Amazon's New AI-People Disclosure: What Sellers Must Tag Before Upload (July 2026 Rule)
Amazon now requires a tag for photorealistic AI-generated people. Learn which images need the tag, which do not, and how to add it before you upload.
10/28/202619 min read
AI-generated models are now a normal part of e-commerce photography. A brand can show a garment on a realistic model without booking a studio, a photographer, and a model for every colourway. That speed is useful, but it has created a new compliance step on Amazon: if a photorealistic AI-generated person appears in your listing image, Amazon requires you to tag that image before upload.
This article explains that rule in plain terms, with the checklists and worked examples you need to apply it to your own gallery. It is not legal advice, and it does not replace checking Seller Central for your category. It is a practical guide to help you decide which images need a tag, which do not, how to add the tag correctly, how to check that it survived your export, and what to tell an editor so the tag is not lost in retouching.
What changed in July 2026
According to reporting in July 2026, Amazon began requiring sellers to tag photorealistic AI-generated people in listing images. The reports link the change to New York's synthetic-performer law, which took effect on 9 June, and describe a metadata keyword that sellers must add to the image file.
The keyword reported is contains-synthetic-performer. It is added to the image metadata, in the IPTC or XMP subject field, often written as dc:subject in metadata tools. In other words, the tag is not text printed on the image itself. It is information carried inside the image file, so Amazon's systems can read it when you upload.
This distinction catches many sellers out, so it is worth stating plainly. Writing "AI model" in your bullet points, adding a note to your product description, or stamping a watermark on the image is not the same thing as adding this metadata tag. Those visible notes may help shoppers, and they may be good practice in their own right, but they do not put the required keyword into the file's metadata. Conversely, adding the metadata keyword does not by itself put visible text on the image. The rule described in the reporting is about the file, not about the caption.
This matters because many sellers first heard about AI disclosure in general terms — "label your AI images" — without a clear instruction about what to tag, where to put the tag, or which images are exempt. A general label on a listing page is not the same as the metadata tag described here. If you use AI people, you need to handle the file itself, not just add a note to your description.
The 60-second decision rule
Before you open any metadata tool, sort every image in the gallery into three piles. This one sorting step prevents most mistakes, because the common error is not a technical failure — it is tagging the wrong set of files.
Tag it: the image contains a photorealistic person who was generated by AI. It does not matter whether the person is the hero of the shot or a smaller figure in a lifestyle scene. If a viewer could reasonably think that person is real, and the person is in fact AI-generated, put the image in this pile.
Do not tag it with this keyword: the image contains no AI-generated person at all. That includes a real person photographed in a studio, a real person whose photo was retouched with AI tools, an AI-generated background with no people, a flat lay, a ghost mannequin shot, and a packshot with no person in it.
Check first, then decide: the image is ambiguous. Perhaps you received it from a freelancer and do not know whether the model was real or generated. Perhaps the person is partly cropped, heavily stylised, or only a hand or a silhouette. Perhaps an AI tool "enhanced" a real photo so much that you are no longer sure what is real. Do not guess. Ask the person or tool that made the image, check the original project file if you have it, and, until you have an answer, treat the image as needing a tag. Tagging an ambiguous image while you check is safer than uploading it untagged on an assumption.
Apply this rule image by image, not listing by listing. A single listing often mixes all three piles: a tagged AI-model hero image, an untagged ghost mannequin secondary image, and an untagged real-model detail shot. Tagging the whole gallery because one image needs it, or leaving the whole gallery untagged because most images do not need it, are both errors.
Which images need the tag
Based on the reporting, the tag applies to photorealistic AI-generated people. In practical terms, that includes:
- An AI-generated model wearing, holding, or using your product, where the person looks like a real person
- An AI-generated "customer" or lifestyle person in a scene, where the person is photorealistic and could be mistaken for a real person
- An AI-generated person in a main or secondary image, if that person is photorealistic
The key test is realism and personhood together. If a viewer could reasonably think the person in the image is a real person, and that person was in fact generated by AI, the image is the kind of image the rule is aimed at.
