Last reviewed
Jun 9, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v3 · 5,669 words
Add missing citations, update stale details, or suggest a clearer explanation.
The Interim Measures for the Management of Generative Artificial Intelligence Services (Chinese: 生成式人工智能服务管理暂行办法, Shengchengshi rengong zhineng fuwu guanli zanxing banfa) are a binding administrative regulation issued by the Cyberspace Administration of China (CAC) together with six other central government agencies on 10 July 2023, taking effect on 15 August 2023.[1][2] The Interim Measures are widely described as the world's first comprehensive, dedicated regulation on generative artificial intelligence systems offered to the public, predating the European Union's AI Act in legal effect by roughly eleven months.[3][4]
The Measures regulate any provider that offers generative AI services to the public in mainland China, requiring those services to undergo a security self-assessment, register their algorithms in the CAC's algorithm filing system, label synthetic content, restrict generation of content deemed harmful to "Core Socialist Values" and state interests, and adopt user protection and complaint handling procedures.[1][5] The character string 暂行办法 ("interim" or "provisional measures") signals their temporary status pending a more comprehensive national Artificial Intelligence Law, which has been repeatedly drafted but not enacted as of mid-2026.[6][7]
Although technically still in force, the Interim Measures have been substantially supplemented by the Measures for Labeling of AI-Generated Synthetic Content (effective 1 September 2025), the TC260-003 Basic Security Requirements for Generative AI Services (issued 29 February 2024, with mandatory national standards taking effect in 2025), and the AI provisions added through the October 2025 amendment to the Cybersecurity Law (effective 1 January 2026).[8][9][10] Together these instruments form the operational core of what scholars and law firms call China's "vertical, technology-specific" approach to AI governance.[11][12]
The Interim Measures are the fourth layer in a stack of Chinese cyberspace regulations enacted between 2017 and 2023, each building on the previous one. Understanding them requires the prior context.[13]
China's general digital governance framework rests on three statutes: the Cybersecurity Law (passed 2016, effective 1 June 2017), the Data Security Law (passed June 2021, effective 1 September 2021), and the Personal Information Protection Law or PIPL (passed August 2021, effective 1 November 2021).[14] PIPL is the most directly relevant to generative AI training. It is China's first comprehensive personal information statute, comparable in scope to the European Union's General Data Protection Regulation, and it imposes consent requirements, data minimisation duties, cross-border transfer restrictions and pre-processing risk assessments on automated decision-making.[15] Generative AI providers training on personal data are bound by PIPL, and the Interim Measures repeatedly cross-reference it.[1]
On 31 December 2021 the CAC, together with the Ministry of Industry and Information Technology (MIIT), the Ministry of Public Security (MPS) and the State Administration for Market Regulation, finalised the Administrative Provisions on Algorithm Recommendation for Internet Information Services, which took effect on 1 March 2022.[16][17] These provisions established the algorithm filing system that the Interim Measures later piggyback on. Operators of algorithmic recommendation systems with "public opinion attributes or social mobilisation capacity" were required to register algorithm summaries, designs, data sources and security self-assessments with the CAC. The 2022 provisions prohibited fake news generation by algorithm, required users to be offered an option to disable personalised recommendations and imposed special protections for minors and the elderly.[17]
The CAC, MIIT and MPS jointly issued the Provisions on the Administration of Deep Synthesis of Internet Information Services on 25 November 2022, with effect from 10 January 2023.[18] Deep synthesis is defined as the use of "generative and/or synthetic algorithms, such as deep learning and virtual reality, to produce text, graphics, audio, video or virtual scenes."[18] The Deep Synthesis Provisions require conspicuous labels on synthetic content, require platforms to identify and block harmful deep synthesis content, prohibit generation of false news without authorisation and require real-name verification of users of deep synthesis services.[19] Many of these provisions were then incorporated by reference into the Interim Measures, which explicitly defer to the Deep Synthesis Provisions on content labelling.[1]
