AI Disclosure Metadata for Commerce Video: Tag the Asset, Not the Campaign

AI disclosure metadata: the short answer

AI disclosure metadata is the machine-readable flag that says a video or image was made with generative AI, written into the asset itself rather than the campaign form around it. For commerce video teams in 2026, that flag is now a delivery requirement: platforms read it on upload, regulators expect it to be machine-readable, and an untagged synthetic performer can hold up a whole catalogue.

The change is easy to miss because nothing on the surface looks different. You still upload a file and you still fill in a form. What moved is where the truth lives. Amazon Ads' own guidance tells advertisers that when creative containing synthetic performers is produced outside Amazon, they must mark it during upload by selecting the Contains synthetic performers option in the Assets section, and that for third-party ad-served video the advertiser is responsible for including a compliant disclosure directly within the creative.

Read that twice, because it splits your pipeline in two. Assets generated inside the platform's own tools get labelled for you. Everything your team makes, which is most commercial video worth shipping, arrives as an unverified file that has to declare itself. Most teams already keep a {{link}} for the markets they ship into, but a jurisdiction list does not tell you which field to write inside the file, or which export will silently drop it.

Most teams already keep a 2026 labelling rules checklist for the markets they ship into, but a jurisdiction list does not tell you which field to write inside the file, or which export will silently drop it.

What the platform actually asks the advertiser to do

Strip the legal framing and an asset can only be in one of three states. It was made with the platform's generative tools and never left the console, in which case identification and disclosure happen automatically. It was made with those tools, then downloaded, edited and re-uploaded, in which case the automatic identification no longer applies and you must declare it yourself. Or it was made entirely in your own stack, in which case you declare it and, where you serve the creative, carry the disclosure in the frame.

The second state is the one that catches production teams. Generating a base clip in a platform tool and then grading, cutting or compositing it in your own timeline breaks the chain of custody the platform was relying on. The output is still synthetic, the file no longer carries the platform's evidence of it, and the obligation quietly transfers to you. Any workflow that treats a generated clip as an ingredient rather than a finished ad lands in this state by default, which is nearly every serious commercial pipeline.

The legal trigger is narrower than the panic around it suggests. New York's requirement, effective 9 June 2026, applies to photorealistic or AI-generated depictions of fictitious people shown to viewers in that state. A generated kitchen behind a real kettle is not a synthetic performer. A generated hand modelling a ring is. The test is whether the human in frame exists, not how much of the frame was generated.

The vocabulary that decides whether you tag at all

If you are going to write a machine-readable flag, write the one the standards bodies already defined. IPTC's Digital Source Type vocabulary separates media created algorithmically using an artificial intelligence model trained on captured content from media merely augmented, corrected or enhanced using a generative model, such as with inpainting or outpainting. It carries a third term for a composite where at least one element is generative, and a separate term for human editing done with non-generative tools.

Those terms map almost exactly onto the marketplace carve-outs. A real performer retouched with AI sits under human or generative editing rather than generative creation, which is why platforms exempt it. A wholly invented model sits under generative creation, which is why it must be flagged. The difference is not cosmetic, because it decides whether a shot needs consent paperwork, a disclosure, or nothing at all. That distinction is the same one a {{link}} already forces you to make before a generated face fronts a brand.

The practical benefit of borrowing the standard vocabulary is that one classification serves every channel. Write it once at the master and the same value answers a marketplace upload field, a European transparency obligation and an agency's own provenance record. Invent your own tagging scheme and you will re-litigate the same question in every export, usually under deadline, usually with a different answer.

That distinction is the same one a clearance workflow for generated faces already forces you to make before a generated face fronts a brand.

Why your encode chain quietly deletes the disclosure

Here is the part almost nobody plans for. Embedded and sidecar metadata is fragile. Transcodes, format conversions, aspect-ratio derivatives and platform re-encodes routinely drop fields they consider non-essential, so a keyword you wrote into the master can be simply absent from the nine-by-sixteen cutdown that actually shipped. Video suffers far more than stills here, because every placement gets its own encode and every encode is another chance to lose the flag.

Provenance standards acknowledge this directly. C2PA binds a manifest to content in two ways: a hard binding, which is a cryptographic hash of the bytes and therefore will not match a transcoded rendition, and a soft binding computed from the content itself, which the specification describes as useful for identifying derived assets and asset renditions. The manifest also carries a digital source type field inside its action assertions, so the same standard value can travel cryptographically rather than as a deletable keyword.

Two habits fix most of the leakage. Keep the highest-fidelity master as the single source of the classification, and generate every derivative through a tool chain you have actually tested for metadata retention rather than one you assume retains it. Then spot-check the output. Reading the metadata off a final file takes seconds and costs nothing next to a suppressed listing or a rejected ad.

The operational conclusion is unglamorous. Never treat the disclosure as done because you wrote it once. Verify it on the file you are about to upload, not the file you exported three steps earlier. This is where an {{link}} stops being paperwork and starts being the thing that keeps you shippable.

This is where an asset management schema for generative pipelines stops being paperwork and starts being the thing that keeps you shippable.

Where the disclosure gate belongs in the pipeline

Put the classification decision as early as possible and the verification as late as possible. Classification belongs at the shot level during generation, while the person who made the frame still knows whether the human in it exists. Ask that question two weeks later in a delivery scramble and you will get a guess. Verification belongs immediately before upload, after the final encode, on the exact deliverable rather than on its ancestor.

Between those two points, carry the value in your own records as well as in the file, because the file may lose it. Pair the tag with the {{link}}, because a marketplace flag proves compliance to one platform while a signed record proves it to everyone downstream, including the client who inherits the asset a year later.

Ownership matters as much as sequence. If the classification is nobody's named job it becomes the editor's problem at six in the evening on delivery day, which is exactly when it gets skipped. Give it to the person who signs off the cut, and make the answer a required field in the delivery ticket rather than a line in a wiki nobody opens.

The regulatory direction supports building this once rather than per platform. Article 50 of the EU AI Act places transparency duties on providers and deployers of systems that generate synthetic content, including marking outputs in a machine-readable form. Marketplace upload fields are one implementation of that idea, and they will not be the last one you have to satisfy.

Pair the tag with the provenance record buyers now request, because a marketplace flag proves compliance to one platform while a signed record proves it to everyone downstream, including the client who inherits the asset a year later.

A five-step asset-level disclosure gate

One: classify every generated shot at creation using the standard vocabulary, recording whether a photorealistic human was generated, a real human was AI-edited, or no human appears. Two: write the classification into the master's metadata and into your asset record, so losing one does not lose both. Three: after every encode, re-read the deliverable and confirm the field survived the round trip.

Four: at upload, complete the platform's own declaration rather than assuming your embedded metadata satisfies it, because the two are separate mechanisms and a platform will honour its own field first. Five: for any placement where you serve the creative yourself, put a compliant disclosure in the frame, since there is no upload form standing between you and the viewer.

None of this makes AI video harder to make. It makes it possible to ship at volume without a compliance review blocking every batch. For any {{link}}, the gate has to run per SKU and per locale, or a single untagged master will propagate the same defect across thousands of live listings before anyone notices.

The teams that move fastest through the next two years of rules will not be the ones with the best legal memo. They will be the ones whose pipeline cannot export an untagged synthetic human in the first place.

For any catalogue-scale product video programme, the gate has to run per SKU and per locale, or a single untagged master will propagate the same defect across thousands of live listings before anyone notices.