Most sellers read one line about Amazon’s AI-image rule and stopped: it is not retroactive. Existing images are fine. Nothing to do.

That reading is correct today and wrong in about six weeks, because the thing that ends the exemption is not a date. It is you opening the listing.

What the rule actually is

On July 22, 2026, Amazon began requiring sellers to identify images and videos containing photorealistic AI-generated people before adding them to listings and A+ Content. The mechanism is unusual enough that it is worth stating precisely.

Element	Requirement
Keyword	`contains-synthetic-performer` — exact string
Field	`dc:subject` (XMP), written with an IPTC-compatible metadata editor
When	Before the file is uploaded to Amazon
Scope	Product images and A+ Content, worldwide stores
Result	Amazon displays an indicator telling customers the image includes AI-generated people

This is not a Seller Central checkbox. It is a string embedded in the file itself. Amazon’s forum guidance puts it plainly: when you include this metadata, Amazon adds an indicator to inform customers that the image includes AI-generated people.

The pressure behind it is legislative rather than commercial. According to one legal analysis of the rule, New York General Business Law § 396-b took effect on June 9, 2026, carrying penalties of $1,000 for a first violation and $5,000 for subsequent ones. Amazon’s practical exposure, though, runs through its own Business Solutions Agreement — image suppression, listing suppression, Account Health violations, and in the worst case account deactivation.

The rule is narrower than teams assume

The most expensive mistake here is over-tagging, and it comes from skim-reading. Four categories are explicitly outside the requirement:

Real people, even when altered with AI tools. Relighting, retouching and AI-assisted cleanup of a photograph of an actual person do not trigger the tag.
Non-photorealistic figures. Illustration, cartoon, stylised rendering.
Characters from expressive works — films, television, streaming, documentaries, video games.
Images with no people at all. Product-only renders, however synthetic, are out of scope.

What triggers the tag is one specific thing: a photorealistic human generated entirely by AI, not based on a real person.

So “tag everything to be safe” is not a safe default. It attaches a public AI-disclosure indicator to imagery of real photographed people, which is a different kind of accuracy problem — and one your brand team will notice before your compliance team does.

Why “not retroactive” expires on contact

Here is the part that turns a July policy into a Q4 operations problem. Asked directly about previously uploaded images, Amazon’s forum response was that compliance is mandatory going forward for new uploads, and that it is best practice to update older images if you plan to make any edits to those listings anyway.

Read that against a normal Q4 calendar. Between now and Black Friday, most catalogues get seasonal hero images, refreshed lifestyle shots, updated A+ modules, holiday bundles and new variation imagery. Every one of those is a new upload against an existing listing.

The grandfather clause does not survive the refresh. It was never a permanent exemption for your catalogue — it was a temporary exemption for files you do not touch, arriving immediately before the quarter in which everybody touches everything.

How large the surface actually is

The obligation is per file, so the workload is a function of how many image files a live listing carries. We measured it rather than estimating it.

Using Helium 10 Product Research on 2026-09-11, we pulled the top 100 listings by 30-day child-ASIN sales in each of eight US categories — 800 listings in total — and read the image count on each.

Category	Median images	Mean images	% with ≥7 images
Beauty & Personal Care	9	9.38	91.0%
Baby Products	9	10.12	87.0%
Home & Kitchen	8	8.07	88.0%
Health & Household	8	9.01	93.0%
Electronics	8	7.96	78.0%
Clothing, Shoes & Jewelry	7	7.10	72.0%
Sports & Outdoors	7	7.33	66.0%
Toys & Games	7	7.02	65.0%

Pooled across all 800 listings, the median is 8 images and the mean is 8.25. 640 of 800 listings — 80.0% — carry seven or more images. No listing in the sample carried zero. The eight categories together account for 6,599 image files.

Two things stand out.

The first is how flat the distribution is. Medians run from 7 to 9 across categories as different as Toys and Beauty. There is no vertical where this is a small job and no vertical where it is an outlier job — the top two by median, Beauty and Baby, are also the categories where photorealistic human models appear most routinely in lifestyle imagery.

The second is that 8 is a floor, not a total. Helium 10 counts images on the listing. A+ Content modules sit on top of that, and the rule covers A+ explicitly. A brand running full A+ on its hero ASINs is auditing meaningfully more than eight files per listing.

One methodology note, because it changed what we published. Helium 10 also returns number_of_videos and influencer_videos, and it is tempting to subtract one from the other to isolate seller-uploaded video. That subtraction is invalid: in 480 of our 800 rows (60.0%), influencer_videos exceeded number_of_videos, including rows reporting zero videos and fifteen influencer videos. The two fields are independent counts, not nested ones, so we report image counts only. Sample caveat: Helium 10 reported more than 100 qualifying listings in every category, so these are top-100-by-sales samples, not category censuses.

The failure mode nobody checks

Writing the tag is the easy half. Keeping it is where this quietly breaks.

dc:subject is XMP metadata, and metadata does not reliably survive an image pipeline. Adobe’s own community documentation describes Export As and Save for Web stripping XMP from exported files, while Save As preserves it. Most teams producing Amazon-ready JPEGs use exactly the paths that strip it, because those are the paths built for small web files.

The result is a compliance failure with no symptom. Your designer tags the master correctly. The export step drops the keyword. The file that reaches Amazon is untagged and looks, to every human in the chain, identical to a tagged one. Nothing errors. Nothing warns.

This is why the verification step has to sit at the end of the pipeline rather than the beginning. The only file whose metadata matters is the one you actually upload.

The Operator Read

The date to plan against is not July 22. It is whenever your Q4 image refresh starts.

Run the classification pass before the creative work, not after. Sort human-containing assets into photographed real person, AI-edited real person, non-photorealistic figure, and fully AI-generated photorealistic person. Only the fourth needs the tag. Do this first and the rest of the work is mechanical.
Do not tag defensively. Over-tagging puts a public AI-disclosure indicator on pictures of real people. “When in doubt, tag it” is the wrong instinct here; when in doubt, find out who is in the picture.
Test your export path with a known-good file. Tag one test image, push it through the exact route your team uses to produce upload-ready files, and read the metadata back. If the string is gone, the problem is your pipeline, and no amount of careful tagging upstream will fix it.
Verify the uploaded artefact, and keep the read-back. The evidence you want later is proof of what the file contained when Amazon received it, not proof that someone tagged a master in September.
Treat your Q4 refresh list as your compliance list. They are the same list. Every listing getting new imagery this quarter is a listing whose exemption ends the moment the new file lands.

If you want a second pair of eyes on which listings are actually getting touched this quarter — and what else those pages are giving away — our Amazon listing audit works at the asset and offer level, and our pricing is published in full, no call required.

The rule arrived in July with a grandfather clause. Q4 is the quarter that clause was never going to survive.

Amazon's AI-people image tag is not retroactive — until you re-upload for Q4