By the time generative chatbots based on large language models drew global regulatory attention with the launch of OpenAI's ChatGPT in November 2022, China therefore already had two specialised algorithmic regulations on the books and could move quickly to adapt them to large-model services.[20]
On 11 April 2023 the CAC released a draft "Administrative Measures for Generative Artificial Intelligence Services" for public consultation, with a comment period closing on 10 May 2023.[2][21] The April draft was notable for the severity of its obligations. It required providers to "ensure the authenticity, accuracy, objectivity and diversity" of training data as an absolute standard; it imposed a three-month deadline for providers to eliminate the reappearance of illegal content; it allowed administrative fines of 10,000 to 100,000 yuan (roughly USD 1,400 to USD 14,000); and it permitted authorities to order both suspension and termination of services.[21][22] The draft also implicitly placed liability for output content directly on providers and barred user profiling based on input prompts.[21]
Following the consultation period, the final version received internal CAC approval on 23 May 2023 and was publicly issued on 10 July 2023 over the names of seven agencies.[2] Analysts at the Future of Privacy Forum, Carnegie Endowment for International Peace and law firms including Hogan Lovells documented that the final version softened several requirements after pushback from scholars, the Chinese AI industry (with Baidu reportedly lobbying quietly for changes) and likely from the National Development and Reform Commission, which co-signed the final version.[21][23][24] Major changes included:
Several obligations went in the opposite direction. The final version added an explicit duty on users (not only providers) to uphold Core Socialist Values, introduced a written service-agreement requirement between providers and users, codified extraterritorial reach over foreign services accessible to Chinese users, and expanded the list of issuing agencies from one (the CAC, in the draft) to seven (in the final), distributing regulatory authority across the government.[21]
The Interim Measures are unusual among CAC instruments in that they bear the seals of seven national agencies, signalling broad inter-ministerial agreement and dispersing implementation authority.[1][4][27] The seven issuers are:
Inclusion of NDRC alongside content control agencies has been read by China observers as a deliberate signal that the regulation is intended as much to develop the Chinese generative AI industry as to constrain it.[23][28] The Asia Society Policy Institute and Carnegie analysts have noted that the seven-agency structure also makes future revision negotiation-intensive, contributing to the regulation's "interim" character.[7][23]
Article 2 of the Measures sets the scope. The Measures apply to "the use of generative AI technology to provide services to the public within the People's Republic of China, including services generating text, images, audio, video and other content."[1][5] Three features are worth noting.
First, the public-facing trigger narrows scope considerably. Article 2 expressly states that organisations and individuals that research, develop or apply generative AI technology without providing services to the domestic public are not within scope. Internal corporate use, university research and embedded AI in non-consumer software fall outside the Measures, although they remain bound by PIPL and the Cybersecurity Law.[5][29]
Second, extraterritorial reach is broad. The Measures apply to any service accessible to mainland Chinese users regardless of the provider's location.[21][25] In practice, however, foreign providers such as OpenAI, Anthropic and Google have not obtained CAC algorithm filings; their services are blocked in mainland China and the Measures function more as a registration gate than as an enforcement mechanism against offshore actors.[30]
Third, the Measures are vertical and content-specific: they regulate one class of AI technology (generative models producing text, images, audio, video or similar content) rather than the entire AI sector. This contrasts with the horizontal, risk-tiered approach of the eu ai act, which applies a single framework across all AI uses.[11][12] See further the comparison section below.
