Amazon's AI Image Rule: How to Tag AI-Generated People

On July 23, 2026, Amazon told sellers that images containing photorealistic AI-generated people have to carry a tag identifying them as such. It applies to Amazon's worldwide stores, and it covers product images, lifestyle images, video, and A+ content.

Coverage of this has been loud and unhelpful. Two things are getting mangled: what the "label" actually is, and who can actually fine you. Both matter, because the seller reading "AI images now need a disclaimer" reasonably concludes they need to stamp visible text on their photos. You don't. If you do, you've just damaged your main image for no reason.

The label is invisible

There is no badge. Nothing gets written on the picture.

What Amazon wants is a keyword buried in the image file's metadata, the invisible information every photo carries alongside the pixels: camera model, date, copyright and so on. Specifically, you add the keyword contains-synthetic-performer to the field called dc:subject.

Here's the part that makes this manageable, and that almost no write-up mentions: dc:subject is the "Keywords" field. It's the same box you've been ignoring in Photoshop's File Info dialog and in Adobe Bridge for years. Adobe apps write to it every time you type a keyword. You don't need to buy anything, learn a new tool, or hire anyone.

So the actual instruction, the one that took me longer to find than it should have, is:

Open the image in a tool that can edit IPTC/XMP metadata, find the Keywords field, and add contains-synthetic-performer as a keyword. Spell it exactly like that. Then upload.

That's the whole step.

Amazon says that when the metadata is present, it will show shoppers an indicator that the image contains AI-generated people. Worth knowing: sellers who have tagged images report not yet seeing that indicator appear, so the customer-facing half looks like it's still rolling out. The tagging requirement is live regardless.

Most of your images don't need this

The exemptions are broader than the panic suggests, and getting them wrong costs you real work for nothing.

You do not need the tag when:

• The image shows a real person, even if you used AI tools to edit, retouch, relight or clean up the shot. A real model is a real model.
• The AI generated the background, scene or product render, but there's no AI human in it. The kitchen counter your product is sitting on doesn't count.
• There are no people in the image at all, which is most product photography.
• The people aren't photorealistic: illustrations, cartoons, obvious graphics.
• It's a character from a film, show, game or other creative work.

You do need it when a photorealistic human who does not exist appears in your image or video. The AI-generated "model" holding your product. The synthetic family using it at the beach. That's the target.

Read that list again if you've been assuming AI-heavy workflows are all in scope. Retouching a real photoshoot with AI is exempt. Generating a fake person is not. The line is the human, not the tool.

Nobody mentions the step where it breaks

Here's what will actually catch people out, and it's the reason to treat this as a workflow change rather than a one-off task.

Metadata is fragile. It travels alongside the image rather than inside the picture itself, and a great many export, resize and compression steps throw it away by default. That's normally a feature, since it strips out camera serials, GPS coordinates and editing history you don't want published. Image compressors, "export for web" settings, and some design-tool downloads routinely discard metadata to shrink the file.

Which means this failure mode is entirely realistic: you carefully add the keyword, then run the image through your usual resize-and-compress step before upload, and the keyword silently disappears. You've done the work and you're not compliant. Nothing warns you.

Test your pipeline once. Tag one image, push it through your normal export and compression route exactly as you always do, then read the metadata back on the file you would actually have uploaded. If the keyword survived, your workflow is fine forever. If it didn't, tag the image after compression instead, and you're fine forever.

That single test is the difference between this being a two-minute change and a compliance problem you don't discover for months.

If you're handling images in bulk, the standard tool is ExifTool, which can write the field from the command line. Adding -XMP-dc:Subject+="contains-synthetic-performer" appends the keyword without wiping any keywords already on the file, and reading the same field back afterwards is how you confirm it stuck. Anyone doing your image processing will know the tool. For everything else, the Keywords box is faster.

Who can actually fine you

This is where the reporting has been sloppiest, and it's worth being precise, because the two things are unrelated in practice.

