AI-generated model photos went from novelty to normal in about two years, and the rules are now catching up. In 2026 two separate regimes started asking sellers to disclose AI-generated imagery — not to stop using it, just to label it. Most sellers have no idea either exists, because the labels live inside the image file where no photo editor shows them.
This guide covers what each rule actually says, who it applies to, what the label technically is, and how to write it — including a free browser tool that does it in seconds.
The New York Law Behind It: GBL §396-b
New York General Business Law §396-b, in force since 9 June 2026, requires clear and conspicuous disclosure when advertising uses a synthetic performer — an AI-generated human likeness that a reasonable person could mistake for a real one. The penalties are per violation: $1,000 for the first, $5,000 for each subsequent one.
A product listing image showing a photorealistic AI-generated model wearing your garment is advertising with a synthetic performer, which is why the law reaches ordinary ecommerce sellers rather than only ad agencies. And because marketplaces can't police one state's buyers separately, the practical enforcement happens at the platform level — which brings us to Amazon.
Amazon's Implementation: a Keyword Inside the File
Since 23 July 2026, Amazon requires that any listing image containing a photorealistic AI-generated person carry the keyword contains-synthetic-performer in the image's embedded XMP metadata (the dc:subject field) — worldwide, not just for New York.
Three things about the scope, because they decide whether this applies to a given image:
- It is about people, not AI in general. A flat lay, a ghost-mannequin shot or a product-only image needs no synthetic-performer keyword, even if AI edited the background.
- Real people are exempt. If the model in the photo is a real person you photographed — or an AI image built from a real person you have rights to — the image is not a synthetic performer.
- It is invisible. The keyword is metadata, not a watermark. Buyers never see it; the platform and regulators can read it.
dc:subject XMP editing, the AI image generators don't write it, and a browser's canvas strips whatever metadata a file had. Sellers who are trying to comply usually can't find the field they're supposed to fill in.
Google's Parallel Rule: Provenance, Not Performers
Separately, Google Merchant Center asks that AI-generated or AI-edited product imagery retain the IPTC provenance term DigitalSourceType: trainedAlgorithmicMedia. This one applies to AI-generated imagery regardless of whether a person appears — it marks how the image was made, not what it shows. Same problem as Amazon's keyword: it lives in metadata nobody's editor exposes, and it must survive your export pipeline.
What Exactly Gets Written
| Tag | Where it lives | When it applies | Asked for by |
|---|---|---|---|
contains-synthetic-performer | XMP dc:subject (keywords) | Image shows a photorealistic AI-generated person | Amazon (since 23 Jul 2026), implementing NY GBL §396-b |
DigitalSourceType: trainedAlgorithmicMedia | XMP Iptc4xmpExt | Image is AI-generated or AI-edited (people or not) | Google Merchant Center |
How to Comply, in Practice
If your images come from CatalogX
You're done — every image exported from CatalogX carries the IPTC provenance term, and images showing an AI-generated person also carry Amazon's keyword. It's written on both the app's downloads and the API's output, free on every plan. The mechanics are described on the AI disclosure & provenance feature page.
If your images come from anywhere else
Use the free AI Image Disclosure Tagger: drop in a JPEG, say whether it shows an AI-generated person, and download the same file back with the correct tags written. It runs entirely in your browser — no signup, no upload, and the pixels are untouched; only metadata is added.
Two honest caveats
- Labels can be stripped downstream. Metadata survives copying, uploading and downloading, but an editor that re-saves the file may discard it. Tag as the last step before upload.
- Disclosure is not accuracy. The tag says how the image was made; marketplace policy still requires that the image truthfully shows the product being shipped. An AI image of a slightly different garment is a policy problem no metadata can fix — which is why garment fidelity, not just disclosure, is the thing to demand from an AI photo tool.
Writes Amazon's synthetic-performer keyword and the IPTC provenance term into your JPEG. No signup, images never leave your device.
Open the free tagger →
This Is Not Legal Advice
This article describes published platform policies and one state statute as they stood on 15 September 2026, and what CatalogX's tools mechanically write into files. Rules differ by jurisdiction and change over time; your compliance position is yours to confirm, ideally with counsel if meaningful revenue rides on it. What we can guarantee is the mechanical part: the tags, written correctly, into your files.