If a product image, video, or A+ module shows a photorealistic person created entirely by AI and not based on any real human, Amazon is asking sellers to embed the keyword contains-synthetic-performer in that file’s IPTC metadata before uploading it. In practice that means writing the string into the file’s Keywords field — the IPTC property that modern editors store as XMP dc:subject. One command does it:
exiftool -XMP-dc:Subject+=contains-synthetic-performer -overwrite_original image.jpg
Everything below is the part the news coverage skipped: which files need the tag, which field it goes in, how to do a whole catalog at once, and how to confirm the write landed. Data checked 2026-07-25.
Where this rule comes from — and what we could not verify
Amazon notified third-party sellers on Wednesday, July 23, 2026. The notice sits inside Seller Central’s product image and shopping video guidance, which is login-gated — we could not retrieve Amazon’s original wording from any public page. Every statement about Amazon’s requirement below is therefore written from press reporting, not from an official page we read ourselves:
- CNBC, 2026-07-23 — Amazon makes sellers label AI-generated people in images after NY law, reporting from a copy of the seller notice.
- eWeek, 2026-07-24 — Amazon Requires Sellers to Disclose AI-Generated People in Product Images.
- Quartz, 2026-07-24 — Amazon is requiring sellers to label AI-generated people in product images.
Per that reporting, sellers must identify qualifying product images and A+ content — including videos and enhanced graphics on product pages — and tag them with the metadata keyword using an IPTC-compatible editor that supports embedded metadata. Amazon said it will use the signal to display a disclosure on listings where applicable. The legal, IPTC, and ExifTool facts below are a different matter: those come from primary sources, linked in place.
Which images need the keyword
This is where most sellers will get it wrong, because the scope is far narrower than “we used AI.” The trigger is a synthetic performer: a photorealistic human who does not exist. Editing tools are irrelevant; the question is whether the person in the frame is real.
| Your creative | Tag it? | Reason |
|---|---|---|
| Photorealistic model generated entirely by AI, not based on a real person | Yes | Exactly what the rule targets |
| Photo of a real model, retouched or enhanced with AI | No | Real people are out of scope even when AI-edited |
| AI-generated background or scene, no person in frame | No | No person, no synthetic performer |
| Product-only shots, flat lays, packaging renders, infographics | No | No person |
| Cartoon, illustrated, or stylized characters | No | Not photorealistic |
| TV, movie, or video-game characters | No | Explicitly excluded |
| Real person’s body, AI-generated face | Judgement call | Not addressed in the reported scope |
| Photorealistic AI hands or arms demonstrating the product | Judgement call | Not addressed in the reported scope |
The last two rows are our reading, not Amazon’s wording — the reported scope splits “created entirely by AI” from “real person edited with AI” and says nothing about hybrids or partial bodies. If a frame shows a photorealistic human likeness that no real person sat for, tagging it is the conservative call.
Two practical consequences:
- Per file, not per listing. A listing with one AI lifestyle shot and six product-only shots needs the tag on one file.
- Before upload, not after. The metadata must be inside the file when Amazon ingests it. Re-tagging a local copy does nothing to an image that is already live — you have to re-upload.
If you generate AI model shots in the first place, our AI product photography workflow covers that side; this page is the step that now sits between generation and upload.
Which metadata field the keyword goes in
The reporting says “metadata keyword” and “IPTC-compatible editor.” Those two phrases point at one specific field. In the IPTC Photo Metadata Standard (2025.1, revision 1, dated 2025-11-26), the property named Keywords is defined as:
| Layer | Where the keyword is stored |
|---|---|
| XMP (current) | dc:subject, a Bag of text values, namespace http://purl.org/dc/elements/1.1/ |
| Legacy IIM | Dataset 2:25 Keywords, max ~64 bytes per keyword |
So when an IPTC-compatible editor shows a Keywords box and you type a term into it, the value lands in XMP dc:subject. That is the field to target. The IPTC Photo Metadata User Guide notes that some fields are duplicated across IIM and XMP and software has to keep the two in sync — which is why the belt-and-braces command below writes both copies.
Two details decide whether your tag is readable. The string is a literal identifier — all lowercase, hyphens, no spaces, since “Contains Synthetic Performer” is a different keyword. And it is one keyword added alongside any you already use, not a sentence; at 28 characters it fits inside the legacy IIM limit of roughly 64 bytes.
Method 1: ExifTool (free, scriptable, handles batches)
ExifTool is the practical choice because it is the only free option that does a whole catalog in one line. The current release is 13.59, published 2026-05-27; it runs on macOS, Windows, and Linux.
Tag a single file
exiftool -XMP-dc:Subject+=contains-synthetic-performer -overwrite_original image.jpg
+= adds an entry to a list tag instead of replacing the list, so existing keywords survive. -overwrite_original skips ExifTool’s default of leaving an image.jpg_original backup beside every file — per the documentation, use it only when you already have separate backups.