Article 4 is the most politically central provision of the Measures. It requires providers and users of generative AI services to "abide by laws and administrative regulations, respect social morality and ethics" and to comply with five enumerated principles.[1][31] In summary, generated content must:
"Core Socialist Values" (社会主义核心价值观) is the formal Communist Party of China value framework first articulated at the 18th Party Congress in 2012 and consisting of twelve values across three levels: at the national level prosperity, democracy, civility and harmony; at the social level freedom, equality, justice and rule of law; and at the personal level patriotism, dedication, integrity and friendship.[32] Article 4(1) of the Interim Measures gives this political programme legal teeth in the context of AI output. In combination with the prohibitions on subversion, separatism and unauthorised "false information," it constitutes what observers including Asia Society scholars and Lawfare analysts describe as "censorship by design."[33][7] In practice, CAC large-model filing reviews focus heavily on whether a model refuses prompts on topics including the Tiananmen Square events of 1989, the political status of Taiwan, the treatment of ethnic minorities in Xinjiang and Tibet, and criticism of senior Party leadership.[28][33]
Article 4 also creates a duty to design models with non-discrimination in mind, a rare formal recognition of algorithmic bias in Chinese law that parallels (though it does not match the detail of) the EU AI Act's fundamental-rights provisions.[11]
Article 7 governs training data. Providers must use lawfully sourced data, may not infringe intellectual property rights held by others, must obtain consent where personal information is used, and must take "effective measures to improve the quality of training data, enhancing the authenticity, accuracy, objectivity and diversity of training data."[1][5] The shift from the draft's hard guarantee to an effort-based standard was the single most important relaxation between draft and final.[21]
Article 8 governs data annotation. Providers must establish clear annotation rules suited to the type of service, conduct quality assessments of annotations, sample-verify annotation correctness and provide necessary training to annotators on lawful and ethical labelling.[1] The TC260-003 standard issued in 2024 fleshed out these obligations, requiring at least 4,000 sample manual reviews of pre-processed training data and at least 100 sample reviews of processed data, with a 5 per cent illegal-or-harmful-content threshold that disqualifies a dataset for training use.[9][34]
Article 12 mandates conspicuous labelling of generated images, videos and other content "in accordance with the Provisions on Administration of Deep Synthesis of Internet Information Services."[1][5] In other words, the Interim Measures defer to the older Deep Synthesis Provisions on the technical specifics of synthetic-content labelling, until the 2025 Labelling Measures specifically replaced and elaborated those rules.
Article 17 is the operational core of the Measures and the provision through which most public information about Chinese generative AI services flows. It requires that "providers of generative AI services with public opinion properties or the capacity for social mobilisation must conduct security assessments in accordance with relevant national provisions and complete algorithm filing, modification and cancellation filing procedures in accordance with the Administrative Provisions on Algorithm Recommendation for Internet Information Services."[1][5]
In practice the CAC operates a two-track filing regime for generative AI:
The CAC published its first public list of approved large-model filings on 2 April 2024, after months of internal review of major Chinese models including Baidu's Ernie Bot, Alibaba's Tongyi Qianwen, Baichuan's Baichuan models, 01.AI's Yi family, Moonshot AI's Kimi, and ByteDance's Doubao.[28] By the end of 2024, 238 generative AI services had been filed under the large-model track in that year alone (compared to 64 in 2023, when the regulation first took effect).[23][24] By April 2025 the running total reached 346 publicly disclosed filings, climbing to roughly 748 national-level filings and 435 local-level registrations by the end of 2025.[36][39] The CAC reports that approximately 250 to 300 new generative algorithmic tools receive approval each month.[36]
Analysts caution that the number of filings is a poor proxy for innovation: Alibaba/Ant Group leads with 66 filings (reflecting many subsidiaries), Tencent has 52, NetEase 37, Inspur 28, Baidu 24, ByteDance 20 and SenseTime 19; meanwhile deepseek, the most internationally prominent Chinese model maker since the January 2025 release of DeepSeek-R1, has only 3 filings.[36]
