Short answer. There is no single global rule, but the three regimes that matter converge on two mechanics: a machine-readable mark embedded in the file, and a human-visible disclosure wherever a person could be misled. The EU AI Act's Article 50 transparency obligations apply from 2 August 2026 and cover synthetic audio, image, video and text plus a deepfake disclosure duty. China's labelling Measures, in force since 1 September 2025, require both explicit visible labels and implicit metadata labels, in more prescriptive detail. The United States has no comprehensive federal statute — California's AI Transparency Act becomes operative 2 August 2026 for large providers, and the FTC polices deceptive practice under existing authority. For a publisher shipping worldwide, the workable policy is to mark everything and disclose wherever a reasonable viewer might be deceived.
Most compliance failures here are failures of stage rather than intent: marking was left to the publishing team instead of the production pipeline, and by the time anyone asked, the generation context was gone. This guide summarises what each regime requires, where responsibility splits between model builder and publisher, and what a single global policy looks like. It is a practitioner's summary, not legal advice.
What the three regimes have in common
Read side by side, the three instruments are less different than their drafting suggests. Each separates a technical duty from a communicative one. The technical duty is to embed something in the file a machine can read — metadata, a watermark, a cryptographic manifest — so artificial origin travels with the content. The communicative duty is to tell a human being, in a form they will notice, when they are looking at something synthetic they might otherwise take as real.
Each also splits responsibility along the supply chain: the party that builds the generative system carries the marking duty, the party that publishes the output carries the disclosure duty. Most brands are in the second category, which is the source of the most common error here — assuming that because a model vendor watermarks its outputs, the publisher has nothing left to do.
| Regime | In force | Core duties |
|---|---|---|
| EU AI Act, Article 50 | 2 August 2026 | Providers mark synthetic audio, image, video and text machine-readably; deployers disclose deepfakes and AI-generated text published to inform the public on matters of public interest; chatbots reveal they are machines |
| China — CAC Labelling Measures + GB 45438-2025 | 1 September 2025 | Explicit labels visible to users, plus implicit technical markers in file metadata; duties on information service providers and distribution platforms |
| California AI Transparency Act (SB 942, amended by AB 853) | 2 August 2026 (platforms from 1 January 2027) | Covered providers above one million monthly users: a free AI-detection tool, an optional visible disclosure, and latent provenance in generated image, video and audio |
Those are the operative headlines, not the full scope of any instrument; the carve-outs decide real cases.
What does EU AI Act Article 50 require?
Article 50 sets transparency obligations for particular categories of system rather than for AI generally. Four duties matter to a content producer. Systems interacting directly with people must make clear the person is dealing with an AI system. Providers of systems generating synthetic audio, image, video or text must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. Deployers who generate or manipulate deepfake content must disclose it. And deployers publishing AI-generated text to inform the public on matters of public interest must disclose that too, unless the content underwent human review with editorial responsibility retained by a person or organisation.
The definition of deepfake is broader than popular usage. It covers AI-generated or manipulated image, audio or video content resembling existing persons, objects, places, entities or events, which would falsely appear to a person to be authentic or truthful. That reaches a synthetic voice of a real spokesperson, an AI-generated shot of a real building, and a manipulated recording of a real event — not only face-swapped video of a public figure. Disclosure has to reach the viewer at first exposure at the latest, clearly and distinguishably.
The obligations apply from 2 August 2026. The European Commission has indicated a transition for the marking obligation in respect of generative systems placed on the market before that date, and the AI Office has been finalising a code of practice on marking and labelling as a compliance route.
The exemption most often missed: the marking duty does not bite where the AI system performs an assistive function for standard editing and does not substantially alter the input data or its semantics. Colour grading and noise reduction sit comfortably inside that. Replacing a background, changing what a person appears to say, or generating a shot that was never filmed do not.
Why is China's regime stricter?
China's regime is older and more specific. In March 2025 the Cyberspace Administration of China, with the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration, released the Measures for Labeling AI-Generated Synthetic Content together with the mandatory national standard GB 45438-2025. Both took effect 1 September 2025.
The Measures require two kinds of label. Explicit labels are perceivable by users — text, audio cues or graphics telling a person the content is AI-generated, such as an "AI-generated" marker on a piece of text or a corner label on an image. Implicit labels are technical markers added to the file's metadata, not readily perceivable, recording the synthetic origin and the provider. The obligations run to internet information service providers and to platforms distributing online content, so a distribution platform carries its own detection and labelling responsibilities rather than relying on the uploader.
