Provisions on the Administration of Deep Synthesis of Internet Information Services
English · 中文
互联网信息服务深度合成管理规定 — CAC, MIIT and MPS Order No. 12: adopted at the CAC's 21st executive meeting of 2022 on 3 November 2022, signed and promulgated 25 November 2022, in force 10 January 2023. Five chapters, twenty-five articles. China's first binding rule on synthetic media, and still the origin of the labelling duty that the 2025 measures later gave technical detail to.
In one line
Anyone offering deep-synthesis services to the public in China must verify users' real identity, review both the input and the output, attach a label to every synthesised item and a prominent label where the result could make the public mistake it for reality, never let anyone strip those labels, and — if the service has public-opinion attributes or social-mobilisation capacity — file the algorithm under the recommendation-algorithm rules. The provisions set no fine of their own.
Who and what it governs
- Scope (Article 2) — applying deep-synthesis technology to provide internet information services inside mainland China. Where other laws or administrative regulations provide otherwise, those apply.
- Deep-synthesis technology (Article 23) — using generative and synthesising algorithms such as deep learning and virtual reality to produce text, images, audio, video or virtual scenes; the list is expressly open-ended ("including but not limited to") and names six families: text generation, style transfer and question-answering; text-to-speech, voice conversion and voice-attribute editing; music generation and scene-sound editing; face generation, face swap, person-attribute editing, face manipulation and pose manipulation; image generation, enhancement and restoration; and three-dimensional reconstruction and digital simulation, used to generate or edit digital humans and virtual scenes. The same article defines the three actors — provider, technical supporter and user (深度合成服务使用者) — together with training data and immersive simulated scenes, the last being highly realistic virtual scenes generated or edited with the technology that participants can experience or interact with.
- Two duty-holders — the service provider (深度合成服务提供者) and the technical supporter (技术支持者). Article 19 makes technical supporters file on the same terms as providers; Articles 14, 15 and 21 place duties on both.
- Regulators (Article 3) — CAC coordinates nationally and locally; the telecom authority (MIIT) and public security act within their remit.
Key provisions (the ones that change behaviour)
| Article | What it requires |
|---|---|
| 6 | No organisation or individual may use deep-synthesis services to make, copy, publish or spread information banned by law, or to endanger national security and interests, damage the state's image, infringe the public interest, disrupt economic and social order or violate others' lawful rights. Providers and users must not use the service to produce or spread false news information; republishing news made with deep synthesis must draw on news released by licensed internet news information source units. |
| 7–8 | The provider bears primary responsibility for information security, with management systems covering user registration, algorithm-mechanism review, science-and-technology ethics review, publication review, data security, personal-information protection, telecom-fraud prevention and emergency response, backed by safe and controllable technical measures. It must publish its management rules and platform pledges, complete service agreements, and remind technical supporters and users prominently of their information-security obligations. |
| 9 | Real-identity verification of users by mobile phone number, ID-card number, unified social credit code, or the national public service for network identity authentication; no information-publishing service may be provided to a user who has not been verified. |
| 10 | Review users' input data and the synthesised result, by technical or manual means. Maintain a feature library for identifying illegal and harmful information, with entry standards, rules and procedures, and keep network logs. On finding illegal or harmful information: dispose of it by law, keep records, report promptly to the cyberspace administration and other competent departments, and act against the user by law and contract — warning, function restriction, service suspension or account closure. |
| 11 | Build a rumour-response mechanism: on discovering false information made, copied, published or spread through the service, take rebutting measures promptly, keep records and report. |
| 12 | Provide convenient user appeals and public complaints and reports, publish the process and response deadlines, and handle and feed back in time. |
| 13 | App stores and other distribution platforms must carry out listing review, daily management and emergency response, verify a deep-synthesis app's security assessment and filing, and on breach refuse listing, warn, suspend or remove it. |
| 14 | Providers and technical supporters must manage training data and protect its security; where training data contains personal information, the personal-information rules apply. Where they offer editing of biometric information such as faces and voices, they must prompt the user to inform the person being edited and obtain that person's separate consent. |
| 15 | Regularly review, assess and verify the mechanism of the generating and synthesising algorithms. Where a model, template or other tool can (i) generate or edit biometric information such as faces and voices, or (ii) generate or edit non-biometric special objects or scenes that may touch national security, the state's image, national interests or the public interest, the actor must carry out a security assessment, itself or through a professional institution. |
| 16 | For information generated or edited with the service, the provider must add a label by technical means that does not interfere with the user's use, and retain log information as required by law and national rules. |
