from this framing. Because every Chinese AI firm operating at scale has passed through the gate, all are structurally aligned with state content and security standards. US sanctions targeting individual firms for "bad behavior" miss the regulatory structure, since every firm has already passed through alignment-conditioning. And export controls, per the PEAT — Proactive Elite Alignment Theory feedback loop, deepen rather than disrupt the system.
Comparison with other Chinese market-access regimes
Market-access registration is not new to AI in China. Authorities have imposed licensing, filing, and approval regimes across technology sectors for decades, including telecommunications and pharmaceuticals, and content regulation of films, television, games, and publications is similarly long-established. Sun argues that what distinguishes the generative AI registration system is that its compliance requirements reach into production inputs, not just finished outputs. Films must clear review before screening and manuscripts before publication, but in those cases content is the inspection target; the generative AI filing regime instead requires applicants to document training data provenance, pass content safety benchmarks, and file emergency response protocols, so that the inspection extends to the training corpus, the training methodology, and the post-training safeguards. On this reading the generative AI gate filters for alignment at a layer that other industry gates, including content regulation, do not reach.
Relation to other concepts and policy
The Incentive Architecture Export Controls Cannot Reach (Sun, Lawfare, May 6 2026) treats the registration system as the structural anchor of the Chinese AI incentive architecture set out in PEAT — Proactive Elite Alignment Theory and as the operative regulatory backbone behind Chinese AI Policy. Registration also functions as the structural prerequisite for "sovereign" framing on the Chinese side, connecting to Sovereign AI (Product Concept). The Cyberspace Administration of China (CAC) is the administrative authority for the regime.
Relationships
- deployed-by: Cyberspace Administration of China (CAC) (CAC) + 6 ministries (per the 2023 Interim Measures)
- supports: PEAT — Proactive Elite Alignment Theory (registration is the structural anchor of the framework)
- supports: The Incentive Architecture Export Controls Cannot Reach (Sun, Lawfare, May 6 2026) (canonical source)
- related: Export Controls (AI) (sits in tension with US export-control policy frame), Sovereign AI (Product Concept), Brussels Effect (analogy: jurisdictional regulatory spillover), Chinese AI Policy, US-China AI Competition: Different Races, Different Metrics
Sources
- The Incentive Architecture Export Controls Cannot Reach (Sun, Lawfare, May 6 2026) — Sun, The Incentive Architecture Export Controls Cannot Reach, Lawfare, May 6 2026 (canonical anchor)
- 2023 Interim Measures for the Management of Generative AI Services: cac.gov.cn (English translation: chinalawtranslate.com)
- CAC end-2025 registration count: cac.gov.cn
- CSET analysis of 31-risk-five-category framework: cset.georgetown.edu
- Revised Cybersecurity Law (effective Jan 1, 2026): npc.gov.cn