Articles 9, 14, 15 and 17 collectively establish a security and content moderation regime that operates alongside the algorithm filing system.[1][5] Article 9 designates providers as legal "producers of online information content," obliging them to take responsibility for the content their models generate (a doctrine that effectively rejects the Section 230 style platform immunity familiar from US law). Article 14 requires providers to take immediate action including model optimisation and reporting to authorities when they discover illegal content, and to keep records of complaints. Article 15 establishes a complaint and reporting channel for users and outside parties.[1][5][25]
The technical detail of security assessment is supplied by TC260-003 Basic Security Requirements for Generative AI Services, issued on 29 February 2024 by the National Technical Committee 260 on Cybersecurity (TC260) of the Standardisation Administration of China.[9][34] TC260-003 enumerates more than 30 categories of safety risk, ranging from violations of Core Socialist Values to algorithmic discrimination, exposure of personal information, and intellectual property infringement, and prescribes the specific sampling, testing and documentation procedures that satisfy Article 17 security assessment obligations. Although TC260 standards are technically voluntary (recommended) national standards, in practice the CAC has treated TC260-003 as the operational baseline for filing approval.[34]
The Interim Measures formally apply to foreign generative AI services accessible to mainland Chinese users, but several practical features limit foreign exposure.[21][25] First, the major US, European and Japanese model providers do not generally offer services in mainland China; OpenAI, Anthropic and Google AI are blocked, and Chinese versions of foreign services typically run through Chinese subsidiaries with localised infrastructure. Second, registry data show foreign companies operate only 19 generative algorithmic tools in the public registry, a 0.5 per cent share, and these are mostly infrastructure or productivity utilities (Amazon Web Services, Canva, IKEA, GE Healthcare).[36] Third, where foreign companies wish to offer generative AI services in China, they typically partner with a Chinese filed entity rather than seek filings of their own.
Cross-border data transfer obligations under PIPL, the Data Security Law and the CAC's 2022 cross-border transfer security assessment rules constitute an additional layer that foreign providers face when training models or operating services involving Chinese personal information.[15][29]
Article 21 provides the operative penalty clause. It defers to existing statutes for most enforcement: violations of the Cybersecurity Law, Data Security Law, PIPL and Law on the Progress of Science and Technology are penalised under those statutes' regimes (which include fines, suspension of services, public criticism and, where criminal conduct is established, referral for criminal investigation). Where no statute provides for a penalty, the relevant agency may issue warnings, public criticism, orders to rectify within a specified period and, if rectification is refused or the violation is serious, orders to suspend services.[1][25][26]
The draft's monetary fines (10,000-100,000 yuan) and explicit power to terminate services were removed from the final version, which is widely read as a deliberate softening to encourage industry investment.[21][22] Critics including Angela Huyue Zhang have argued that the Measures function more as a "signalling device" pro-growth messaging than as a substantial enforcement tool, while Carnegie Endowment analysts note that enforcement focuses heavily on politically sensitive content rather than uniform application of all provisions.[23][40]
Public enforcement actions accelerated through 2024 and 2025. Notable examples include:
The September 2025 Cybersecurity Incident Reporting Measures further require generative AI providers to report "relatively serious" incidents affecting 10,000 or more users (or recurring five times within 24 hours) to provincial authorities within four hours.[39]
The Interim Measures shaped the structure and pacing of the Chinese generative AI industry from late 2023 onward.[28][36]
In the months immediately following the August 2023 effective date, the CAC granted a first wave of filings to politically reliable, well-resourced providers including Baidu (Ernie Bot 3.5, then 4.0), Alibaba (Tongyi Qianwen), 360 Search, iFlyTek, Baichuan and SenseTime. Many startups reported a four to nine month internal preparation effort to assemble keyword lists, evaluation question sets and self-assessment reports compliant with TC260-003.[28] Industry observers noted that the filing process effectively functioned as a soft licensing regime, slowing market entry and concentrating early advantages in incumbents with substantial legal and compliance resources.