| 17 | Where the service may cause the public to confuse or misidentify the content, a prominent label must be placed in a reasonable position or area of the content, signalling that deep synthesis is involved. The listed services are: (i) simulated-person text generation or editing such as intelligent dialogue and intelligent writing; (ii) synthesised or imitated voice, or editing that markedly changes a person's identity features; (iii) face generation, face replacement, face manipulation and pose manipulation, or other image/video generation or editing that markedly changes identity features; (iv) immersive simulated scenes; (v) other services with generation or marked-editing functions. For deep-synthesis services outside those categories the provider must supply a prominent-labelling function and prompt the user that it can be applied. |
| 18 | No organisation or individual may use technical means to delete, tamper with or conceal the deep-synthesis labels required by Articles 16 and 17. |
| 19–20 | Providers of services with public-opinion attributes or social-mobilisation capacity must perform filing, amendment and cancellation under the Provisions on the Administration of Algorithmic Recommendations; technical supporters do the same by reference. Those who have filed must display their filing number in a prominent position on the website or app and provide a public-information link. Developing and launching new products, applications or functions with those attributes requires a security assessment under national rules. |
| 21 | CAC, the telecom authority and public security inspect within their remit; providers and technical supporters must cooperate and supply necessary technical and data support. Where a material information-security risk is found, the authorities may require suspension of information updates, of new account registration, or of other related services, with rectification to remove the hazard. |
| 22 | Penalties. Violations are punished under the relevant laws and administrative regulations, with heavier punishment where serious consequences result; conduct amounting to a public-security breach is punished by the public security organs, and criminal liability follows where a crime is committed. |
| 24 | Providers and technical supporters engaged in internet publishing, network cultural activities or network audiovisual programme services must also meet the rules of the press-and-publication, culture-and-tourism and broadcasting authorities. |
The labelling core, and what came after it
Articles 16, 17 and 18 are where China's AI-content labelling regime began: a technical label that must not get in the user's way, a prominent label where confusion is possible, and a prohibition on removing either. That three-part structure is the ancestor of the 2025 Measures for Labelling AI-Generated Synthetic Content and its mandatory standard GB 45438-2025, which supply the format detail the provisions themselves leave open; the 2026 Anthropomorphic Interaction Services Measures Article 18 likewise points back at the labelling duty rather than replacing it. Both earlier instruments remain in force — reading the three as a stack rather than a replacement is this repository's characterisation, not a statement in the texts. — C
What the official Q&A adds, and what it does not
The CAC answered reporters on the day the provisions were posted (gov.cn policy interpretation, 2022-12-12; the page credits the CAC site). Three things belong on the record, and one absence matters more than any of them.
- Where the rule came from. The Q&A's background answer cites 《法治社会建设实施纲要（2020-2025年）》 as having expressly called for rules on new technology applications — its words are 制定完善对算法推荐、深度伪造等新技术应用的规范管理办法 — which is the common origin of these provisions and of the algorithm-recommendation provisions.
- Who files and who assesses: restated, not delimited. Question seven repeats the trigger and sends the procedure to the Provisions on the Administration of Algorithmic Recommendations; question eight repeats it for new products, new applications and new functions and sends the assessment to "relevant national regulations" (按照国家有关规定). Neither supplies a test for "public-opinion attributes or social-mobilisation capacity" — the delimitation is in the instrument the referral points at — Article 2 of the Security Assessment Provisions (CAC and MPS, issued 2018-11-15, in force 2018-11-30), which lists ten categories of service and a two-limb catch-all, and is now carded here → social-mobilisation-security-assessment-provisions.md.
- Technical supporters are named inside the referral. They file, amend and cancel "by reference to the service provider" (参照服务提供者), and filing numbers must be displayed prominently with a link to the public information — matching Article 19 rather than extending it.
The rest paraphrases Articles 6–19 and 21: the primary-responsibility list, input-and-output review backed by a feature library, the rumour-response mechanism, the three-part labelling duty, and the inspection powers — where a material information-security risk is found, the authorities may require suspension of information updates, of new account registration, or of other services. Nothing in it reaches further than the text. — B
Relationship to other laws
- Legal basis named in Article 1 — Cybersecurity Law, Data Security Law, Personal Information Protection Law and the Measures for the Administration of Internet Information Services.
- Article 19 imports the filing machinery of the Provisions on the Administration of Algorithmic Recommendations; the same route reappears in Article 17 of the generative-AI measures and Article 26 of the anthropomorphic-services measures.
- Article 14's separate-consent rule for faces and voices is the operative portrait and voice protection in this area, sitting alongside the Personal Information Protection Law's sensitive-information rules.
- Against the EU: the AI Act's transparency clause covers synthetic content (machine-readable marking) and deep fakes (disclosure), applicable from 2026-08-02 — see
../eu/ai-act-transparency.md. The EU duty is to disclose; these provisions additionally require identity verification of the person doing the synthesising (Article 9), consent from the person being synthesised (Article 14), review of inputs as well as outputs (Article 10), and a rumour-response mechanism (Article 11), and they bind app stores to gate the distribution channel (Article 13).
