本頁4 分鐘閱讀
概述
The Deep Synthesis Provisions took effect in January 2023. The Measures for Labeling AI-Generated Synthetic Content, in force since September 1, 2025, require both visible labels and hidden metadata labels on AI-generated text, images, audio, video and virtual scenes. Together they make China one of the first jurisdictions to require AI content labeling across the whole distribution chain, from the tools that generate content to the platforms that spread it.
深入探討
The Provisions on the Administration of Deep Synthesis Internet Information Services were issued in late 2022 by the Cyberspace Administration of China (CAC), the Ministry of Industry and Information Technology and the Ministry of Public Security, and took effect on January 10, 2023. They define deep synthesis broadly as technology that uses deep learning or virtual reality to generate or edit text, images, audio, video or virtual scenes. Providers must verify users' real identities, set management rules, review content, and keep mechanisms for dispelling rumors. When they offer tools for editing biometric information such as a face or voice, they must prompt users to obtain the separate consent of the person being edited. They must also file their algorithms with the CAC. Content that could confuse or mislead the public, such as synthetic faces, cloned voices or realistic scenes, must carry a conspicuous label, and no one may remove or conceal these labels. The 2025 Labeling Measures, issued by the CAC with three other agencies and accompanied by a mandatory national standard, GB 45438-2025, make labeling more specific. Explicit labels are visible or audible cues such as text notices, audio prompts or graphic marks. Implicit labels are embedded in file metadata and include information such as the provider's name and a content ID. Distribution platforms must check uploads for metadata. They display a notice when AI generation is confirmed, and they apply possible or suspected labels when the user declares it or detection suggests it. App stores must check that apps comply. Western approaches are narrower or newer. The EU AI Act's Article 50 imposes transparency and marking duties on AI-generated content and deepfakes, but they apply later than China's. In the United States, deepfake rules are mostly state laws on election ads and intimate imagery, while standards such as C2PA content credentials are voluntary. Labels do not solve detection: metadata can be stripped by screenshots or re-encoding, which is why China layers platform detection and user declarations on top.
戰略影響
風險與安全
災難性和日常的人工智慧危害都取決於誰了解風險以及誰能夠採取行動。
更明確的決策
民眾和專業素養決定強而有力的安全政策在政治上是否可行。
突破炒作
清晰的解釋可以減少炒作、實驗室公關和模糊道德劇場的影響。
The Future of China's Deep Synthesis Rules and AI Content Labeling
Enforcement of the 2025 Labeling Measures is still new, and it is not yet clear how consistently platforms will detect unlabeled content or how regulators will treat false positives. Watermarking technology keeps improving, but no method yet survives every edit, so labeling will probably continue to rely on several signals together. As the EU's transparency obligations take effect and more countries consider labeling rules, technical interoperability between China's metadata standard and schemes like C2PA may become a practical question for global platforms.
現實世界的實施
A face-swap app must add a conspicuous label to its output videos and require users to obtain consent from the person whose face is being edited. It must also verify its users' real identities.
An AI image generator must embed hidden metadata in each file, such as the provider's name and a content identifier. Text generation tools must show visible cues, such as notices at the start or end of generated text or in the interface.
A short-video platform that detects AI-generation metadata in an upload must show a label telling viewers the content is AI-generated. If there is no metadata but its systems detect signs of synthesis, it adds a suspected label.
A creator uploading an AI-made voiceover to a Chinese platform is expected to declare it as AI-generated using the platform's labeling tool, and must not strip existing labels.
風險與防護欄
將存在風險視為科幻小說，同時能力複合。
混淆了表面產品安全與高度自治下的對準。
只給非英語和非專業觀眾留下低品質的資源。
實施路線圖
單獨的產品危害、誤用和失控/失調風險。
詢問哪些證據會改變您對時間表和嚴重性的看法。
比起行銷主張，更喜歡主要來源和具體評估。
確定一條行動路徑：職業、政策、資金或技能——而不僅僅是意識。
不斷探索
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the China's Deep Synthesis Rules and AI Content Labeling quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
常見問題
What is China's Deep Synthesis Rules and AI Content Labeling?
China regulates deepfakes and other synthetic media through two layers of rules. The Deep Synthesis Provisions took effect in January 2023. The Measures for Labeling AI-Generated Synthetic Content, in force since September 1, 2025, require both visible labels and hidden metadata labels on AI-generated text, images, audio, video and virtual scenes. Together they make China one of the first jurisdictions to require AI content labeling across the whole distribution chain, from the tools that generate content to the platforms that spread it.
When did China's Deep Synthesis Provisions take effect?
The Deep Synthesis Provisions were issued in late 2022 and took effect on January 10, 2023, before the generative AI Interim Measures.
When did the Measures for Labeling AI-Generated Synthetic Content take effect?
The Labeling Measures, together with the mandatory national standard GB 45438-2025, took effect on September 1, 2025.
What is an implicit label under China's labeling rules?
Implicit labels are embedded in metadata and are not directly visible. Explicit labels are the visible or audible cues shown to users.
What should a platform do when an upload has no AI metadata but its detection suggests synthesis?
Platforms use tiers. Confirmed content with metadata gets a clear AI-generated notice, and content flagged by detection gets a possible or suspected label.
Under the Deep Synthesis Provisions, what must be obtained before a user edits another person's face or voice?
Providers offering biometric editing tools must prompt users to obtain the separate consent of the person whose face or voice is being edited.
繼續學習
相關指南
為此主題精選的更多指南