Amazon's AI-Generated People Rule: What Your Creative Team Has to Tag

Your retoucher sends back a lifestyle set on Friday. One of the frames has a person in it who does not exist, because the brief said "model-free" and the pipeline obliged. The image is good. It is also, as of July 2026, an asset Amazon expects you to label before it goes anywhere near a listing, and nobody in that chain, not the retoucher, not the brand manager, not whoever does the uploads, has been told which field the label goes in.

That is the actual shape of this problem. It isn't a question about whether AI creative is allowed on Amazon. It is a question about who in your production line is responsible for a metadata keyword.

What Amazon told sellers in July

On 22 July 2026, Amazon notified third-party sellers that product images or videos containing AI-generated people have to be labeled, and that the requirement extends to A+ content. CNBC reported the notice the following day. Sellers add specific metadata keywords to the asset before uploading it, and Amazon says it will add an indicator on the listing to tell shoppers when content features AI-generated people, where applicable.

The driver is a New York law requiring disclosure when synthetic performers stand in for human actors in advertising, which took effect in June 2026. Governor Kathy Hochul called it first-in-the-nation, which is usually a reliable signal that other states are drafting their own version. We'd plan for that rather than for this staying a New York problem.

Worth saying plainly before the detail: this is not a ban. Amazon has not told anyone to stop producing AI creative. It has created a disclosure obligation for one narrow category of asset, and the cost of getting it wrong falls on whoever owns the Amazon creative production pipeline, which in most brands we work with is three different people who have never been in the same meeting.

What triggers the rule, and what doesn't

The scope is narrower than the panic around it. The trigger is a photorealistic AI-generated person. Not AI in the pipeline, not AI retouching, not AI anything else.

Asset	Needs the label?
Lifestyle image with a photorealistic person generated by AI	Yes
Video with an AI-generated presenter or model	Yes
A+ content module containing an AI-generated person	Yes
Photo of a real model, retouched or upscaled with AI tools	No
Photo of a real model with an AI-replaced background	No, the person is real
Product-only white background shot, AI-generated or not	No, there's no person
Licensed film, television or video game characters	No, explicitly excluded
Illustrated or obviously stylized figures	Not photorealistic, so no

Read the two exclusions carefully, because they are where the arguments will happen. A real person edited with AI is out of scope. Fictional characters from movies, TV and games are out of scope. Everything in between, which is most of what a generative pipeline produces when you ask it for "a woman in her thirties holding the product," is in scope.

The label, and the reason it disappears

Seller-facing guidance across the agency and tooling world has converged on one keyword written into the image's XMP subject metadata, spelled in lowercase with hyphens: contains-synthetic-performer. We'll be honest about the sourcing here, because it matters for how much weight you put on it. Amazon's own notice sits inside Seller Central behind a login, and the news coverage that confirms the policy does not name the field or the string. So treat the exact keyword as well-attested secondary guidance and confirm it against the current notice in your own Seller Central account before you build a process around it.

The mechanical problem is more interesting than the policy one. XMP metadata is fragile. Export a PSD to a web-optimized JPEG through the wrong preset and the subject field is gone. Run the file through a compression service, a DAM that re-encodes on ingest, or a bulk resizer, and it's gone. Hand the asset to a freelancer who opens it, crops it and saves a copy, and it's gone.

Across our client portfolio the creative chain usually has at least one of those steps in it, and often three. So the label is not a one-time action on an asset. It's a property the asset has to still have at the moment of upload, which means somebody has to check it at the end of the line rather than the beginning.

Assets Amazon generated versus assets you generated

There's a split here that catches people out, and it runs along the question of whose tool made the person.

Amazon has been shipping generative creative features inside its own ad console for a while now, and the guidance circulating since July says Amazon handles disclosure itself for synthetic performers produced by its own tools: it identifies them and surfaces the indicator without the advertiser doing anything. Assets you produced somewhere else and brought in are the ones that need a declaration from you at upload. That asymmetry makes sense from Amazon's side, since it knows what its own models made.

Practically, it means the risk concentrates exactly where you'd expect. A team using Amazon's in-console creative tools has less to track. A team with an external generative pipeline, a contractor, or an agency producing lifestyle sets off-platform carries the whole obligation. If you run Sponsored Brands and Sponsored Display creative through an outside studio, that's the stream to audit first, because ad creative and listing creative are usually produced by different people on different timelines and only one of them has seen a policy update.

We'd add one caution to that, consistent with everything else here: the automatic-handling claim comes from the same secondary guidance as the keyword itself, not from the news coverage. Don't build a process that assumes Amazon's tools cover you until you've read the current notice yourself.

Who actually owns this in a real pipeline

We've found this breaks along the same seam every time. The person who knows an image contains a synthetic performer is the person who generated it, usually a creative contractor. The person who uploads it is usually an account or catalog manager. Neither has the other's information, and the metadata field is invisible in every interface either of them looks at.

Four changes close that gap, and none of them needs a tool purchase:

1. Put the declaration in the delivery brief. Every creative hand-off should state, per file, whether it contains a photorealistic AI-generated person. That is a line in a spreadsheet, not a project.
2. Write the keyword at the source, then verify at the destination. The contractor tags it. The uploader confirms the tag survived. Two steps, because one of them will fail.
3. Audit the back catalog, not just new work. The rule is about what's on the listing today. If you shipped generative lifestyle imagery in late 2025 or the first half of 2026, before any of this existed, those assets are live and untagged right now.
4. Keep the export preset fixed. Pick one metadata-preserving export path and ban the others. Most of the losses we've traced came from somebody using a different save-for-web preset in a hurry.

Point four sounds like the trivial one. It's the one that actually fails, because it depends on a habit rather than a decision.

What this does not tell you

Two real gaps, and we'd rather name them than pretend the policy is fully legible.

Amazon says it will show the shopper-facing indicator where applicable, and it hasn't published the criteria for when that is. So you cannot currently model what the disclosure badge does to your conversion rate, because you cannot reliably predict whether it appears. Anyone telling you the badge costs you a specific percentage of conversions is guessing.

The enforcement side is also thin. Coverage of the notice describes the obligation without describing a penalty schedule, and the secondary guidance that mentions asset removal and account-health consequences is extrapolating from how Amazon handles other media policies. That extrapolation is probably right. It is still extrapolation, and it isn't a reason to panic-audit a catalog in a weekend.

Is this worth your time this quarter?

If your listing creative is photographed, no. You have a one-line policy note and nothing to do. If any part of your gallery, video or A+ content came out of a generative pipeline, then yes, and the work is an afternoon of auditing rather than a project.

The brands that will have trouble with this are the ones who moved to generative lifestyle imagery for the cost reason, at volume, without a record of which assets came from where. That was a reasonable call at the time. It's now a documentation problem, and documentation problems are cheap to fix early and expensive to fix after a listing gets suppressed.

For what it's worth on the creative question underneath all this: the main image is still the asset that decides whether anyone clicks, and we've never seen a synthetic model beat a well-lit real one on a product people wear or hold. When we rebuilt True Classic's Amazon launch to $1.1M in its first 40 days, the creative that carried it was photographed. If you're choosing a pipeline on cost alone, test the asset before you commit the catalog to it.

Tag what needs tagging, verify it at upload, and stop treating metadata as somebody else's field. If you want somebody to go through the catalog and tell you which assets are exposed, book a call and we'll do that read.