Amazon's AI Image Disclosure Rule, and How to Add the Tag

What does Amazon's AI image rule actually require?

In its late-July announcement Amazon told sellers this, in its own words: "you must now tag product images that contain photorealistic AI-generated people with specific metadata before you add them to your listings and A+ content in any of our worldwide stores". The instruction is specific. Use an IPTC-compatible metadata editor to add the keyword contains-synthetic-performer to the dc:subject (XMP) field of the image or video file. Not the filename, not the alt text, and not a field in Seller Central. Inside the file itself.

Amazon says it will "add an indicator to inform customers that the image includes AI-generated people, where applicable". That is the company's own wording and it is worth keeping in the future tense. There is no published example of the indicator, no screenshot from a live detail page, and nothing describing where it sits. A forum post is also the whole of the documentation, so there is no help page to check any of it against.

Amazon has not said what happens to untagged content that should have been tagged. There is no stated rejection behaviour and no suppression warning, and no surface in Seller Central where you can confirm Amazon read the tag after upload. You can prove the keyword is in your file before you upload it. After that you are trusting the pipeline. Treat the tag as a cheap insurance step rather than something you can audit.

When do you not need to tag an image?

Amazon lists four situations where no tagging is needed. Read them as a set. If any one of them describes your file, the rule does not reach it. Most seller imagery falls outside the rule completely, because most seller imagery has no people in it at all. The trigger is narrow and specific: a photorealistic person who was generated by AI and is not a real human being.

Two of those exceptions cover most of the AI imagery sellers actually make. An AI-generated background behind a real product photo is out of scope. A fully AI-rendered studio shot of a bottle on a plinth is out of scope. A stylised illustration of a person, drawn rather than photographed, is out of scope because it is not photorealistic. What is in scope is the AI model holding your product, or the AI hand applying your cream.

An Amazon moderator has added two useful clarifications in the seller forums. Amazon treats simple AI retouching, in its examples background removal, colour correction and lighting, like traditional photo editing, so it needs no declaration. And already-published content is grandfathered: "You are not required to include this information retroactively on already-published content. Compliance is mandatory going forward for new uploads." So the tag matters the next time you upload, not for your back catalogue.

• ✓Content that only features real people, even if they have been altered with AI tools
• ✓Content that only features characters from movies, TV programs, streaming content, documentaries, video games, or other expressive works
• ✓Content that does not feature any people
• ✓Content that does not feature photorealistic people

How do you add contains-synthetic-performer, tool by tool?

One detail decides whether your command works or quietly destroys your existing keywords. In the IPTC Photo Metadata Standard, Keywords maps to dc:subject as an unordered bag of text, which is to say a list. So you append to it. In ExifTool that is the += operator. A plain = assigns instead, and silently wipes every other keyword already on the file, including anything your studio or your VA depends on.

Video is genuinely doable, which most coverage of this rule skips. Amazon's instruction covers image or video files. ExifTool's supported formats table lists XMP as read, write and create for MP4 and MOV as well as for JPEG and PNG. Older copies of the ExifTool manual still circulating on other sites show MOV as read only. That is out of date, and it is the reason people conclude video cannot be tagged at all.

How do you check the tag survived your export?

Read the file back. Running exiftool -XMP-dc:Subject image.jpg prints the keywords that are actually in the file. If contains-synthetic-performer comes back in that list, you are done. If nothing prints, the tag is not there, whatever the app told you when you saved. Run this on the exact file you are about to upload, not on the master you edited.

The reliable way to lose the tag in Photoshop is File, then Export, then Export As. That path strips keywords and description even though the dialog offers metadata options. Save for Web (Legacy) has a Metadata dropdown, and anything short of All can drop your keywords, so set it to All. PNG saves have historically dropped metadata too. None of this is consistent enough to trust from memory, which is exactly why the read-back check matters.

The fix is sequencing. Write the metadata last, after every crop, resize and export step is finished, then verify. Every stripping risk in the pipeline sits between creation and final export, so tagging early guarantees rework. Re-exporting carries its own risk. It is easy to break Amazon's baseline image specs in the same pass, so check the frame fill and the longest side before you upload.

Does the A+ Content uploader add the AI tag for me?

For those assets, yes. The Creative Assets uploader in A+ Content Manager presents checkboxes for "AI-generated" and "AI-generated people". Tick the AI-generated people box and Amazon applies the necessary metadata to the asset for you. That is the only place Amazon has documented a control for this, and it is worth using, because it removes the export-strips-metadata problem for those assets entirely.