The filing system also gave the Chinese state continuing visibility into model architecture and training data, a capacity that the Shanghai CAC and other regional offices have used to compel rapid model updates when test prompts elicit politically problematic outputs. Carnegie's analysis suggests that the CAC's filing decisions correlate strongly with content moderation performance rather than with capability benchmarks.[23]
deepseek, whose December 2024 release of DeepSeek-V3 and January 2025 release of DeepSeek-R1 substantially reset international perception of Chinese AI capabilities, has only three CAC filings.[36] DeepSeek's rapid international visibility prompted intensified rather than relaxed state oversight: Zhejiang provincial authorities reportedly began screening prospective investors in DeepSeek before meetings with company leadership, instructed headhunters to halt talent recruitment from the company, and required certain employees to surrender passports owing to state-secrets classifications.[23]
Baidu's Ernie family of models is a useful case study in interaction with the filing system. Ernie Bot was the first widely available Chinese chatbot to clear large-model filing (in August 2023). Baidu maintains separate filings for the consumer-facing Ernie Bot and for enterprise API offerings of the same underlying model, illustrating that the filing requirement attaches to the service rather than the model.[28] Baidu also reportedly lobbied during the consultation period for the relaxations adopted in the final version, reflecting the political weight of leading domestic AI companies in shaping the rules they would later operate under.[23][24]
Foreign-developed models adapted for the Chinese market, such as the Meituan and Tencent Cloud offerings of Llama derivatives, have filed under the standard algorithm filing track rather than the large-model track when they were repositioned away from public-facing consumer use, illustrating providers' strategic choices about whether to subject themselves to the more onerous tier.[36]
Several instruments adopted between 2024 and 2026 supplement the Interim Measures and reinforce their effect.[8][9][39]
TC260-003 Basic Security Requirements for Generative AI Services was issued by the National Technical Committee 260 on Cybersecurity on 29 February 2024 and provides the operational specifications for security assessments under Article 17, including the 4,000-sample manual review, the 5 per cent illegal content threshold and the more than 30 enumerated risk categories.[9][34]
The Measures for Labeling of AI-Generated Synthetic Content (人工智能生成合成内容标识办法), jointly issued on 7 March 2025 by the CAC, MIIT, MPS and NRTA, took effect on 1 September 2025.[8][43][44] Together with mandatory national standard GB 45438-2025, the Labelling Measures elaborate the labelling duties that the Interim Measures' Article 12 had delegated to the older Deep Synthesis Provisions. Providers must add explicit labels (visible text, voice cues, watermarks or markers indicating "AI-generated") and implicit labels (embedded metadata, including provider identifiers and content reference numbers) to all generated text, images, audio, video and virtual scenes. Platforms must verify metadata on uploaded content and add warning labels when detection reveals likely synthetic origin. Users publishing AI content must proactively declare it. Maliciously deleting or altering labels is prohibited. Providers must retain six-month logs of unlabeled content.[8][43]
The Data Annotation Security Specification (GB/T 45674-2025) and a series of Emergency Response Guidelines released in September 2025 added incident classification and reporting duties, while the AI Security Governance Framework Version 2.0 issued on 15 September 2025 set out higher-level ethical and traceability principles.[39]
On 28 October 2025 the Standing Committee of the National People's Congress adopted amendments to the Cybersecurity Law, effective 1 January 2026, formally incorporating AI provisions into the country's general cybersecurity statute. These amendments support algorithm research and development, require AI ethics review and risk assessment for major projects and explicitly recognise AI within the cybersecurity legal regime.[7][45] Although a comprehensive Artificial Intelligence Law (人工智能法) was placed on the legislative agenda in 2023 and 2024, it was dropped from the 2025 schedule, and no official draft has been released as of mid-2026.[6][7] The Interim Measures therefore remain in force as the primary regulation specifically governing generative AI services in China.
A related public-facing 2025 initiative, the State Council Action Plan for AI Plus (人工智能+行动计划, distinct from the State Council's broader sectoral AI guidance), encouraged government, state-owned enterprise and commercial deployment of AI tools, with explicit reference to compliance with the Interim Measures and filing under Article 17.[39] For comparison with parallel public AI policy actions in other jurisdictions, see ai action plan.
