Amazon Makes Sellers Tag AI Generated People in Listing Images

A single state law in New York has just rewritten how every Amazon seller on the planet has to handle a generated face. On Wednesday 22 July 2026, Amazon notified sellers that any product image, product video or A+ Content asset containing a photorealistic person generated entirely by artificial intelligence must now carry a machine-readable disclosure in its metadata, written by the seller, before the file is uploaded. The policy went out as documentation rather than as a seller-wide alert, which is the first reason most operators have not registered that it exists.

The operational point is small and unforgiving. This is not a checkbox in Seller Central, not a declaration in a listing feed, and not something Amazon detects for you. It is a keyword you write into the file itself, and if it is missing, the asset is simply non-compliant. For European operators the temptation is to file this under American problems. That would be a mistake, and the reason is in the scope.

The mechanic: one exact string, in one exact field

The requirement is to open the asset in an editor that supports IPTC metadata and write the keyword contains-synthetic-performer into the dc:subject field of the image’s XMP metadata, before upload. Stripped of the terminology, that is one IPTC-compatible keyword, flagging a synthetic performer, embedded in every qualifying asset ahead of the upload.

Three details in that sentence carry all the operational weight.

First, the string is exact. What is documented is a single keyword: no variants, no synonym list, no plural form. A tag that reads “synthetic-performer” or “contains synthetic performer” is not the tag Amazon asked for.

Second, the field is exact. XMP dc:subject is the keywords field, which means the tag lives in the same place as whatever taxonomy your creative team already writes there. It also means the tag survives or dies with your export settings: any pipeline step that strips metadata on save, resizes through a tool that flattens XMP, or rebuilds the file from a design source will quietly remove the disclosure you just added.

Third, the coverage is broad. The requirement applies to product images, product videos and A+ Content. A+ Content matters here more than it looks, because A+ modules are where generated lifestyle scenes with human models most often live, and because A+ assets are frequently produced by an agency rather than by the seller who is accountable for them.

The exclusions are narrower than sellers assume

The rule targets one thing: photorealistic people generated entirely by AI. The carve-outs around it are specific, and there are four of them. Fictional characters from film, television or video games do not require the tag. Real people do not require the tag, even when their appearance has been altered or enhanced with AI tooling. Nor do assets with no people in them, or synthetic people rendered in a clearly illustrated, non-photorealistic style.

Read that list again from the perspective of a seller who has spent the last two years building a creative workflow, because the boundary is thinner than the summary suggests. A photograph of a real model whose skin, hair or background was cleaned up with a generative tool sits outside the rule. A fully generated model who looks like a photograph sits inside it. A stylised, obviously illustrated character sits outside it. The dividing line is not “did you use AI”, it is “is there a photorealistic human here who does not exist”. That distinction is a judgement call, and it is being delegated to whoever exports the file.

Why a New York statute reaches a European seller

The legal trigger is New York General Business Law section 396-b, amended by S.8420-A / A.8887-B, in force since 9 June 2026. It requires disclosure of synthetic performers in advertising. The penalties are 1,000 US dollars for a first violation and 5,000 US dollars for each subsequent one.

Amazon’s own framing is minimal, a single line: “recent legislation requires disclosure when images or videos in advertisements contain photorealistic AI-generated people.” Amazon has declined to say more about how the requirement will be implemented, so that one sentence is the whole of the company’s public reasoning.

Here is the part that converts a New York statute into a Madrid, Milan or Manchester problem. Amazon is applying the requirement across its worldwide stores, not only in the United States. Amazon did not build a New York-only compliance path; it built a platform policy and pointed it at a state law. That is the pattern European operators should recognise from GDPR going the other direction: the strictest jurisdiction in a global catalogue sets the default for everyone, because maintaining two asset libraries is more expensive than maintaining one.

So the practical reading for a European seller is that this is a global platform requirement with a US legal origin, and the fine schedule is the least interesting part of it. The listing-level consequence of a non-compliant asset on a marketplace that suppresses content it cannot verify is the part worth planning around, and it is also the part Amazon has not yet described.

The reason your European colleagues have not heard about this

Several operators have asked us the same question this week: if this is a global requirement, why did nothing arrive in our inbox? The answer explains most of the confusion around this story.

The rule lives on a Seller Central help page titled “how to tag media that contains an AI-generated person”. It was not pushed as a general notification to every seller account. It is documentation, and documentation is pull, not push: you find it if you go looking, or if a consultant, an aggregator newsletter or the trade press puts it in front of you. The seller-side guidance now circulating on the policy says as much in plain terms: “this is a brand-new policy, so most sellers haven’t seen it yet.”

For a European seller that produces creative in-house or through a local agency, this is the whole risk. The requirement is enforceable in principle from the moment it is published, but the discovery mechanism is a help article in a system most sellers only open when something breaks. Nobody is going to tell your retoucher.

The tension Amazon has not resolved

The awkward part of this announcement is that Amazon is on both sides of it. Amazon sells generative AI creative tools to the same sellers it is now asking to label the output. And by Amazon’s own measurement, that output works: Amazon’s own internal US measurements across 2024 and 2025, reported by Forbes, show Sponsored Brands campaigns using AI-generated images delivering a 10.3% higher return on ad spend than campaigns without them.

That is a genuine bind rather than hypocrisy. Amazon built the tool, published the performance number, and now has to comply with a law that says the tool’s output must be declared. But it does leave sellers holding an unpriced question: if a shopper-visible marker eventually appears on listings, does the 10.3% advantage survive the disclosure?

Because that marker is coming. Amazon has said it plans to add a visible indicator on qualifying listings, and has given no detail on timing or on the exact criteria that trigger it. That is the single most consequential open item in this story, and it is the one Amazon has said least about.

What is confirmed, and what is not

Confirmed: the keyword, the field, the asset types, the exclusions, the New York statute and its penalties, the worldwide application, and Amazon’s intention to show a visible indicator. Not confirmed: when the indicator appears, what it will say, how Amazon will detect an untagged synthetic performer, and what happens to an asset that should have been tagged and was not. Amazon has said nothing about the enforcement mechanism, so anyone describing it to you is speculating.

The honest summary for an operator this week is that the compliance action is cheap and the uncertainty is not. Writing one keyword into one metadata field costs minutes per asset. Discovering in three months that your export pipeline has been silently stripping it, across a catalogue built with generated models, costs considerably more.

Sources

• Forbes, “Amazon Requires Sellers To Label AI-Generated People In Listing Images”, 25 July 2026: https://www.forbes.com/sites/gabrielalinzainescu/2026/07/25/amazon-requires-sellers-to-label-ai-generated-people-in-listing-images/
• CNBC, “Amazon makes sellers label AI-generated people in images after NY law”, 23 July 2026: https://www.cnbc.com/2026/07/23/amazon-makes-sellers-label-ai-generated-people-in-images-after-ny-law.html
• Quartz, coverage of Amazon’s AI-generated image labelling requirement, 24 July 2026: https://qz.com/amazon-sellers-label-ai-generated-people-product-images-072426
• eWeek, coverage of Amazon’s AI-generated image labels and the New York law: https://www.eweek.com/news/amazon-ai-generated-product-images-labels-new-york-law/
• Goat Consulting, seller guidance on Amazon’s AI-generated image policy: https://www.goatconsulting.com/amazon-policy/amazon-ai-generated-images
• New York General Business Law section 396-b, as amended by S.8420-A / A.8887-B, in force 9 June 2026