The $1,000 and $5,000 figures are New York State civil penalties, not Amazon penalties. They come from New York's synthetic performer disclosure law, an amendment to General Business Law §396-b that took effect on June 9, 2026: $1,000 for a first violation, $5,000 for each one after that. It's a state law about disclosure in advertising, and it's what prompted Amazon to move.

Amazon has not published a deadline, an enforcement mechanism, or a stated consequence for untagged images. It said to do it; it didn't say what happens if you don't. I'd rather tell you that than invent a cliff edge.

So the honest risk picture is: a legal exposure that comes from a state statute and applies to your advertising, and a separate platform policy where the consequences are unknown and Amazon reserves the right to decide later. Neither is a reason to panic. Both are reasons not to leave it.

The pragmatic reading is that Amazon rolled this out globally rather than fencing it to New York, which tells you it intends the tag to be the default state of a compliant catalogue. Policies that start as advisory and later grow teeth are not exactly unheard of.

What to do this week

1. Work out whether you're even affected. Do any live images or videos show a photorealistic person who was generated rather than photographed? For a lot of catalogues the answer is none, and you're done in five minutes.
2. List the ones that are. Include A+ content and video, not just the main image gallery. The requirement covers all of it, and A+ is where AI lifestyle imagery tends to hide.
3. Add the keyword. contains-synthetic-performer in the Keywords field. Exact spelling.
4. Test that it survives your export. Once. This is the step that actually decides whether you're compliant.
5. Put it in the brief. If a freelancer, agency or AI tool produces images for you, the tag belongs in your standard instructions from now on. Retrofitting is the expensive version.
6. Decide your policy on synthetic models generally. Now that the disclosure is going to be visible to shoppers, "should we use AI people at all" becomes a conversion question rather than only a production-cost one.

The honest summary

This is a small, precise task dressed up as a scary one. One keyword, in a field your software already has, on the subset of images featuring people who don't exist. The genuine risk isn't the rule. It's assuming the tag survived a pipeline that quietly removed it.

The wider pattern is worth noticing, though. Compliance requirements are increasingly landing inside the file rather than on the listing page, where you can't see them and won't get an error. That's a category of problem that doesn't announce itself.

Knowing which of your listings changed, and when, is the part we build for. Seller Sphere tracks your catalogue and tells you when something moves, so the changes you didn't make, and the ones that quietly failed, don't sit there for a quarter before anyone notices.

Frequently asked questions

What is Amazon's new AI image rule? From July 23, 2026, Amazon requires sellers to tag images and videos containing photorealistic AI-generated people. The tag is a keyword, contains-synthetic-performer, added to the image's dc:subject metadata field before upload. It applies across Amazon's worldwide stores and covers product images, video and A+ content.

Do I have to put a visible label on my product images? No. Nothing is written on the picture itself. The disclosure is invisible metadata attached to the file. Amazon says it will display its own indicator to shoppers where applicable, so you should not add text or badges to the image yourself.

Where exactly does contains-synthetic-performer go? In the dc:subject XMP field, which is the same thing as the "Keywords" field in Adobe Bridge and Photoshop's File Info dialog. Any IPTC-capable metadata editor can write it. Type it exactly as contains-synthetic-performer.

Do I need the tag if I used AI to edit a photo of a real model? No. Images of real people are exempt even when AI tools were used to edit or retouch them. The tag is only for photorealistic people who were generated rather than photographed. AI-generated backgrounds, scenes and product renders without an AI human are also exempt, as are non-photorealistic illustrations.

What happens if I don't tag my AI images? Amazon has not published a deadline, enforcement mechanism or stated consequence. Separately, New York's synthetic performer disclosure law (General Business Law §396-b, effective June 9, 2026) carries civil penalties of $1,000 for a first violation and $5,000 for each subsequent one. Those are state penalties for advertising disclosure, not Amazon fines.

Why would the tag disappear after I add it? Because metadata is commonly stripped by resizing, compression and "export for web" steps, which discard it by default. If you tag an image and then run it through your usual export pipeline, the keyword can be removed silently. Test your workflow once by reading the metadata back on the final file, and tag after compression if it doesn't survive.