What Article 50 actually requires
Two separate duties sit in Article 50, and they land on different parties.
Providers of generative AI systems must mark synthetic image, audio, and video outputs in a machine-readable format so they are detectable as artificially generated. In practice this means embedded provenance metadata — C2PA Content Credentials, invisible watermarking, or equivalent. This is the model vendor's job, not yours.
Deployers — that's you, the merchant publishing the image — must disclose that content is artificially generated or manipulated when it depicts real-looking people, objects, places, or events. The disclosure must be clear, distinguishable, and delivered no later than first exposure.
Non-compliance with the transparency obligations carries fines of up to €15 million or 3% of global annual turnover, whichever is higher — plus exposure to unfair-competition claims from competitors in several member states.
Note what the regulation does not do: it does not mandate a specific badge, wording, or placement. There is no official EU logo. You choose the format; regulators judge whether an average shopper would notice and understand it.
The exemption that decides most of your catalog
Article 50(2) exempts AI systems performing "an assistive function for standard editing" or that do "not substantially alter the input data provided by the deployer or the semantics thereof." For product photography, that single sentence separates routine retouching from synthetic content.
Generally exempt (assistive editing)
- Background removal or replacement with plain white
- Colour correction and white balance to match the real product
- Dust, lint, and sensor-spot cleanup
- Wrinkle reduction on a garment you actually photographed
- Upscaling, sharpening, cropping, format conversion
- Shadow cleanup and reflection removal
Requires disclosure
- A photorealistic AI model wearing your garment
- Fully generated product shots with no source photograph
- Synthetic lifestyle scenes that read as a real location
- Swapping a real model's face, body, or ethnicity
- Ghost-mannequin output that invents unseen garment interiors as a realistic scene
- Generated colourways depicting a variant you never produced
The dividing question is semantic, not technical: does the finished image assert something to the shopper that the camera did not capture? A white background asserts nothing. A smiling person who does not exist asserts a great deal.
Borderline cases — heavy AI background generation that reads as a real studio or location — should be labeled. The cost of an unnecessary label is small; the cost of a missing one is a regulatory finding.
Where to put the disclosure without killing conversions
The most common merchant fear is that a label depresses conversion. The evidence so far points the other way when the label is specific rather than ominous. Vague warnings read as risk; precise statements read as confidence.
| Placement | Compliance strength | Conversion impact |
|---|---|---|
| Caption directly under the image | Strong | Minimal |
| Small corner badge on the image itself | Strong | Minimal |
| Line in the product description | Adequate if above the fold | Negligible |
| Sitewide policy page only | Insufficient | None |
| Full-width interstitial before viewing | Strong | Significant drop |
Wording matters more than placement. Compare these two, both technically compliant:
- Weak: "Warning: this image was created by artificial intelligence."
- Strong: "Shown on an AI-generated model. The garment, colour, and fabric are photographed from the actual product."
The second discloses fully and reassures the shopper about the thing they actually care about — whether the product itself is real and accurately represented.
Machine-readable marking: what to check with your vendor
Even though the marking duty falls on the AI provider, you inherit the practical consequences. Platforms increasingly read provenance metadata automatically, and stripped metadata can trigger their own auto-labels — often less flattering than yours.
Ask any AI imaging vendor three questions:
- Does the output carry C2PA Content Credentials or an equivalent machine-readable marker? Get the answer in writing; it is your evidence of upstream compliance.
- Does the marker survive your pipeline? Most CDN resize and re-encode steps strip metadata by default. Test an end-to-end round trip from generation to live product page.
- Is there an audit trail per image? When a regulator or marketplace asks which images were AI-generated, a per-asset record beats reconstructing history from memory.
Run one generated image all the way to your live PDP, then inspect it with a Content Credentials verifier. If the provenance is gone, your image-optimisation step is stripping it — configure the CDN to preserve metadata for AI-generated assets, or keep your own database record as the fallback audit trail.
Platforms like Retouchable that maintain a per-asset generation record make this considerably easier than reconstructing provenance from a folder of exported files.
A practical compliance workflow
Blanket-labeling every image is the wrong response — it dilutes the signal and misrepresents genuinely photographed products. Classify instead.
Illustrative distribution for a mid-size apparel brand using AI models for secondary gallery shots. Your split will differ; the point is that most catalogs are majority-exempt.
- Inventory. Tag every image in your DAM as camera-original, assistively edited, or synthetic.
- Classify against Article 50(2). Apply the semantic test: does it assert something the camera did not capture?
- Template the disclosure. One approved caption per image type, so nobody improvises wording per listing.
- Automate the join. Bind the label to the image record, not the listing, so the disclosure follows the asset into every channel — feed, marketplace, email, ad.
- Verify metadata survives. Re-check after any CDN or theme change.
- Keep the record. Retain generation logs; they are your defence if the classification is challenged.
What this means outside the EU
The AI Act applies where output is placed on the EU market, regardless of where you are established. A US brand shipping to Germany is in scope for those listings.
Practically, most brands converge on one global standard rather than maintaining EU-only image variants. Maintaining two catalogs — one labeled, one not — multiplies asset management overhead and creates the exact inconsistency that draws scrutiny.
The wider direction of travel supports that choice. Major marketplaces already require AI-content declarations, ad platforms auto-detect and label synthetic media, and several non-EU jurisdictions have transparency bills in progress. Building the labeling discipline once, against the strictest standard in force, is cheaper than retrofitting it per market.
Treat disclosure as product information rather than legal boilerplate, and it stops being a tax. Shoppers who know the model is generated but the fabric photograph is real return fewer items than shoppers who feel misled after delivery.