This is also why the rule is easy to miss. A seller may buy AI model images from a tool, receive finished files, and upload them without ever seeing the metadata. If the tool did not add the tag, or if the seller exports a new version that strips metadata, the uploaded file may arrive untagged. Always check the file you actually upload, not just the file you first received.
Which images do not need this tag
The same reporting describes important exemptions. You do not need this particular people tag for:
- An AI-generated background with no people in it
- AI retouching, colour correction, background removal, or background replacement applied to a photo of a real person
- A photo of a real person that has been edited with AI tools, where the person themselves is real
- A fictional character who is clearly not a photorealistic person
This distinction is important for editing services and for sellers who use AI tools for ordinary retouching. Cleaning a background, correcting colour, or removing a distraction from a real model's photo is not the same as generating a person. If no AI person is in the image, the contains-synthetic-performer tag does not apply to that image on the basis described here.
That said, do not use this exemption as a loophole. If you generate a person, tag it. If you are unsure whether a person in an image counts as photorealistic and AI-generated, treat it as needing a tag until you have checked with the tool that made it or with Seller Central guidance for your category.
Worked examples: six galleries, six answers
Abstract rules are easy to agree with and hard to apply. These six common cases show how the decision rule works in practice. In each case, the question is the same: does this exact file contain a photorealistic AI-generated person?
1. AI model wearing a hoodie, hero image. The model was generated by an AI tool, the pose and fabric look real, and the image is the first image a shopper sees. Tag it. This is the central case the rule is aimed at, and it is also the image most likely to be checked, because it is the main image.
2. Real model, AI-cleaned background. You photographed a real model in a studio, then used an AI tool to remove a distracting background and replace it with clean white. The person is real, even though the background is AI-generated. Do not add the people tag on the basis described here. The AI work changed the setting, not the person.
3. Real model, AI retouching only. Skin retouching, wrinkle smoothing, colour correction, and stray-hair cleanup were applied to a photo of a real model. The person remains a real person. Do not add the people tag. Keep a note of what was retouched anyway, because a buyer or a marketplace reviewer may ask, but retouching a real person is not generating a person.
4. Ghost mannequin product shot. The garment is shown on an invisible mannequin form, with no visible person, no face, and no hands. There is no person to tag. Do not add the people tag. This is one reason ghost mannequin images are useful as secondary images: they show shape and fit without raising the AI-person question at all.
5. Lifestyle scene with an AI-generated person in the background. Your product is a backpack on a table in the foreground, and a photorealistic AI-generated person is walking past in the background, slightly out of focus. Tag it. The rule is about whether a photorealistic AI person appears in the image, not about whether that person is the subject. Background figures still count if they look real.
6. Clearly stylised illustration, not a photograph. The image is a cartoon-style character, an obviously painted figure, or a low-detail avatar that no viewer would mistake for a real person. On the basis described in the reporting, this is a fictional character who is clearly not a photorealistic person, so this particular tag does not apply. If you push the style toward photorealism — realistic skin, realistic eyes, realistic lighting — you cross back into "check first" territory, and you should treat the image as needing a tag until you confirm otherwise.
Two edge cases deserve a direct warning. First, a partially visible person still counts if the visible part looks like a real person: an AI-generated hand holding a bag, or an AI-generated torso in a cropped shot, should not be dismissed just because the face is out of frame. Second, swapping a real model's face or body with an AI-generated one changes the answer. If any photorealistic part of the person in the final image was generated rather than photographed, check first and tag unless you can confirm the final person is real.
How to add the tag, step by step
Adding metadata sounds technical, but the steps are straightforward if you work from a copy of the file and check your work:
1. Identify every image that contains a photorealistic AI person. Go through the full gallery, not just the main image. Secondary images, lifestyle images, and comparison images can all contain people.