Write both the XMP and the legacy IIM copy
exiftool -XMP-dc:Subject+=contains-synthetic-performer \
-IPTC:Keywords+=contains-synthetic-performer \
-overwrite_original image.jpg
Amazon has not published which container it reads. Writing both costs nothing and removes the guess. Legacy IIM is best supported in JPEG and TIFF; if that write fails on another format, the XMP copy is the one that matters.
Tag an entire folder
exiftool -r -ext jpg -ext jpeg -ext png \
-XMP-dc:Subject-=contains-synthetic-performer \
-XMP-dc:Subject+=contains-synthetic-performer \
-overwrite_original \
~/listing-images/ai-model-shots/
-r recurses into subdirectories, -ext restricts processing to the extensions you name. The -= before the += is deliberate: -= removes a matching entry, so re-running the batch cannot leave the keyword listed twice. ExifTool prints how many files it updated — compare that number against how many you expected to change.
Videos and A+ assets
The reported scope covers videos and A+ graphics, not just stills. ExifTool’s format table lists MOV and MP4 as read/write and XMP as creatable, so the same syntax applies:
exiftool -XMP-dc:Subject+=contains-synthetic-performer -overwrite_original clip.mp4
Verify these file by file — video containers vary far more than JPEG, and a silent no-op is easy to miss in a folder of MP4s.
Method 2: a metadata editor with a Keywords panel
For a handful of hero images, any editor exposing the IPTC Keywords field will do — the field name is standardized even though menu paths are not. Common places to look: the Keywords panel or File Info dialog in Adobe Bridge, Lightroom Classic, and Photoshop; the IPTC stationery pad in Photo Mechanic; the IPTC/XMP panels in the free XnView MP and digiKam.
Two cautions. Menu paths and labels shift between versions, so confirm the field you edited is the one labelled Keywords under IPTC, not a proprietary “tags” feature that writes only to a local catalog database. And — the reason the next section exists — a GUI showing the keyword in its panel has not proven the keyword is inside the file: some applications hold keywords in a sidecar .xmp file until you explicitly write metadata to the file.
How to verify the keyword actually landed
Run this on a file you just tagged:
exiftool -s3 -XMP-dc:Subject image.jpg
-s3 prints values only. You should see your keywords, one per line, including contains-synthetic-performer. An empty result is a failure, not a pass — it means nothing was written.
To see which container each copy lives in:
exiftool -G1 -a -s -Subject -Keywords image.jpg
-G1 prefixes each result with its group (XMP-dc, IPTC), -a shows duplicates across groups, -s uses short tag names. This confirms both copies exist.
To audit a folder, list every file still missing the keyword:
exiftool -r -q -if '$Subject !~ /contains-synthetic-performer/' -p '$FileName' ~/listing-images/
-if evaluates a condition per file and -p prints a format string only for files that pass, so the output is a plain list of filenames needing work. Swap !~ for =~ to list the ones already tagged. Run this after every batch — it is the cheapest way to catch a command that silently matched zero files because of a wrong path or extension filter.
Auditing a catalog that is already live
New uploads are the easy case. The harder question is which of your existing listing images contain AI-generated people.
- Inventory your live creatives per ASIN. Bulk-pulling images beats clicking through detail pages — see downloading Amazon product images in bulk, or a one-click extension such as ASINCrate (overview). One limitation: images pulled from a live listing have already been through Amazon’s image pipeline, so use them to identify which creatives show AI-generated people, not to inspect the metadata you originally embedded.
- Sort against the table above. Most sellers find the affected set is small — AI lifestyle and model shots, not the whole gallery.
- Tag the master files, not the downloads. Run the batch command against your original renders.
- Verify, then re-upload. Confirm zero files come back from the audit command, then re-upload. Since the gallery is open anyway, it is an efficient moment to work through your listing optimization checklist.
- Bake it into the workflow. Put the tag step between “export render” and “upload” so this never becomes a cleanup project again.
Optional: other IPTC signals for AI-generated media
Amazon’s reported requirement is the keyword and only the keyword. But IPTC defines richer standardized properties for synthetic media, and adding them is harmless, machine-readable provenance:
- Digital Source Type. IPTC recommends that images made by trained AI algorithms carry
Iptc4xmpExt:DigitalSourceTypeset tohttp://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia. Related values exist for composites with synthetic elements and for digital art.
exiftool -XMP-iptcExt:DigitalSourceType=http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia \
-overwrite_original image.jpg
- Generation details. The IPTC user guide has a section on applying metadata to AI-generated images, with fields for AI Prompt Information, AI Prompt Writer Name, AI System Used, and AI System Version Used. ExifTool added the new XMP-iptcExt generative AI tags in version 13.40, released 2025-10-24, so keep ExifTool current if you want to write them.
