China requires explicit and implicit AI labels
China's labelling measures have required both a visible label and a metadata or watermark label since 1 September 2025. What to attach before you distribute.
Almost every AI labelling framework in the world asks for one label. It either wants something a human can see, or something a machine can read, and the arguments in each jurisdiction are about which one counts. China's Measures for Labeling AI-Generated Synthetic Content resolved that argument by refusing it. Since 1 September 2025, alongside a mandatory national standard, the requirement has been both: an explicit label a person can perceive, and an implicit label carried inside the file.
That date is worth sitting with. It is eleven months earlier than the EU AI Act's Article 50 obligations, which applied on 2 August 2026, and earlier than California's marking obligation, which landed on the same day after being pushed back. If you distribute AI-assisted creative into China at all, the deadline you have been diarising for the EU passed a year ago somewhere else.
Two labels, two failure modes
The pairing is not bureaucratic duplication. The two labels fail in opposite circumstances, which is exactly why requiring one is a weaker rule than requiring both.
An explicit label is perceptible to the person consuming the content: text on or around the media, an audio announcement, a visible marker. It survives everything that happens to a file downstream, because it is part of what the file is. Screenshot it, re-encode it, repost it to a different platform, and a burned-in line of text is still there. What an explicit label cannot do is be read at scale. No platform's ingest pipeline reliably parses a sentence rendered into pixels, and nothing about it is machine-verifiable.
An implicit label lives in the file's metadata or in a watermark embedded in the content. It is readable at scale by any system that knows to look. It also disappears without a trace the first time something touches the file that was not built to preserve it — and that includes ordinary, entirely innocent operations. Metadata is stripped by most re-encodes. A screenshot destroys it completely. Passing a file through a chat app before uploading often loses it.
Put them side by side and the design intent becomes obvious.
| Explicit label | Implicit label | |
|---|---|---|
| Audience | The viewer | Platforms, regulators, downstream tools |
| Lives in | Visible pixels or audible audio | Metadata or an embedded watermark |
| Survives a screenshot | Yes | Metadata, no. A robust watermark, sometimes |
| Survives a platform re-encode | Yes | Metadata, usually not. Watermark, depends on implementation |
| Machine-readable at scale | No | Yes |
| Verifiable | No | Depending on how it was applied |
Neither column is sufficient on its own. That is the whole argument for requiring both, and it is the same argument the European Commission has been making from the other direction when it says no single technique currently satisfies the machine-readable marking standard on its own.
What "implicit" actually asks of your toolchain
The explicit label is a production task. Someone adds a line of text, or a spoken notice, and it ships. It is annoying, it competes with your creative, and it is entirely within your control.
The implicit label is a supply-chain problem, and this is where teams discover they do not know what their own pipeline does. The label has to exist in the delivered file. Whether it does depends on a chain of tools, most of which you did not select for this property:
- Whether the generating model writes anything into the output at all.
- Whether the file format you export to can carry it.
- Whether your editor preserves what arrived rather than discarding it on ingest.
- Whether your export settings strip metadata as a size optimisation.
- Whether anything between export and upload re-encodes the file.
Any one of those steps can silently drop the record, and several routinely do. The reason this feels invisible is that nothing warns you. A stripped file is not marked as stripped. It looks identical to a file that never carried a label, which is the state a lot of published work is actually in without anyone having decided that.
The practical consequence: "we use a model that supports labelling" is not a compliance position. "We inspected the delivered file and the label was present" is. Those are different sentences and only the second one is checkable.
Before you distribute into China
A workable pre-flight, in order:
- Decide the explicit label copy once, per format. One short line for vertical video, one for square, one for static images, one spoken variant for audio-only. Reuse them. A team that rewrites the disclosure per asset produces inconsistent disclosure, which reads worse than a plain repeated one.
- Burn it in rather than relying on a platform toggle. A text overlay rendered into the frame travels with the file. A platform-side label does not survive the file leaving that platform, and a caption does not travel at all.
- Inspect the actual export for the implicit label. Not the project, not the source generation. The file you are about to hand over. This is a thirty-second check and it is the only step that tells you the truth.
- Find the hop where it dies, if it is missing. Export at each stage and inspect each one. The first stage where the label disappears is your culprit, and in most pipelines it is a re-encode nobody thought of as an edit.
- Record what you attached, per asset. Not for the platform. For yourself, six months later, when someone asks what a specific published file carried.
Worth stating plainly for anyone assembling in Versely: there are no watermarks on any plan, which is a deliberate output-quality choice and means there is no Versely-applied mark to rely on as your implicit label. The visible label is your job, and the editor is where you add it. What arrives on source material from upstream tools is a separate question, and the inspection step above is how you answer it rather than assume it.
Three regimes, one export
The temptation is to build a China-specific pipeline. Resist it, because the requirements overlap more than they conflict.
| Regime | In force | Visible label | Machine-readable marking |
|---|---|---|---|
| China labelling measures | 1 Sep 2025 | Required | Required |
| EU AI Act Article 50 | 2 Aug 2026 | Deployers label deepfakes on first exposure | Providers mark synthetic output |
| California AI Transparency Act (SB 942, as amended by AB 853) | 2 Aug 2026 | — | Latent provenance marking by covered providers |
A file that carries a burned-in disclosure and an intact machine-readable mark satisfies the visible and the marking limb in all three columns. It is one export discipline, not three. The differences that remain are about who owes the duty — provider versus deployer is the distinction that decides whether a given obligation is yours at all — and about timing, where the EU gave systems already on the market before 2 August 2026 until 2 December 2026 for the marking obligation specifically.
For anyone tracking why so much of this converges on China first, the reason video models tend to launch there ahead of Western markets and the reason its labelling rules arrived first are the same reason: the regulatory and release cycles are running on a different clock.
FAQ
Does an implicit label mean a C2PA manifest specifically?
Not necessarily. The requirement described is for a label carried in metadata or in a watermark. C2PA Content Credentials are one well-known way to satisfy a metadata-borne marking requirement and are widely deployed, but the measures are written around the function rather than around one consortium's specification. If you need to know which specific implementations are accepted, that comes from the mandatory national standard rather than from any summary of it, including this one.
If my video already says "AI-generated" in the caption, is that the explicit label?
A platform caption is weak for this and weaker still once the file moves. Captions do not travel with the asset, they are not present when the video is downloaded or reposted, and on several platforms they are collapsed by default. A burned-in line is the version that holds. Use the caption too — two signals cost nothing — but do not make the caption the only one.
Does this apply to content that is only lightly AI-assisted?
That threshold question is exactly the part I cannot resolve from an English translation, and the honest answer is that "how much AI triggers the duty" is decided by the measures' own definitions and the accompanying standard, not by intuition. If your work sits near the line — a live-action plate with a generated background, or a real presenter with a synthesised voice — that is a question for someone reading the Chinese text, not for a checklist.
Do I need a different export for each market?
Usually not. The overlap table above is the reason: a burned-in visible disclosure plus intact machine-readable marking is the superset. The place a per-market export becomes genuinely necessary is language — an explicit label a viewer cannot read is not a label — and that is a localisation task rather than a compliance architecture one.