2. Work on a copy. Keep your original AI output untouched, and add the tag to the copy you will upload. That way, if metadata is stripped by an export, you can redo it from a known good file.
3. Open the file in an editor that can write IPTC or XMP metadata. Many photo editors and metadata tools can do this. Look for a keywords, subject, or metadata panel. In a typical editor this is a File Info or metadata panel; in a catalogue tool it is the keywording panel; in a command-line tool it is the subject or keywords field.
4. Add contains-synthetic-performer as a keyword or subject term. Spell it exactly, with hyphens and no spaces: contains-synthetic-performer. Do not add it as visible text on the image, and do not bury it in a long paragraph of unrelated keywords where it could be lost on export. Add it as its own separate keyword term, alongside any other keywords you normally use, so it survives as a discrete field rather than as one word inside a sentence.
5. Save and re-check. Reopen the exported file and confirm the keyword is still present in its metadata. Some export settings strip metadata by default. If your export strips it, change the export setting or use a different export workflow.
6. Upload the checked file to Amazon. Then check the listing after processing, and keep a record of which files were tagged, so a future re-export does not silently remove the tag.
If you outsource editing, tell your editor clearly: "This image contains an AI-generated person. Please keep the contains-synthetic-performer metadata in the delivered file." Do not assume an editor will know, or that a standard export will preserve it. Ask for confirmation that the delivered file still carries the tag.
The export trap: where tags most often disappear
Most missed tags are not forgotten at the tagging step. They are lost later, in an ordinary export. Watch for these four moments:
- A "save for web" or optimised export that strips metadata to reduce file size. If your export dialog has a metadata option, set it to preserve all metadata, or at least keywords and copyright, and then verify the exported file.
- A format change. Converting between formats, or re-saving a JPEG at a new quality setting, can drop metadata if the tool does not carry it across. Always check the final file, in its final format, not an earlier version.
- A resize or crop for a marketplace template. Cropping itself does not remove a keyword, but re-exporting after the crop can. Treat every re-export as a new file that needs a fresh metadata check.
- A round trip through another app. Sending an image through a messaging app, a shared drive preview, or a design tool that recompresses uploads can strip metadata. Download the file back and check it before you upload it to Amazon, rather than assuming the file you sent is the file you received.
A simple habit prevents all four: never upload a file you have not just reopened and checked. The check takes under a minute per file and is described in the next section.
How to check the result yourself
You do not need to take anyone's word that a file is tagged, including an AI tool's or an editor's. Check the exact file you intend to upload, after your final export, with these three checks:
Check 1 — Reopen the file in your editor. Open the exported JPEG or PNG in the same metadata panel you used to add the tag. Confirm that contains-synthetic-performer is listed as a separate keyword or subject term, spelled exactly. If it is missing, or if it appears merged into a sentence of other words, fix it and re-export.
Check 2 — Inspect it with a metadata reader if you have one. A free metadata viewer, or a command-line tool such as ExifTool, will show the IPTC keywords and XMP subject fields directly. You are looking for the exact term in one of those fields. If the tool shows no keywords at all, your export stripped the metadata, even if the image itself looks perfect.
Check 3 — Compare the uploaded file with the file you checked. If you rename, recompress, or re-export the file after checking it, check it again. The file Amazon receives is the file that matters, so your last check should always be on the last version. Keep a short log — filename, tagged yes or no, date checked, and who checked it — so a future team member does not re-export an untagged version by mistake.
If any check fails, do not upload yet. Fix the metadata, re-export with metadata preserved, and run the checks again. A listing delayed by a day for a metadata fix is a smaller problem than a listing live with an untagged AI person.
Common mistakes, and how to avoid each one
Mistake 1: Tagging only the main image. Secondary and lifestyle images are part of the listing too. Fix: work through the full gallery in order, image by image, and mark each file tagged or not tagged in your log before you upload anything.