The Interim Measures occupy a distinctive position in global AI regulation. Three regulatory archetypes can be identified.[11][12][46]
The European Union AI Act, adopted 13 June 2024 and entering progressively into force from 1 August 2024 through 2027, is a horizontal, risk-tiered framework. It defines AI broadly and applies a single regulation across all AI uses, classifying systems as prohibited, high-risk, general-purpose or limited-risk and imposing graduated obligations. Generative AI ("general-purpose AI") is one regulated category among many. The EU model embeds transparency obligations and fundamental-rights protections at the centre of compliance, with national market surveillance authorities and the AI Office acting as enforcers and fines reaching up to 7 per cent of global turnover. See eu ai act and the gpai code of practice for general-purpose AI providers.[11][46]
The Chinese Interim Measures, by contrast, are vertical and content-specific. They target one technology class (generative AI services offered to the public), defer to existing statutes for penalties, route compliance through algorithm and large-model filings, and prioritise information control (Article 4, socialist values) and industrial development (Article 3, balance development and security) ahead of risk classification. The interim character of the Measures and their reliance on national standards and inter-agency campaigns allow rapid adjustment as the technology evolves, but also produce regulatory fragmentation and uncertainty.[11][12][23]
The United States has, as of mid-2026, no comprehensive federal AI statute. The federal approach has shifted from the rescinded Biden Executive Order 14110 (2023) towards a market-light, voluntary-standards regime under the Trump administration's 2025 AI Action Plan, with the National Institute of Standards and Technology's AI Risk Management Framework as a non-binding reference. States including California (SB 942, AB 853), Colorado, Illinois and others have enacted patchwork statutes addressing labelling, automated decision-making, employment discrimination and deepfakes; the Federal Trade Commission and Department of Justice have pursued enforcement under existing consumer protection and antitrust authorities. The Korean Framework Act on the Development of Artificial Intelligence (see korea ai basic act), enacted in early 2025 with effect from 22 January 2026, sits between the EU and Chinese poles by setting a high-level legislative framework while leaving operational detail to subsequent presidential decrees.[46][47][48]
Several Chinese provisions track international developments. Article 4's non-discrimination clause parallels EU AI Act fairness provisions and US state anti-discrimination statutes. Article 12 and the 2025 Labelling Measures parallel California SB 942 (which imposes labelling on AI providers with one million or more users) and Article 50 of the EU AI Act. The TC260-003 4,000-sample manual review parallels (and exceeds in numeric specificity) the EU AI Act's data quality requirements. Where the Chinese approach diverges sharply is on the political content of Article 4(1) and on the role of mandatory filing and security assessment as a gating mechanism for market entry.[11][46]
The Vrije Universiteit Amsterdam, Cambridge University and Belgravia Law have all published comparative analyses noting that the EU and Chinese approaches are converging on labelling and transparency but diverging on the scope of permissible content, the basis of liability (provider-as-publisher in China versus operator-of-system in the EU) and the role of central administrative gatekeeping.[11][12][49]
The Interim Measures for the Management of Generative AI Services consolidated, for the first time anywhere in the world, a single binding framework specifically for public-facing generative AI. Issued by seven agencies under the leadership of the CAC, the Measures combine content rules rooted in Core Socialist Values, training data and labelling requirements, mandatory algorithm and large-model filings, security assessment and a deferred penalty regime. Their relaxation between draft and final version signalled a deliberate decision to encourage Chinese industry investment in generative AI while preserving information-control levers. Their evolution through TC260-003 (2024), the Labelling Measures (2025), the AI Security Governance Framework (2025) and the amended Cybersecurity Law (2026) shows the Chinese state shifting from principle to implementation. While the Measures remain "interim" in name, they have become, in operation, the most substantive nationwide generative AI regulation in force, and their algorithm filing register has produced the most detailed public dataset on commercial generative AI products in any jurisdiction.