These do not substitute for contains-synthetic-performer, and nothing in the reporting suggests Amazon reads DigitalSourceType. Treat them as hygiene for your own asset library and for other platforms that adopt the IPTC vocabulary.
What the New York law actually says
New York’s Synthetic Performer Disclosure Law (S.8420-A / A.8887-B) amended General Business Law § 396-b. It was signed 2025-12-11 and took effect 2026-06-09. Per the statute:
- Anyone who, for a commercial purpose, produces or creates an advertisement must conspicuously disclose in it that a synthetic performer appears — but only where that person has actual knowledge.
- Civil penalties are $1,000 for a first violation and $5,000 for each subsequent violation.
- Exemptions include advertisements for expressive works such as films, TV shows, and video games where the synthetic performer use is consistent with the work; audio-only advertisements; and AI used solely for language translation of a human performer.
- Media outlets and platforms are exempted from liability for third-party advertising content.
The last two points explain the mechanics: the duty falls on whoever creates the advertisement — the seller — while the platform is shielded, and the actual-knowledge standard is why the tagging burden sits with you. The statute’s expressive-works exemption also matches Amazon’s reported carve-out for TV, movie, and video-game characters, which is a useful sanity check on edge cases.
Common mistakes
- Tagging everything AI-touched. Background removal, upscaling, relighting, and retouching a real model are all out of scope.
- Typing the keyword by hand. One typo or capital letter and it no longer matches. Copy-paste or script it.
- Putting it in the wrong field. Pasted into the caption, title, or description field, it is not in the field being read.
- Trusting the GUI. Confirm with
exiftool -s3 -XMP-dc:Subjectinstead of the panel showing what you just typed. - Assuming the tag is retroactive. Live images keep whatever metadata they had at upload, and files pulled back down from a listing are not the files you re-upload — fix the masters.
- Forgetting videos and A+ assets. The reported scope names them explicitly, and they live in a different folder from the main gallery.
Frequently Asked Questions
What exactly is the metadata keyword Amazon wants?
The literal string contains-synthetic-performer, all lowercase with hyphens, added as a keyword in the file’s IPTC metadata before upload. Treat it as case-sensitive and exact — Amazon has not published its matching rules. Per reporting by CNBC (2026-07-23) and eWeek (2026-07-24) it applies to product images, videos, and A+ content containing photorealistic AI-generated people. Data checked 2026-07-25.
Which metadata field does contains-synthetic-performer go in?
The IPTC Keywords field, which the IPTC Photo Metadata Standard 2025.1 defines as XMP dc:subject (a Bag of text values), with a legacy IIM equivalent at dataset 2:25. In ExifTool that is -XMP-dc:Subject+=contains-synthetic-performer; in a GUI editor it is the Keywords box in the IPTC panel.
Do I need to tag photos of real models edited with AI?
No. Per the reported scope, the requirement covers photorealistic people generated entirely by AI and not based on a real person. Images of real people stay out of scope even when edited with AI tools. Also excluded: TV, movie, and video-game characters, images with no people, and non-photorealistic characters.
Can I add the keyword without paid software?
Yes. ExifTool is free and open source and tags a single file or an entire folder from one command. Free desktop editors with IPTC panels, such as XnView MP and digiKam, work for small volumes. Confirm the result with exiftool -s3 -XMP-dc:Subject image.jpg either way.
Does this apply to videos and A+ content?
Per the reporting, yes — the notice covers product images and A+ content including videos and enhanced graphics. ExifTool lists MOV and MP4 as writable and XMP as creatable, so the same command applies to video files; verify each one individually, because video containers behave less predictably than JPEG.
What are the penalties behind this rule?
The underlying statute is New York General Business Law § 396-b, effective 2026-06-09, setting civil penalties of $1,000 for a first violation and $5,000 for each subsequent one where an advertiser with actual knowledge fails to disclose a synthetic performer. Amazon has not published a separate enforcement schedule for sellers; treat listing-level consequences as unknown until it does.
Conclusion
The operational core is three steps: decide whether a file shows a human who does not exist, write contains-synthetic-performer into that file’s IPTC Keywords field, and verify with exiftool -s3 -XMP-dc:Subject before upload. The scope is narrower than most coverage implies — real models edited with AI are out, stylized characters are out, product-only shots are out.
What Amazon has not published is worth tracking rather than guessing at: which container it reads, how it matches the string, and what happens to a listing that should have been tagged and was not. Until that appears on a page readable without a login, write both copies, keep master files tagged, and re-verify after every batch. For adjacent 2026 listing changes see the 75-character title limit; for the protection side of listing hygiene see listing hijacker removal and the suspension appeal playbook.