Research opportunity · Regulation-in-Practice: Rights, Compliance & Developmental Steering
China's AI filing register: how often are filings sent back, and whose models do apps use?
Question
Using the successive public lists of filed generative-AI services and registered applications published by China's internet regulator, together with interviews with applicants and filing consultants, how often do filings go through rejection and rework before approval, how often do listed services later disappear, and how concentrated is the dependence of registered applications on a small number of filed base models?
Institutional practice unknownOne PhD chapter or paperLow to medium costPartly addressed
Why it matters
The filing count is rising quickly, but counts of successful filings say nothing about how hard the gate is or what the regime does to market structure. Rejection and rework rates would show whether filing works as registration or as de facto licensing; the declared-model information that registered apps must display would show whether the regime is channelling the application market onto a few approved models.
The debate lists this among what we still do not know: Chinese rejection or sanction rates under the filing regime
Evidence that motivates it
- China's CAC reported 1,112 generative-AI services filed and 731 applications/functions registered by 31 August 2026, up from 988 and 598 at the end of June.
Filing and registration counts rose from 988 and 598 to 1,112 and 731 between June and August 2026, but the claim explicitly says nothing about rejection rates, adoption or market share. - China's CAC notices require launched generative-AI applications or functions to display the filed/registered service or model they use, including the model name and filing or launch number.
Launched applications must display the filed service or model they use, with its name and filing number, which makes model dependence observable; the claim does not license any measured concentration, which is what this study would supply.
A possible design
Two strands. (1) Register panel: download every published batch of the regulator's filing and registration lists (and provincial lists) since 2023, link entries across batches to build a panel, and record for each registered application the base model it declares (from the published list if given, otherwise from the in-app display the rules require, checked for a random sample). Measure concentration of applications on base models (shares and a standard concentration index) by sector and over time, and record services that drop out of later lists or appear in published enforcement-campaign notices. (2) Process evidence: 20–30 semi-structured interviews with applicant firms, compliance consultants and lawyers, recruited through industry networks, on the number of submission rounds, reasons for return, and time to approval; triangulate with any published regulator statements. Report rework as ranges from interviews, clearly labelled as such, not as an official rejection rate.
Data needed
Public CAC and provincial filing announcements; app-store pages and in-app displays for a sample; interview access in China (Mandarin required); ethics approval with attention to interviewee risk.
Feasibility
One PhD chapter or paper. Low to medium: researcher time, Mandarin-speaking assistance, travel for interviews or remote interviews. The regulator publishes only successful filings, so the register strand alone cannot give a rejection rate; the interview strand is essential and will give estimates, not counts. Interviewee confidentiality is a real constraint.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 12 October 2026, checked again on 13 October 2026).
Analysts have compiled the public registration lists and broken them down by firm, region, sector and ownership (Trivium China, April 2025), and academic work has described the registry descriptively by category, disclosure period and registrant concentration (Pi, Li and Singh 2026). Earlier work described the registry's design from its user manual (Carnegie, 2022) and its de facto licensing character from practitioner accounts (ChinaTalk). None reports rejection or rework rates, none counts services that later drop out of the lists, and none measures which base models registered applications depend on; Trivium only speculates that many large-model registrations are lightly fine-tuned copies of a few foundation models.
How strong this evidence is
- Who searched and checked: Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert; an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Every check so far was done by AI; no person or domain expert has checked it yet.
- What was read: none of the 5 nearest works below is recorded as judged from an abstract only. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by web search on the date shown; Chinese-language academic databases (CNKI, Wanfang) were not reachable, and the Chinese-language search returned only news pages, so Chinese scholarship on the register is likely under-covered. The ChinaTalk post carries no date on the page; 2024 is inferred from its reference to August 2024 figures. Absence of a hit is not proof that the work does not exist. Literature check of 2026-10-13: CNKI and Wanfang were not reachable, so Chinese academic scholarship remains under-covered; Chinese web results were consultancy and platform guides (which could inform the interview strand but are not research). A 2026 Journal of Contemporary China article ("Redefining China's Approach to AI Governance", DOI 10.1080/10670564.2026.2733081) surfaced but could not be opened, so whether it uses register data was not checked. The final ~22,000 characters of Pi et al. were not read. Absence of a hit is not proof that the work does not exist.
- Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.