Mistake 2: Trusting the AI tool's output without checking. Some tools add disclosure metadata, some do not, and some add it to an intermediate file that is stripped on download. Fix: check the downloaded file yourself, using Check 1 or Check 2 above, before you treat the image as compliant.
Mistake 3: Writing the tag on the image instead of in the metadata. A visible "AI-generated" caption or watermark does not put the keyword into the IPTC or XMP subject field. Fix: add the keyword in the metadata panel, and treat any visible note as a separate, additional choice — not a substitute.
Mistake 4: Adding the tag to a real person's photo. Over-tagging creates its own confusion: it tells Amazon's systems that a real person is synthetic. Fix: sort with the 60-second decision rule first. Tag only files that actually contain a photorealistic AI person.
Mistake 5: Burying the keyword in a paragraph. If the term is pasted inside a long description field, or joined to other words without separation, a system reading the subject field may not recognise it as a discrete keyword. Fix: enter it as its own keyword term, exactly spelled, in the keywords or subject field.
Mistake 6: Losing the tag in retouching. You send a tagged file for background cleanup or colour correction, and the editor returns a beautiful file with no metadata. Fix: use the editor brief in the next section, and check the returned file before upload. Make metadata preservation a delivery requirement, not an afterthought.
Mistake 7: Renaming a file and assuming the rename tagged it. Adding "AI" or "tagged" to a filename is useful for your own organisation, but a filename is not metadata. Fix: use filenames only as a tracking aid, and rely on the metadata check for compliance.
Mistake 8: Assuming one marketplace's rule covers all marketplaces. This article covers the Amazon reporting described above. Other marketplaces and your own website may have different disclosure expectations. Fix: check each channel's guidance separately, and keep your tagging log channel-specific if the requirements differ.
What to send your editor: a brief you can copy
Most tagging failures happen in handover, when a seller knows an image contains an AI person but never tells the editor. A short, explicit brief removes that gap. You can adapt this wording for any editor or retouching service:
"Please note: the following files contain a photorealistic AI-generated person and must carry the metadata keyword contains-synthetic-performer in the IPTC/XMP subject field of the delivered file: [list the filenames]. Please preserve that keyword through all retouching and export, and confirm in your delivery message that the keyword is still present in each delivered file. The following files do NOT contain an AI-generated person and should not receive that keyword: [list the filenames]. Please do not add any visible 'AI' text to the images themselves unless I ask for it separately."
That brief does four useful things. It names the exact files, so the editor does not have to guess. It gives the exact keyword spelling, so there is no typo. It separates tagged from untagged files, so real-model images are not over-tagged. And it asks for confirmation, so you receive a checkable statement rather than an assumption.
When the edited files come back, do two things before upload. First, run the metadata check on each returned file. Second, compare the returned images visually with the originals, to confirm that retouching did not change the person in a way that changes the tagging answer — for example, replacing a real model's features with generated ones, or vice versa. If anything about the person changed, re-apply the decision rule to the new file.
Auditing listings that are already live
If you already use AI models, your existing listings need the same treatment as new ones. Work through them in priority order rather than trying to fix everything in one sitting:
1. List every live listing that uses an AI person. Start from your AI tool's export history, your editor's delivery folders, or your own gallery files — whichever record is most complete. If you are unsure about a listing, put it on the list and check it.
2. Prioritise main images first. The main image is the most visible image and the first a shopper and a reviewer sees. Fix untagged main images before secondary images, then work through the rest of the gallery.
3. Download the live file if you can, or check your upload copy. The file that matters is the file Amazon received. If your upload copy is untagged, assume the live file is too, and prepare a tagged replacement.
4. Replace with a properly tagged version. Upload the checked, tagged file through your normal listing-update route. Do not try to edit metadata on a file that is already live on Amazon; replace the file with a verified version.
5. Log the fix. Record the listing, the image position, the date replaced, and the filename of the tagged version. That log is what stops the same image from being re-exported untagged in a future refresh.