The main image gallery is different. No Amazon post or help page describes an equivalent control on the standard gallery upload path. That is an observed absence rather than a stated policy, so treat it as a working assumption: for gallery images, assume the metadata has to be in the file before you upload it. Writing the keyword yourself costs one command and works everywhere, including any path where a checkbox might appear later.

Amazon is not detecting synthetic people in your uploads and tagging them for you. The checkbox is you declaring it. The file metadata is you declaring it. Whatever tool made the image, Picmato included, the tagging step belongs to the seller. Build it into the handoff: the person who exports the final file is the person who runs the check.

Where did this rule come from? New York's synthetic performer law

Amazon's rule landed a few weeks after a New York law took effect. The synthetic performer disclosure duty was added to General Business Law section 396-b by bill S8420-A and its Assembly twin A8887-B. Governor Hochul signed it on 11 December 2025 as Chapter 617 of the Laws of 2025. It takes effect "on the one hundred eightieth day after it shall have become a law", which is 9 June 2026. If you go looking for the bill, note that S1228C reads almost identically, was never enacted, and is still sitting in committee.

The duty itself is short. A person who advertises "shall conspicuously disclose in such advertisement if a synthetic performer is in such advertisement, where such person has actual knowledge". A synthetic performer is a digitally created asset made with generative AI. It is intended to give the impression of a human performance by someone who is not a recognisable real performer. The statute sets the civil penalty at $1,000 for a first violation and $5,000 for any subsequent violation.

The exemptions explain why this landed on you rather than on Amazon. The publishing medium, meaning the newspaper, network or platform carrying the advertisement, is exempt. Audio-only advertisements are exempt, and so is using AI purely to translate a real performer's speech. There is no private right of action, so enforcement comes from the state, not from shoppers. Whether a product detail page image is an advertisement under the statute has not been tested. This is not legal advice.

Amazon, the EU AI Act and Walmart, side by side

The EU AI Act's Article 50 transparency obligations became applicable across the EU on 2 August 2026, and were not delayed. The Digital Omnibus on AI, reported as Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026. It pushed the high-risk deadlines out to 2 December 2027 and 2 August 2028, and left Article 50 alone. It did add a grace period: systems placed on the EU market before 2 August 2026 have until 2 December 2026 before the Article 50(2) marking duty applies. The regulation is on EUR-Lex under that number, and it entered into force on 27 July 2026.

Who owes what matters here, because two duties get mashed together constantly. The machine-readable marking duty in Article 50(2) sits on the provider, the party that develops the AI system and puts it on the market. That is your image tool's vendor, not you. As a seller you are a deployer, and your duty is Article 50(4). If you generate or manipulate a deep fake, disclose that the content is artificially generated or manipulated, clearly and at first exposure.

The penalty band gets misquoted constantly. For an Article 50 breach it is up to EUR 15 million or 3 percent of total worldwide annual turnover, whichever is higher, under Article 99(4). The widely repeated "EUR 15 million or 7 percent" does not exist in the Act: 7 percent pairs with EUR 35 million and covers prohibited practices under Article 5. Walmart, meanwhile, asks for nothing. Its April 2026 AI compliance note and its US image guidelines page, last updated 12 May 2026, require only that AI images be truthful, accurate and not misleading.

Why did Amazon ask for a keyword instead of the standard AI field?

There is a purpose-built field for exactly this job, and Amazon did not use it. The IPTC Photo Metadata Standard defines Digital Source Type. Two of its controlled values matter here: trainedAlgorithmicMedia, for media created algorithmically by an AI model trained on captured content, and compositeWithTrainedAlgorithmicMedia, for augmentation or correction using a generative model. That property is what Adobe, Google and the C2PA content credentials ecosystem write when they mark AI output.

Amazon asked for a free-text keyword in dc:subject instead. That choice has one advantage and one real drawback. The advantage is that every keyword editor in existence can set it, including free ones, so no seller is locked out by tooling. The drawback is that it is a weak signal: a careless export wipes it and a typo defeats it, and it carries no information about how the image was made. It also means a tool that already writes correct AI provenance can still fail Amazon's rule.

So check both ends. Confirm the exact spelling, contains-synthetic-performer, lowercase with hyphens, and confirm it survived the export by reading the file back. Keep one tagged master per asset, so a re-upload never becomes a re-tagging job. And keep the scope in mind. This is a narrow rule about photorealistic AI people, not a general disclosure duty for AI-assisted product photography.