What actually counts as an AI-generated product image
Article 50 targets content that is artificially generated or manipulated in a way that a reasonable viewer would take as a real depiction. That framing does most of the work. Editing a photograph is not generating one.
The practical test: did the AI invent visual information that was never in front of a camera, and would a shopper reasonably believe it was photographed?
Generally in scope
- AI-generated lifestyle scenes and backdrops
- AI models wearing your garment
- Ghost mannequin effects synthesized from a flat lay
- Products composited into rooms or environments that were never shot
- Generated variant colorways presented as photographs
Generally out of scope
- Background removal and white-background cutouts
- Color correction and white balance
- Dust, lint, and blemish removal
- Upscaling, sharpening, denoising
- Cropping, straightening, resizing
The second column is ordinary retouching. It has been standard practice since long before generative models, it does not fabricate a scene, and it does not mislead a buyer about what the product is. A cutout of a real shoe on white is still a photograph of that shoe.
Generative fill used to extend a real photo — widening a backdrop to hit a 1:1 crop, filling a corner of a seamless sweep — sits between the two columns. The safe read: if the fill is a continuation of a surface that genuinely existed in the frame, it behaves like retouching. If it adds objects, environments, or people, treat it as generated and label it.
The two obligations people keep conflating
Article 50 imposes two separate requirements, and they have very different scopes. Most of the confusion in brand Slack channels comes from treating them as one rule.
| Obligation | Form | Who it applies to | Scope |
|---|---|---|---|
| Machine-readable marking | Embedded metadata, invisible to shoppers | Providers of the AI system | Essentially all synthetic images |
| Visible disclosure | On-page label a shopper can read | Deployers — you, the brand | Deepfakes: realistic depictions of identifiable real people or events |
The first is largely handled upstream. If you use a reputable AI imaging tool, the provenance marking should already be embedded in the file it hands you — your job is to verify it survives your pipeline, not to create it.
The second is where brands have real work to do, and it's narrower than most assume. A generated backdrop behind a coffee grinder does not depict an identifiable real person. An AI model rendered from scratch, who resembles no actual human, generally doesn't either. But if you generate imagery based on a real person's likeness — a founder, an athlete, a licensed ambassador — you are squarely in visible-disclosure territory.
Most catalogs need metadata integrity across the board and visible labels on a small subset. Audit for the subset rather than stamping a disclaimer on every product page — over-labeling trains shoppers to ignore the label and can itself read as misleading.
How C2PA provenance metadata works
The machine-readable requirement is met in practice through C2PA — the Coalition for Content Provenance and Authenticity standard, backed by Adobe, Microsoft, Google, OpenAI, and the major camera manufacturers. It is the de facto answer to "how do I mark this in a detectable way."
A C2PA manifest is a cryptographically signed record attached to the image file. It records what created the image, whether a generative model was involved, and what edits were applied afterward. Because it's signed, tampering is detectable — you can't quietly strip the "AI-generated" assertion and keep a valid manifest.
The catch, and it's a serious one for e-commerce: most image pipelines destroy it.
Standard image optimization — the exact thing every e-commerce team does for page speed — strips non-essential metadata by default. A brand can generate a properly marked image, upload it, serve it through a CDN, and end up shipping a stripped file to every shopper. The obligation sits with the AI provider, but a pipeline that silently discards provenance undermines the whole regime and leaves you unable to prove what you did.
Fixing it is a configuration problem, not a rebuild: most image CDNs and build tools have a metadata-preservation flag that is simply off by default.
Building labeling into your catalog workflow
The efficient approach is to classify at generation time rather than audit at publish time. Once an image is sitting in your DAM alongside 30,000 others, working out how it was made is archaeology.
1. Tag at the source
Every image entering your asset library gets one of three flags: photographed, retouched, or generated. This is a single field, and it answers every compliance question you will be asked later. Tools that produce generated imagery should write this automatically — Retouchable, for instance, distinguishes retouching operations from generative ones in its job records, so the classification comes from the pipeline rather than someone's memory.
2. Preserve metadata through delivery
Audit each transform step in your delivery chain and enable metadata preservation. Test by pulling a live production image URL and inspecting it with a C2PA verification tool — do not assume the flag worked.
3. Identify your visible-disclosure subset
Filter for generated assets that depict identifiable real people. For most catalogs this is a small, knowable list: ambassador campaigns, founder imagery, licensed likenesses. Those get an on-page label.
4. Write the label like a human
Disclosure needs to be clear and accessible, not buried in a footer. It also doesn't need to be defensive.
| Weak | Better |
|---|---|
| "Image may contain AI elements." | "Model imagery generated with AI. Garment shown is the actual product." |
| "AI" | "Scene created with AI. Product photographed in studio." |
| Buried in terms of service | Caption directly beneath the image |
5. Keep records
Retain generation logs — model used, prompt or input, date, source photograph. If a regulator or marketplace asks how an image was produced, a timestamped record is the entire answer.
What this means beyond the EU
If you sell into the EU, the rules apply regardless of where you're headquartered. But the broader shift is not really regulatory — it's platform-driven, and it's moving faster than legislation.
Major social and marketplace platforms already read provenance metadata and apply their own "AI info" labels automatically, without asking you. That means the label appears whether or not you were planning to disclose — and a brand that gets auto-labeled after saying nothing looks worse than one that disclosed up front.
Meanwhile, existing consumer protection law already covers the substantive risk. Any image that misrepresents the product — a color you don't sell, a fit that isn't achievable, an included accessory that isn't in the box — was actionable long before the AI Act existed. Generative tools make that failure mode easier to hit by accident, which is the real reason to keep generation confined to context rather than product.
Generate the scene, never the product. Backdrops, lighting environments, and staging can be synthetic. The item itself — its shape, color, texture, hardware, and proportions — must match what ships. That single constraint keeps you compliant under transparency rules, consumer protection law, and marketplace policy simultaneously.
A 30-minute compliance audit
You can get a realistic picture of your exposure in one sitting. Work through it in order.
| Step | What to check | Time |
|---|---|---|
| 1 | List every tool in your image pipeline that can generate, not just edit | 5 min |
| 2 | Estimate what share of live catalog images came from those tools | 10 min |
| 3 | Pull three live product image URLs, inspect for a C2PA manifest | 5 min |
| 4 | Identify any generated imagery depicting real, identifiable people | 5 min |
| 5 | Confirm your DAM has a field distinguishing generated from retouched | 5 min |
Step 3 is where most teams get a surprise: the manifest was there at upload and gone at delivery. Step 5 is the one that pays off longest — the brands that will find the next round of rules trivial are the ones that already know, per asset, how each image was made.
None of this requires abandoning AI imagery. Generated backdrops, on-model visuals, and synthetic staging remain the fastest way to build a catalog, and the transparency rules were written to make them trustworthy, not to make them unusable. The cost of compliance is a metadata flag and a database column — not a return to a rented studio.