AI-generated product image disclosure has been enforceable in the EU since 2 August 2026, and the rule most suppliers worry about is not the one that exists. Article 50 does not put an "AI" sticker on every generated product photo. It splits the duty in two: the tool provider must mark synthetic output in a machine-readable way, and you, the supplier publishing the image, must visibly disclose it only when the picture is a deepfake.
Most catalog teams prepare for the wrong check: a caption on a plain studio shot, while a generated living room ships as the customer's actual install. One is paperwork. The other is exposure.
What AI-generated product image disclosure actually requires
Article 50 of Regulation (EU) 2024/1689 is a transparency rule, not a ban. It creates four duties, and only one of them lands on the supplier as a visible label.
| Duty |
Who carries it |
What it looks like in practice |
Date |
| Mark synthetic output in a machine-readable format |
Provider of the AI system (the tool or model maker) |
Metadata, watermark or signal a detector can read |
2 August 2026; 2 December 2026 for systems already on the market before that date |
| Disclose a deepfake visibly |
Deployer (you, if you publish the image) |
Clear, distinguishable label at first exposure |
2 August 2026 |
| Label AI-generated text on matters of public interest |
Deployer |
Rare in product catalogs |
2 August 2026 |
| No retroactive labelling |
— |
Content generated before 2 August 2026 |
— |
Two details catch suppliers. The machine-readable mark belongs to the tool provider: if your background tool writes none, that is a vendor question, not a reason to stamp your catalog. And the visible disclosure duty is narrower than the marking duty, triggered by deepfake content rather than by generation itself.
The Commission's own FAQ on Article 50 puts the deployer duty this way: deployers must clearly label deepfakes and AI-generated or manipulated text published on matters of public interest. Product images are not public-interest text. They become relevant when the image itself is a deepfake.
Myth 1: every AI-generated product image needs a visible label
This myth produces the worst compliance theater: a small "AI" badge burned over the product, doing nothing for the buyer and everything to hurt conversion.
The visible half of AI-generated product image disclosure is a deepfake remedy. Article 50(2) already obliges the provider to mark the output so detection tools can identify it. Article 50(4) adds a human-readable disclosure when the content is a deepfake. Those are different jobs for different people. A generic generated backdrop that no one would mistake for documentary evidence does not need a visible "AI" stamp to satisfy Article 50.
It may still need a marketplace disclosure or an accuracy check. Those are separate rulebooks with separate consequences; do not merge them into one badge.
Myth 2: a generated background turns the whole photo into a deepfake
A deepfake, as defined in Article 3(60), is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.
The Commission's guidance breaks that into three cumulative tests:
- Resemblance. There is a high level of similarity between the image and the thing it simulates.
- Existing. The simulated person, object, place, entity or event resembles something that exists, could plausibly exist, or could have plausibly existed.
- False appearance. The image would deceive or mislead a person about its authenticity or truthfulness.
All three must hold. The guidance also says that AI-generated or manipulated background scenes, special effects and technical pre- and post-processing, in the context of standard production, are not likely to make content falsely appear authentic. A neutral studio sweep is the easy case; a generic living room usually is too.
What flips it:
- A room that recreates a specific real showroom, hotel lobby or a named customer's store.
- A generated person presented as a real model, installer or customer.
- A scene captioned as an installed reference project when no such installation exists.
- A "lifestyle" image that implies a measured fit, finish or material the physical sample does not have.
The last one is not strictly a deepfake problem. It is a product-accuracy problem, and it survives every exemption in Article 50.
Myth 3: the standard-editing exemption covers any retouch
Article 50(2) does not apply when the AI system performs an assistive function for standard editing, or does not substantially alter the input data or its semantics. That is the exemption most editing tools lean on, and it is narrower than "I used AI, so it's fine."
| Operation |
Usually standard editing |
Usually beyond it |
| Crop, straighten, exposure, white balance |
Yes |
— |
| Remove or replace a background |
Yes, when the product pixels are untouched |
— |
| Remove a scratch or dust spot |
Yes |
— |
| Change the product's colour, material or finish |
— |
Yes |
| Add, remove or reshape a part |
— |
Yes |
| Generate a product that was never photographed |
— |
Yes |
| Put a generated person into the shot |
— |
Yes, and run the deepfake test |
The exemption belongs to the provider's marking duty; it does not give you permission to show a product that differs from the one in the carton. A background swap that leaves the measured product alone is the safe side of the line.
