Last reviewed: April 9, 2026

Jurisdictions covered: China (primary), EU and Korea (comparison)

Reading time: 12 minutes

Language note: Regulation text verified against original Chinese on cac.gov.cn (read March 21, 2026).

China’s Deep Synthesis Rules: The Labeling and Watermarking Requirements That Apply Right Now

Your AI-generated image must carry a visible text label at least 5% the height of its shortest side. Your AI-generated audio must embed a Morse code rhythm spelling “AI.” Your AI-generated video must display a watermark throughout playback. These are not suggestions. They are mandatory requirements under China’s national standard GB 45438-2025, enforceable since September 1, 2025.

On February 12, 2026, the CAC demonstrated it means it. In a single enforcement sweep, regulators penalized 13,421 accounts and removed 543,000+ pieces of content for failing to label AI-generated material. Violations included celebrity deepfakes, fabricated disaster footage, and sellers offering tools to remove AI labels from content.

China’s deep synthesis regulatory framework is the world’s oldest and most technically prescriptive. The Deep Synthesis Provisions took effect in January 2023 — 19 months before the EU AI Act entered into force. The 2025 updates (AI Content Labeling Measures + mandatory standard GB 45438-2025) added exact specifications for how every type of AI content must be marked. For the full regulatory landscape, see our China AI Regulation 2026 guide.

Key Takeaways

• Six categories of “deep synthesis” are regulated: text generation, speech synthesis, non-speech audio, biometric image/video (face/voice), non-biometric image/video, and 3D/virtual scenes. If your AI product falls into any of these, the rules apply.
• Three-tier labeling is mandatory: invisible metadata on ALL content (Art. 16), visible labels where confusion risk is high (Art. 17), and anti-tampering protections (Art. 18).
• Biometric consent must be “separate.” If you edit someone’s face or clone their voice, you need standalone consent — not bundled with general terms of service (Art. 14).
• GB 45438-2025 specifies exact watermark dimensions: image text height >= 5% shortest side, Morse “AI” audio marking, JSON metadata with AIGC fields.
• 13,421 accounts penalized in February 2026 for labeling violations. Enforcement is real and at scale.

The Three-Tier Labeling System

The Deep Synthesis Provisions (2023) established a layered labeling approach. The 2025 AI Content Labeling Measures and GB 45438-2025 added technical specifications.

Tier 1: Invisible Metadata on ALL Content (Art. 16)

Every piece of AI-generated or AI-modified content must carry invisible technical marks in its metadata. This applies to all six categories — no exceptions. The marks must be embedded in the file itself, not just displayed on the platform.

GB 45438-2025 specifies the format: JSON metadata with standardized AIGC (AI-Generated Content) fields. Digital watermarks are encouraged but not mandatory at this tier.

Tier 2: Visible Labels Where Confusion Risk Is High (Art. 17)

Five content types require additional visible labeling because they are most likely to be mistaken for real:

1. Text generation/dialogue — users must see an AI disclosure

2. Voice synthesis/cloning — audio must carry an AI label

3. Face generation/replacement/manipulation — video must show a watermark

4. Immersive virtual scenes — AI-generated environments must be disclosed

5. Other significant content modification — any material alteration

For images, GB 45438-2025 requires the visible label text to be at least 5% the height of the image’s shortest side. For audio, the standard specifies a Morse code “AI” rhythm embedded in the audio signal. For video, a persistent watermark throughout playback.

Tier 3: Anti-Tampering Protections (Art. 18)

Nobody may delete, alter, or forge the labels applied under Tiers 1 and 2. This is a separate prohibition — the February 2026 enforcement sweep specifically targeted sellers offering tools to strip AI labels from content.

Art. 10 of the Labeling Measures adds a user dimension: users must self-declare AI-generated content when posting. Platforms must verify these declarations.

The 2025 Updates: What Changed

The regulatory framework evolved in three stages:

Stage 1: Deep Synthesis Provisions (January 10, 2023). Established the definitions, the three-tier labeling concept, biometric consent requirements, and filing obligations. This was the foundation.

Stage 2: AI Content Labeling Measures (effective September 1, 2025). Extended labeling requirements beyond deep synthesis to ALL AI-generated content. Added platform verification duties, app store obligations, and the user self-declaration requirement. Published by four agencies (CAC, MIIT, MPS, SAMR).

Stage 3: GB 45438-2025 (effective September 1, 2025). The mandatory national standard that added technical specifications — exact watermark dimensions, audio marking format, JSON metadata schema. GB (not GB/T) designation means this has near-law authority.

The practical effect: before September 2025, labeling obligations existed but lacked technical specifics. After September 2025, there are exact pixel requirements. Compliance is measurable.

China vs EU vs Korea: Labeling Compared

Our view: China’s labeling regime is the global benchmark. It is the only system with published mandatory technical specifications, active enforcement, and a user opt-out mechanism. Companies building AI content labeling systems should build to GB 45438-2025 specifications — this exceeds what the EU and Korea currently require and positions you for compliance in all three markets.

Disclaimer: This content is for informational purposes only and does not constitute legal advice. Chinese AI regulation changes frequently. Reg Intel is not a law firm and does not provide legal services.

Last verified: April 9, 2026

Compare: EU vs China

For the global keystone comparison across twelve dimensions — algorithm filing vs conformity assessment, content moderation conflicts, asymmetric extraterritoriality, enforcement philosophy, and a five-step dual-market compliance baseline — see EU vs China AI Regulation: Two Systems, Two Philosophies (2026).