Why this matters for the comparison
These provisions, not the AI Act and not the 2025 labelling measures, are where China's synthetic-media regime was set out — a year and a half before the EU's transparency duties began to apply. They also show the method this repository describes as China's: regulate a service and its distribution chain (provider, technical supporter, app store, user), reach the underlying model only through security assessment and algorithm review, and lean on the general laws for punishment.
Official sources
- Full text, CAC (Order No. 12, promulgated 2022-11-25, posted 2022-12-11): https://www.cac.gov.cn/2022-12/11/c_1672221949354811.htm
- Full text, State Council policy document library (成文日期 2022-11-25): https://www.gov.cn/zhengce/zhengceku/2022-12/12/content_5731431.htm
- Announcement of the three departments' joint issuance (gov.cn news): https://www.gov.cn/xinwen/2022-12/12/content_5731429.htm
- Policy interpretation (gov.cn): https://www.gov.cn/zhengce/2022-12/12/content_5731430.htm
Source quotes
Every sentence below is copied from the official text and pinned to the snapshot in snapshots/ that this card was written against. node scripts/check-quotes.mjs re-checks them offline, so a quote that stops matching the source fails a check instead of quietly becoming the record. The prose above is a rendering; these are the anchors.
（六）三维重建、数字仿真等生成或者编辑数字人物、虚拟场景的技术— Article 23, the sixth family this card used to omit (snapshot:snapshots/cn/deep-synthesis-provisions.txt)（四）人脸生成、人脸替换、人物属性编辑、人脸操控、姿态操控等生成或者编辑图像、视频内容中生物特征的技术— Article 23, the biometric family (snapshot:snapshots/cn/deep-synthesis-provisions.txt)采取技术或者人工方式对深度合成服务使用者的输入数据和合成结果进行审核— Article 10, input and output review (snapshot:snapshots/cn/deep-synthesis-provisions.txt)应当提示深度合成服务使用者依法告知被编辑的个人，并取得其单独同意— Article 14, separate consent for faces and voices (snapshot:snapshots/cn/deep-synthesis-provisions.txt)任何组织和个人不得采用技术手段删除、篡改、隐匿本规定第十六条和第十七条规定的深度合成标识— Article 18, the anti-stripping rule (snapshot:snapshots/cn/deep-synthesis-provisions.txt)造成严重后果的，依法从重处罚— Article 22 — heavier punishment, but no amount of its own (snapshot:snapshots/cn/deep-synthesis-provisions.txt)具有舆论属性或者社会动员能力的深度合成服务提供者— the Article 19 filing trigger (snapshot:snapshots/cn/deep-synthesis-provisions.txt)明确提出制定完善对算法推荐、深度伪造等新技术应用的规范管理办法— the policy origin named in the Q&A (snapshot:snapshots/cn/deep-synthesis-provisions-qa.txt)深度合成服务技术支持者应当参照服务提供者履行备案和变更、注销备案手续— supporters file by reference (snapshot:snapshots/cn/deep-synthesis-provisions-qa.txt)应当按照国家有关规定开展安全评估— the assessment referral, which supplies no test (snapshot:snapshots/cn/deep-synthesis-provisions-qa.txt)
Unverified / TODO
- The English rendering of every article name and term is this repository's own translation — no official English text was found. ⚠ [gap:no_official_translation]
- Read the official policy interpretation above and record how the drafters delimit "public-opinion attributes or social-mobilisation capacity" — read 2026-10-11: it does not delimit them. Both mentions restate the trigger and refer the procedure outwards (filing to the algorithm-recommendation provisions, assessment to "relevant national regulations"), so the test has to come from those instruments — see the next item.
- The instrument that operationalises the trigger is identified and carded — and this card had its name and year wrong: it is the Provisions on the Security Assessment of Internet Information Services with Public-Opinion Attributes or Social-Mobilisation Capacity (具有舆论属性或社会动员能力的互联网信息服务安全评估规定) — provisions, not measures, and 2018-11-15, not 2022 — issued jointly by the CAC and the Ministry of Public Security, seventeen articles, no order number. Its Article 2 is the delimitation the AI instruments defer to → card.
- Confirm which MIIT or NRTA rules now apply additionally under Article 24, and whether any have changed since 2023. ⚠
- Check the interaction with the 2025 labelling measures and GB 45438-2025 in practice: which label specification regulators cite for a given case. ⚠
- Close the ⚠ in
compare/01-ai-labeling.mdabout penalties under the deep-synthesis provisions using Article 22: this instrument sets no fine amount of its own — done: the penalty row and the limits bullet incompare/01(both languages) and incompare/05now carry it, and the row moved from B to A.