EU AI Act — marking is the provider's duty, deep fake disclosure the deployer's
Article 50(2) of the EU AI Act applies from 2 August 2026: “Providers of AI systems … generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.” Under Article 111(4), added by Regulation (EU) 2026/1744, providers of systems placed on the market before 2 August 2026 have until 2 December 2026 to comply.
The marking duty sits with the provider of the AI system, the tool that generates the image. A seller who uses such a tool should keep that marking intact. A separate duty sits with the deployer: under Article 50(4) and (5), a deployer must disclose a deep fake clearly and distinguishably, at the latest at first exposure, and the Commission's non-binding Guidelines say machine-readable marking does not meet that duty. Whether a photorealistic AI model in a product image counts as a deep fake is unsettled.
Google Merchant Center — IPTC metadata, kept in the file
Google requires that all images created using generative AI contain metadata indicating that they are AI-generated, using the IPTC DigitalSourceType tag — for example TrainedAlgorithmicMedia or CompositeSynthetic.
Google also asks sellers not to remove embedded metadata such as the IPTC DigitalSourceType property. AI-generated images are accepted in image_link, additional_image_link and lifestyle_image_link.
Amazon — a keyword for AI-generated people
On 23 July 2026 Amazon told sellers to tag images and videos that show photorealistic, AI-generated people with the keyword contains-synthetic-performer in the file's embedded metadata, added with an IPTC-compatible editor. Real people, including AI-edited photos of real people, are excluded.
The trigger is New York General Business Law section 396-b, in effect since 9 June 2026, which requires disclosure of photorealistic AI-generated people in advertising and sets civil penalties of $1,000 for a first violation and $5,000 for subsequent ones.
Zalando — invisible marking, no visible labels
Zalando does not allow visible on-image labels and says not to provide content that requires visible labelling under applicable laws, naming deepfakes. It strongly recommends invisible marking of all AI-generated content, under a heading that says required by December 2026.
Its guidelines also list distortions that are not acceptable: an inaccurate representation of the product, implausible settings and an inconsistent model appearance.
Labelling AI-generated content: metadata or a visible label
There are two ways to label an AI-generated image. A visible label is text or a badge on the picture itself, seen by every buyer. Embedded marking is a value written into the file, such as the IPTC digital source type, which software, platforms and checkers can read but buyers do not see. The platform rules above rely on the second for product images, and one of them does not allow the first on the picture.
The two serve different readers. Metadata lets a platform sort and disclose AI content in its own interface, and it lasts only as long as the file does. A visible label reaches the person looking at the image but changes the picture, which a platform may refuse. Where a rule asks for disclosure that people can perceive, whether a note in the listing text meets it depends on the wording of that rule.
Keeping the AI label intact through editing and resizing
A marker written into a file stays only while that file is kept. Re-saving in an editor can drop it, an export for the web often strips metadata to make the file smaller, a screenshot creates a new file with none, and messaging apps usually recompress images when they are sent. The safe order is to finish every edit first and to check the final file, the exact one that will be uploaded, last.
Every image Herofab delivers already carries the IPTC digital source type inside the file, written without recompression, and the synthetic-people keyword is added where the image shows a person Herofab generated. Two of the free browser tools on this site, Marketplace image resizer and Image compressor, write a new file and do not carry embedded metadata over. AI image label checker reads the markers in a file and lists which are present.
Records worth keeping for every AI product image
A disclosure question often arrives long after an image was published, from a platform review or from a customer. A simple record for each image answers it: the source photo of the real product, the delivered image as downloaded, the tool that made it and the date. The source photo shows what the product really looked like; the original file shows that the marker was present when the image left the tool.
The record also supports the oldest rule of product images, which every platform above repeats: the product shown must match the product that ships. Comparing each AI image with the physical product at full size, for colour, shape, logo and parts, is the check that no marker replaces. An image that fails it is replaced before it is published, whether it is labelled or not.
Checklist
- Keep the metadata: re-saving an image in an editor or a resizer can drop the AI marker that Google asks sellers not to remove.
- For Amazon, images that show photorealistic AI-generated people need the contains-synthetic-performer keyword in their metadata.
- For Zalando, use invisible marking and no visible label on the picture.
- Everywhere: the product in the image must match the product that ships.
Sources
- EU AI Act, Article 50
- Google Merchant Center Help — AI-generated images
- eWeek — Amazon requires sellers to disclose AI-generated people
- CNBC — Amazon makes sellers label AI-generated people
- Zalando Partner University — Image guidelines
Checked 24 September 2026. Rules change; the official pages are the final reference.