If you discover an untagged AI-person image already live, correct it by uploading a properly tagged version, rather than leaving it and hoping it is not noticed. A quiet, prompt correction is the practical response, and it is far easier than explaining a listing problem later.
A simple workflow for teams and repeat uploads
One-off tagging is easy. Repeat tagging, across colourways and seasons, is where errors creep in. A light workflow keeps it reliable without adding bureaucracy:
- Keep two folders per product: an untouched source folder for AI outputs and original photos, and an upload-ready folder for checked, tagged or confirmed-untagged files. Never upload directly from the source folder.
- Use filenames as a tracking aid only. A suffix such as _tagged on an upload-ready filename helps your team find the right file, but it is not a substitute for the metadata itself. The metadata check is still required.
- Keep a one-line log per product. Columns can be as simple as: image position, filename, AI person yes or no, tag checked yes or no, date, and checked by. A spreadsheet with those columns is enough, and it answers the "did we tag this?" question months later without reopening every file.
- Re-check on every refresh. New colourways, seasonal crops, and marketplace template updates all create new exported files. Each new export gets the same metadata check as a first upload, even if the previous version was tagged correctly.
- Assign one checker. If several people export and upload, assign one person to run the final metadata check before upload. A single checkpoint is more reliable than assuming everyone checked.
What happens if you miss it
The reporting does not give a single, guaranteed outcome for a missed tag, and Amazon's own "where applicable" consumer-label criteria are not fully specified in the sources available for this article. That uncertainty is itself a reason to be careful.
Do not rely on a guess about whether a label will appear, whether a listing will be flagged, or whether a penalty will apply. The practical approach is simpler: tag every photorealistic AI person before upload, keep a record, and check Seller Central for your category and marketplace. If you discover an untagged AI-person image already live, correct it by uploading a properly tagged version, rather than leaving it and hoping it is not noticed.
Also remember that disclosure is not only a platform rule. Shoppers increasingly want to know when a model image is AI-generated. Surveys cited in fashion AI coverage have found that many shoppers are comfortable with AI imagery when they are told, but want disclosure. Tagging correctly, and being clear with buyers where appropriate, protects trust as well as compliance.
What about AI models and ghost mannequin images?
AI models and ghost mannequin images serve different purposes, and they should not be confused. A ghost mannequin image shows the garment itself, without a visible person. An AI model image shows a generated person wearing the garment.
That difference matters for Amazon rules as well as for buyers. A ghost mannequin image may be the right choice for a secondary image where you want to show shape and fit without a person. An AI model image may be the right choice for a hero or lifestyle image, provided it is tagged correctly and, where Amazon requires a live model for an adult apparel main image, you follow that separate main-image rule as well. Do not assume that because an AI model image is tagged, it automatically qualifies as a main image in every category. Check both rules: the disclosure rule and the main-image rule.
Use the two formats deliberately rather than by habit. A useful way to decide is to ask what the shopper needs from that image slot. If the slot's job is to show drape, movement, and how the garment sits on a body, an AI model — properly tagged — may serve that purpose. If the slot's job is to show construction, neckline, sleeve shape, or back detail without distraction, a ghost mannequin shot often shows those details more clearly, and it avoids the AI-person question entirely. Many strong galleries use both: a tagged AI-model hero for lifestyle appeal, followed by ghost mannequin and flat-lay images for detail. That mix gives shoppers both the styled view and the unstyled evidence, and it limits the tagging work to the images that actually contain an AI person.
Frequently asked questions
Does the tag show to shoppers on the listing page? On the basis described in the reporting, the tag lives in the file's metadata, not as visible text on the image. Whether Amazon also shows a consumer-facing label, and under what criteria, is the "where applicable" question that was not fully specified in the sources for this article, so do not promise a shopper-facing label outcome either way. Check Seller Central for your category and marketplace for the current position.