The consequence for a company producing content for Chinese platforms is that labelling cannot be a publishing afterthought. The implicit label has to be written when the file is produced, must survive handoff, and is what a platform's automated check looks for. Legal commentary comparing the two approaches identifies this as the main difference: the EU states an outcome, China specifies the mechanism.
Is there a US federal labelling law?
Not a comprehensive one as of mid-2026. Three overlapping sources of obligation exist instead, and treating any one as the whole picture is a mistake.
State legislation. California's AI Transparency Act (SB 942), as amended by AB 853, is the most consequential for content producers, and becomes operative 2 August 2026. Covered providers — generative AI systems with more than one million monthly visitors or users, publicly accessible in California — must offer a free public AI-detection tool for their own outputs, offer users the option of a visible disclosure, and embed latent machine-readable provenance data in generated image, video and audio. AB 853 extends the framework to large online platforms, which from 1 January 2027 must surface provenance data and not knowingly strip it, and eventually to capture-device makers.
Conduct rules. The Federal Trade Commission's long-standing authority over unfair or deceptive acts or practices reaches undisclosed synthetic endorsements, fabricated testimonials and misleading AI-generated claims, with no AI-specific statute needed.
So the question for a US publisher is rarely "is there a labelling law here" and usually "would a reasonable consumer be misled if we did not say this" — a standard older than generative AI that does not wait for a statute.
A publishing policy that satisfies all three
The cheapest defensible position for an organisation publishing across markets is one global policy meeting the strictest applicable requirement. Obligations attach to the file, and files travel.
- Classify every asset by how synthetic it is. Three buckets: AI-assisted editing of real material; AI-generated material depicting no real people or events; and AI-generated or manipulated material resembling real persons, places or events. Only the third is a deepfake in the Article 50 sense.
- Mark at export, not at publication. Write machine-readable provenance into the file as it leaves production. Marking at publication means trusting every downstream distribution path to do it, and one of them will not.
- Decide the visible disclosure by audience risk, then apply it globally. Keeping a labelled and an unlabelled cut of one file is how the unlabelled one reaches the wrong market.
- Make the deepfake rule a hard gate. Any asset depicting a recognisable real person, place or event that was generated or materially manipulated needs disclosure at first exposure and documented consent from anyone depicted.
- Audit the pipeline for metadata stripping. Editing, transcoding and platform upload mostly strip it. Where provenance cannot survive, record the chain of custody yourself.
- Keep a per-asset record. Which model, when, under which licence, what human review, what labels, and where it was published.
The mechanics of provenance itself — what the record must contain, how oversight is specified, how a disclosure position is set centrally — are covered in AI content governance, disclosure and provenance.
What this actually changes in production
Less than the commentary suggests, provided the work happens at the right stage. Marking is a one-time pipeline change at export; visible disclosure is a design decision made once per asset class. The expensive version is the retrofit — discovering after publication that a library of localised variants needs a label, or that nobody recorded which of them used a synthetic voice. For content produced before a policy existed, back-marking is usually impossible because the generation context is gone: re-export from source where it survives, add visible disclosure where it does not, or retire the asset.
How Lifewood approaches this
Lifewood produces AI-generated content commercially and so operates under these rules itself rather than commenting from outside. Marking and disclosure metadata are applied at the delivery stage, the last point at which one process touches every language variant in a batch — and across 50+ languages and 40+ delivery centres in 30+ countries, per-market labelling policies fail at exactly the handoffs a single export-stage rule covers. For a library being retrofitted, the first step is an inventory of what can still be re-exported from source: that determines what is fixable and what has to be retired. See AIGC services, AIGC video production and the delivery methodology.
Sources and further reading
- Article 50, Transparency Obligations — EU Artificial Intelligence Act (consolidated text), and the European Commission's Article 50 FAQ.
- Measures for Labeling of AI-Generated Synthetic Content (English translation) — China Law Translate, March 2025; commentary from Loeb & Loeb LLP and Bird & Bird, 2025.
- SB 942 — California Legislative Information, 2024; AB 853 — CalMatters Digital Democracy, 2025.