Labelling regimes compared — EU AI Act Article 50 vs China's labelling measures

English · 中文

graded per item, sources below·checked 2026-10-11

One question: when AI generates content, how does each side make it show? Short answer: China turned the technical format into a mandatory national standard — finer-grained, and earlier. The EU wrote the duty into its main statute and left the detail to guidance — broader in scope.

Evidence grades (defined in 00-overview.md): A primary text, checked live; B other official source or a direct reading of the primary text; C our own comparison or inference, or not yet checked.

1. The table

2. Four observations you can use directly

2.1 Both are two-tier — but the second tier differs

• China: the measures set the duty and name the elements; a mandatory standard sets the method. What an implicit label must carry is listed by the measures themselves — attribute information, the provider's name or code, and a content identifier, in two sets for production (Article 5) and for distribution (Article 6) — and Article 11 makes conformance to mandatory national standards binding, which is how GB 45438-2025 attaches: published 2025-02-28, in force 2025-09-01 alongside the measures, with 中央网络安全和信息化委员会办公室 as both 主管部门 and 归口部门. The drafters' Q&A describes the standard as setting out the 具体实施方式和操作方法 of the mandatory part, including a reserved metadata field for protection information such as label integrity. Format non-compliance is therefore measurable; the literal key names are not carded here (see §4). — A
• EU: the regulation sets the duty, Commission guidance sets the detail. The guidance is interpretive and does not create new obligations; firms can also show compliance through a voluntary code. — B
• In practice: in China you can be found non-compliant on format; in the EU the argument is more often about whether disclosure happened at all. — C

2.2 The effective dates are almost eleven months apart

• China: 2025-09-01. — A
• EU: applies from 2026-08-02, with a grace period until 2026-12-02 for machine-readable marking of generative systems already placed on the market before that date. That grace period was introduced when the AI Act was amended by the Digital Omnibus on AI (Regulation (EU) 2026/1744). — A
• So on digital labelling, China went first — which cuts against the common intuition that EU legislation is always ahead. — C

2.3 Different centre of gravity

• China aims at content form (text/image/audio/video/virtual scene) and the distribution chain (generator → distribution platform → app store → user). — A
• The EU aims at use situation (chatbots, emotion recognition, biometric categorisation, deep fakes) — "you must know when you are talking to AI" counts as transparency. — C

2.4 China labels the relationship, not only the output

• Order No. 21 attaches the labelling duty to a service and then builds conduct duties around it: no virtual relatives or virtual partners for minors (Article 14), no service goal of replacing social interaction or inducing dependency (Article 10), an exit that must not be blocked by continued interaction (Article 19). — A
• The EU's Article 50 disclosure tells you the counterpart is a machine and stops there; what the machine then does to you is regulated by the AI Act's risk tiers, not by the transparency clause. — A
• So on AI companions the two regimes are not "stricter" and "looser": they take different objects. China regulates the provider's influence on the user; the EU regulates whether you are informed. — C

3. What it means for an ordinary person

• In China: an unlabelled AI-generated video can plausibly be found non-compliant, with a mandatory standard behind it. — B
• In the EU: you must be told when you are subject to emotion recognition, or shown that a deep fake is a deep fake. — A
• Common limit: both can reach platforms at home; neither reaches the same clip hosted abroad. — C

4. Limits and unverified ⚠

• The EU penalty tier is confirmed as Article 99(4) (EUR 15m or 3%); the scope of the 2026-12-02 grace period rests on Regulation (EU) 2026/1744 rather than the original AI Act text. — A
• Chinese penalty provisions are carded now, and the two instruments are not alike. Article 21 of the Generative AI Interim Measures enumerates the Cybersecurity Law, Data Security Law, PIPL and Science and Technology Progress Law and then supplies its own ladder — warning, circulated criticism, order to correct, suspension — for cases the named laws do not reach. Article 22 of the Deep Synthesis Provisions merely refers to the relevant laws and administrative regulations and aggravates where serious consequences result; it has no fallback of its own. Neither states an amount, so the ceilings sit in the general laws — PIPL Article 66 (RMB 50m or 5% of prior-year turnover) and the amended Cybersecurity Law (RMB 10m). — A
• GB 45438-2025's interior is not carded, but its shape is. The element classes come from Articles 5 and 6 of the measures, and the standard is described officially in the drafters' Q&A (questions 4–5: explicit-label methods per content type, the metadata implicit-label method, and a reserved field for 标识完整性 and 内容一致性), which is why this sits at B rather than C. Still unverified: the literal metadata key names, their encodings, and per-format parameters such as label duration and size — openstd serves this standard only as page images, and while the tiles are reachable by script (re-checked 2026-10-12) what comes back is a picture, not text that could be snapshotted. ⚠
• Anthropomorphic services have their own penalty bands now: Article 30 of Order No. 21 reaches RMB 10,000–100,000 for refusal to correct or serious cases, and RMB 100,000–200,000 where harm to life or health results. Both sit far below the Article 99(4) tier of the AI Act, and below the amended Cybersecurity Law's RMB 10 million ceiling. — A

5. Official sources

• EU, quick facts on transparency rules: https://digital-strategy.ec.europa.eu/en/factpages/quick-facts-transparency-rules-ai-systems
• EU, Commission guidelines on Article 50 transparency obligations: https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations
• EU, AI Act consolidated text (Article 50, Article 99): https://eur-lex.europa.eu/eli/reg/2024/1689/oj
• EU, Digital Omnibus on AI, Regulation (EU) 2026/1744: https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng
• China, Q&A on the labelling measures: https://www.cac.gov.cn/2025-03/14/c_1743654685896173.htm
• China, GB 45438-2025 full text: https://openstd.samr.gov.cn/bzgk/std/newGbInfo?hcno=F32EA2A561F1886CD8D606513512D547
• China, expert explainer: https://www.cac.gov.cn/2025-09/05/c_1758792061408012.htm
• China, Generative AI Interim Measures (Order No. 15), full text: https://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm
• China, Deep Synthesis Provisions (Order No. 12), full text: https://www.cac.gov.cn/2022-12/11/c_1672221949354811.htm
• China, Anthropomorphic Interaction Services Measures (Order No. 21), full text: https://www.cac.gov.cn/2026-04/10/c_1777558395078289.htm
• China, the same text in the State Council Gazette: https://www.gov.cn/gongbao/2026/issue_12806/202606/content_7072472.html