Does tagging an AI model make it eligible as a main image? No, not by itself. Disclosure and main-image eligibility are separate rules. Some categories require a live model for an adult apparel main image, and a tagged AI model does not automatically satisfy that separate requirement. Check both rules for your category before choosing a main image.
Do I need the tag for an AI-generated background behind a real model? No, not on the basis described here, because there is no AI-generated person in the image. The background is AI-generated, but the person is real. Keep the distinction person versus background in mind for every image.
Do I need the tag if the AI person is only in a video or a comparison image? This article covers the listing-image reporting described above. Video, A+ content, and storefront images may be treated differently, and the sources for this article do not fully specify those cases. Check Seller Central guidance for each content type before you assume the same rule applies, or does not apply, to it.
My AI tool says its images are "Amazon compliant." Can I trust that? Treat that claim as a starting point, not a verification. Download the file, run the metadata check described above, and confirm the exact keyword is present in the exact file you will upload. If the tool's claim is correct, the check will confirm it in under a minute. If it is not, you have caught the problem before upload.
What if I bought AI model images from a freelancer or an agency? Ask them directly whether the model is AI-generated or a real person, and ask for the answer in writing for your log. If they confirm an AI person, tag the files yourself or instruct your editor using the brief above. If they cannot confirm, treat the image as "check first" and do not upload it as untagged.
What if the same garment appears on both a real model and an AI model? Tag only the AI-model files. The real-model files do not receive this keyword, even though they show the same garment. File-by-file tagging, not product-by-product tagging, is the correct approach.
Should I delete an old, untagged AI-model listing and start again? Not necessarily. The practical fix described in this article is to replace the untagged file with a properly tagged version through your normal listing update. Deleting a listing can lose reviews and ranking history, so replacement is usually the less disruptive correction — but check Seller Central guidance for your situation before deciding.
A simple decision checklist for your next upload
Before you upload a new gallery, ask:
- Does any image contain a person who looks real but was generated by AI? If yes, tag that file.
- Does any image contain only a real person, an AI background, or AI retouching with no AI person? If yes, this particular people tag does not apply to that image on the basis described here.
- Have I checked the metadata of the exact file I am uploading, after my final export?
- If I outsourced the edit, did I tell the editor about the AI person and ask them to confirm the tag is preserved?
- Have I logged which files are tagged and which are not, so a future re-export does not undo the work?
- If any image is ambiguous, have I held it back or treated it as needing a tag until I have an answer?
If you can answer all six, you have handled the main risk in this rule. If you cannot answer the third question, do not upload yet. A few minutes checking metadata is far cheaper than fixing a listing after the fact.
Try it on your own listing
If you are unsure whether your AI model images are tagged correctly, test the process on one listing first — and that is exactly where working with an editor who understands this workflow helps. Send one gallery, flag which images contain an AI-generated person, and ask the editor to return files with the contains-synthetic-performer metadata preserved and confirmed, alongside the usual background cleanup, colour correction, and ghost mannequin work. You can then run the metadata checks above on the returned files yourself, so you know the process works before you apply it across a catalogue.
Fix The Photo Online offers a free trial: one free image, no credit card required, typically completed in 2–4 hours. Standard turnaround is 24–48 hours, rush turnaround is 6–12 hours, and support is available 24/7.
When you send an image, tell the editor whether it contains an AI-generated person, so the metadata can be checked and preserved in the delivered file. Contact: contact@fixthephotoonline.com. You can also start from the Free Trial page or the Photo Editing Services page on fixthephotoonline.com.
Note: this article summarises reporting available in July 2026 and is not legal advice. Amazon's requirements can change, and category-specific guidance in Seller Central should be checked before you upload.
Fix The Photo Online
Professional photo editing & retouching services for everyone. Email: contact@fixthephotoonline.com
© 2026 Fix The Photo Online. All rights reserved.
Services
Company
Contact
Email: contact@fixthephotoonline.com Phone: +1 (419) 506-5585 117 Jefferson Avenue #16, Brooklyn, NY 11216