China's deep synthesis (deepfake) provisions take effect
Providers must verify users' identities, obtain separate consent before generating a synthetic face or voice, and label AI-altered content.
- Government & policy
- Notable
China’s Cyberspace Administration, Ministry of Industry and Information Technology and Ministry of Public Security jointly issued the Provisions on the Administration of Deep Synthesis Internet Information Services on 25 November 2022; they took effect on 10 January 2023, weeks after ChatGPT’s public launch but before that launch had reshaped the Western industry the rules were written to police.
“Deep synthesis” in the provisions covers any service that uses generative techniques — text, image, audio, video or virtual-scene synthesis — to create or alter content. Providers must verify users’ real identities before offering the service, obtain a subject’s separate, explicit consent before generating synthetic biometric data such as their face or voice, and add a conspicuous label to outputs that could otherwise be mistaken for an unaltered recording, either visibly or in embedded metadata. Providers whose algorithms have the “capacity to mobilise public opinion” must additionally file them with the CAC and undergo a security assessment before deployment.
Because the rules were drafted around impersonation, fraud and disinformation using face- and voice-swapping tools, they only partly anticipated the general-purpose chatbot boom that began weeks earlier. China’s more direct response to large language models — interim measures aimed specifically at generative AI services — followed in mid-2023, after a stricter draft was softened to favour industry development over the earlier real-name and fixed-penalty requirements.
The deep-synthesis provisions were nonetheless among the first binding, nationally enforced labelling and consent regimes for synthetic media anywhere, arriving well ahead of comparable proposals debated in the EU and US at the time. Real-name verification, mandatory labelling and algorithm filing became recurring features of Chinese AI regulation, carried forward into the 2023 generative-AI rules and into later requirements for AI-content watermarking.