Nearest work found
- Schaefer, "Seeking the next DeepSeek: What China's generative AI registration data can tell us about China's AI competitiveness," Trivium China (29 Apr 2025)
Compiles 3,739 records from the deep-synthesis and generative-AI registers and analyses them by firm, region, sector (labels assigned by an LLM and spot-checked) and state affiliation. Does not record base-model dependence or rejections; a starting dataset rather than the study proposed here. - Sheehan & Du, "What China's Algorithm Registry Reveals about AI Governance," Carnegie Endowment for International Peace (9 Dec 2022)
Infers what the earlier algorithm registry collects from its user manual and the first batch of 30 public filings (August 2022). Predates the generative-AI filing regime; no data on outcomes of filings. - Bit Wise (pseudonym), "SB 1047 with Socialist Characteristics: China's Algorithm Registry in the LLM Era," ChinaTalk (undated, 2024)
Practitioner account of how generative-AI filing works, describing repeated rounds of rework after rejection; states that no good data on failed filings exist because only successful filings are published. Qualitative, no counts. - Pi, Li & Singh, "Understanding the Role of Algorithm Registers in AI Governance Through Comparative Analysis of China and the UK," arXiv:2606.00035 (submitted 25 Apr 2026)
Compares China's filing (beian) system with the UK's Algorithmic Transparency Recording Standard using regulatory documents, guidelines and registry data. Reports descriptive registry statistics (5,822 unique entries from 3,787 entities as of November 2025; 76% of entities with a single filing; generative-AI service-provider and tech-support categories at 67.3% and 20.0%; registrations by disclosure period; top five registrants). Mentions iterative consultation and testing and app-store removals only in passing. On the first 100,000 characters read, it gives no rejection or rework rates, no delisting counts and no analysis of which base models registered applications use. No interviews mentioned in the abstract. The closest academic use of the register; does not do the proposed design. - Zou & Zhang, "Navigating China's regulatory approach to generative artificial intelligence and large language models," Cambridge Forum on AI: Law and Governance 1, e8 (2025); DOI 10.1017/cfl.2024.4
Doctrinal and policy analysis of the 2023 Interim Measures, including filing duties and the gap between foundation-model and application providers. On the portion read, no filing counts, rejection data or measurement of base-model dependence. Background only.
Literature check
Stands as first published. No study was found that measures rejection or rework rates in China's generative-AI filing process or the concentration of registered applications on filed base models. The closest new work, Pi, Li and Singh (arXiv:2606.00035, 2026), analyses the registry descriptively (5,822 entries from 3,787 entities as of November 2025, trends by disclosure period, category shares, top registrants) but, on the parts read, does not address rejection, delisting counts or base-model dependence. Chinese-language results were practitioner guides and CAC announcements, not research.
Searched: web (extended); Chinese-language web; arxiv.org; ssrn.com; osf.io; socarxiv.org; cambridge.org; tandfonline.com; sciencedirect.com; academic.oup.com; carnegieendowment.org; cset.georgetown.edu. Checked by Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert.
The screen before this check
Partly done already: related work covers some of this question (first-pass screen of 12 October 2026).
Analysts have compiled the public registration lists and broken them down by firm, region, sector and ownership (Trivium China, April 2025), and earlier work described the registry's design from its user manual (Carnegie, 2022) and its de facto licensing character from practitioner accounts (ChinaTalk). None reports rejection or rework rates, and none measures which base models registered applications depend on; Trivium only speculates that many large-model registrations are lightly fine-tuned copies of a few foundation models.
Search queries used
China generative AI service filing registry analysis dataset CAC filed models empirical study algorithm registryChina algorithm registry empirical analysis filings Carnegie Sheehan "algorithm registry" study entries生成式人工智能服务 备案 实证分析 备案清单 统计 研究 论文
Literature check, 13 October 2026:
生成式人工智能 备案 登记 大模型 应用 调用 已备案模型 集中度 研究China generative AI filing CAC registered models empirical analysis dataset [arxiv.org, ssrn.com, papers.ssrn.com, osf.io, socarxiv.org]China algorithm registry generative AI filing licensing regime empirical study journal 2025 2026 rejection approval process interviews大模型备案 退回 驳回 补正 通过率 多轮 审核 周期 企业 经验"Seeking the next DeepSeek" Trivium registration data China AI follow-up analysis foundation models registrations fine-tunedChina generative AI service registration regime market concentration foundation model dependence downstream applications study [cambridge.org, tandfonline.com, sciencedirect.com, academic.oup.com, link.springer.com, journals.sagepub.com, dl.acm.org, carnegieendowment.org, cset.georgetown.edu]
Provenance
RO-LD12-04 v2. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-12 debate on 2026-10-12. Checked by an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.
Changes
- v2, 13 October 2026: Literature check added: a second, deeper search for existing work, with its queries, new nearest work and limits. The question and verdict stand. See v1 as published.