Key takeaways
- Amazon requires the keyword contains-synthetic-performer in the dc:subject field of AI-generated images.
- New York law mandates disclosure for synthetic performers since June 9, 2026, with penalties up to $5,000.
- Tag only photorealistic AI-generated people, not real or stylized images, to avoid false AI notices.
- Image metadata can be stripped during resizing or export, so verify the final upload file retains the tag.
- California's AI Transparency Act starts Jan 1, 2027, requiring platforms to detect AI provenance metadata.
How to tag AI generated images on Amazon is now a file question, not a listing question. If a product image shows a photorealistic person generated with AI, Amazon expects the keyword contains-synthetic-performer inside the image file itself, in the dc:subject XMP field, added with an IPTC-compatible metadata editor, before the upload.
Amazon notified sellers on July 22, 2026. The rule it implements is older, and it is not Amazon rulemaking.
The law behind the Amazon AI image tag
New York amended General Business Law section 396-b through S.8420-A, signed on December 11, 2025 and in force since June 9, 2026. It requires a conspicuous disclosure when an advertisement features a synthetic performer: a digitally created figure, made with generative AI, built to read as a human performance and not recognisable as any identifiable real performer.
The penalty is 1,000 dollars for a first violation and 5,000 dollars for each one after it.
The part that matters operationally is who is on the hook. Legal analysis of the statute puts the obligation on the advertiser and exempts media publishers that only distribute the advertisement. On a marketplace detail page, the advertiser is the seller. Amazon built the mechanism. The exposure sits with you.
What to tag and what to leave alone
Tag it when the image contains a photorealistic person generated with AI.
Leave it alone for edited photographs of real people, for characters from film, television or games, for stylised or non-photorealistic characters, and for any image with no person in it. Reporting on the rollout lists the same exclusions.
Over-tagging is not the safe default it looks like. Amazon adds a disclosure to the listing where the keyword applies, so tagging a photograph of a real model puts an AI notice on a real photograph. The keyword is not a private flag buried in a file. It becomes copy on your detail page.
The pipeline problem nobody checks
XMP metadata survives a copy. It does not always survive a workflow.
Resize scripts, export presets and asset pipelines strip metadata on the way out, because for years metadata was weight to be shaved. If the master file carries the keyword and the derivative you actually upload does not, you have a gap that no listing report will surface, because nothing about the listing looks wrong.
The check is small and worth running once per pipeline. Open the exact file you upload, not the master, and confirm the keyword is still in dc:subject. Then run it again after the next change to your export settings.
The next date is January 1, 2027
The California AI Transparency Act, AB 853, builds the other half of the same system on a staggered schedule. Generative AI providers above one million monthly users have had to embed latent provenance metadata since August 2, 2026: system name, version, and the date the content was created or altered. Large online platforms have to detect that data and surface it to users from January 1, 2027. Capture device manufacturers follow on January 1, 2028. The penalty is 5,000 dollars per violation.
Read it forward. The tools that generate your imagery are already writing provenance into the file. From 2027, large platforms have to read it and show it. The keyword you add by hand today is the manual first layer of something that is going to be machine-read.
The principle underneath
Compliance has moved inside the asset.
For years a listing was compliant when the right fields on the right page held the right values. That is no longer where the check lives. The GPSR responsible person requirement made the same move: a filled field was never the whole test, and the documentation behind it is what gets examined. EPR registration numbers went the same way, from a form to a verified identifier per country.
Your image file is now a compliance document. It carries a declaration, that declaration is legally yours, and a resize script can delete it without anyone touching your account. Treat the creative pipeline the way you treat the compliance folder, and audit the file that ships rather than the one you kept.
And do not assume a detail page stays as you left it. Amazon already rewrites titles over 75 characters when the seller does not act in time. Listing content is increasingly something both parties edit.
Sources
- Amazon Seller Central Forums - Adding contains-synthetic-performer on AI created images sellercentral.amazon.com
- New York State Senate - Bill S8420-A nysenate.gov
- eWeek - Amazon Requires Sellers to Disclose AI-Generated People in Product Images eweek.com
- Cooley - New York Enacts Synthetic Performer Disclosure Law for Advertisements cooley.com
- Morgan Lewis - New California AI Disclosure Rules Become Operative morganlewis.com
Hey! We're the SellerKey Team
Data & product at SellerKey.co
📊 We build SellerKey.
The team behind SellerKey. We write about fees, margin, inventory and ads for Amazon sellers.