There is also a narrow exemption in the Commission's guidance for outputs used in business-to-business or industrial contexts, but it is conditional and it is about the marking duty, not about what you may claim to a buyer. Treat it as a vendor-side argument, not a catalog-wide pass.
Myth 4: if the file is marked, the listing is covered
The Commission FAQ is explicit: for a deepfake, deployers cannot simply rely on the machine-readable marking embedded by the provider. The disclosure has to be clear and distinguishable to a person, visible or audible, at the latest at first exposure.
The machine-readable mark tells a detection system what the file is. The visible label tells a buyer what the picture is not. A buyer scrolling a product page will never open your metadata.
Myth 5: images made before 2 August 2026 are safe forever
Content generated before 2 August 2026 does not need to be labelled retroactively. That is a date-of-generation exemption, not a permanent pass for the SKU.
Regenerate the scene at a new aspect ratio, restyle it for another market, or produce a new variant, and you have new output. Check whether the tool treats it as a new generation, and remember that accurate representation, marketplace rules and the measured product were never exempt.
Amazon already asks for a label Article 50 does not
Marketplace policy is where this stops being theoretical. Amazon's seller guidance in 2026 treats simple AI retouching, such as background removal, colour correction and lighting adjustment, the same as traditional photo editing, as long as the final image accurately represents the product. Images containing photorealistic AI-generated people are different: full bodies, faces and hands must be disclosed, and A+ Content uploads carry explicit "AI-generated" and "AI-generated people" checkboxes.
Two conclusions follow. First, a platform can require a tag that Article 50 does not, so "the regulation doesn't make me label this" is not a defence on a listing. Second, the platform's disclosure is not retroactive either: it applies to new uploads, and the sensible move is to add it the next time you edit an old listing.
Answer the platform question and the regulation question separately: a generated model shot triggers both; a generic backdrop usually triggers neither visible label, but still has to pass the accuracy check.
The three-question test before you publish
Run these three questions for every generated or AI-assisted image. They map to the deepfake criteria.
| # |
Question |
If yes |
If no |
| 1 |
Does a real-looking person appear in the image? |
Add a visible disclosure; treat it as the highest-risk case |
Continue |
| 2 |
Does the scene reproduce an identifiable real place, store, property or event? |
Add a visible disclosure, or replace the scene with a generic one |
Continue |
| 3 |
Would a reasonable buyer read this as a real photograph of the product in that location? |
Add a visible disclosure and keep the proof of what was generated |
No Article 50(4) label needed |
"No label needed" is not "no work needed." You still have to answer for the product: its dimensions, material, finish, capacity and included parts.
The image that actually costs you the order: the product, not the background
The deepfake label is a narrow legal tool with a narrow trigger. The commercially dangerous image is the one where the scene is honest and the product is not.
A generated sofa in a beautiful room is a presentation choice. A generated sofa that appears 20 cm wider than the shipped model is a returns engine. A generated wood grain that reads as solid oak when the carton holds oak veneer is a dispute waiting for a chargeback. No Article 50 disclosure fixes either one, because the buyer's complaint is not "this was AI" — it is "this is not what I bought."
That is why the measured layer has to stay separate from the generated layer. The photograph of the real product, with its measured dimensions locked onto the actual pixels, is the evidence. The scene around it is styling. A product visuals workflow that keeps those two layers apart lets you change the room without changing the facts.
If you generate scenes, run the same checks a buyer would: proportions, part count, hardware, finish and the direction of the grain. A seven-point checklist for a generated product scene catches most of the failures before they reach a listing. And if someone asks you to prove size from a generated render, the honest answer is that AI-generated images cannot establish measured dimensions, no matter how sharp they look.
What goes in the compliance file
Keep one folder per market, not one per campaign. The file should answer a regulator's question and your buyer's question without a meeting.
That last line is not padding: Article 50 and a marketplace policy can both apply to the same file, and satisfying one does not satisfy the other.
How to write a disclosure that does not wreck the listing
Article 50(5) requires the information to be clear and distinguishable, at the latest at the time of first exposure, and to meet accessibility requirements. In practice that means a human-readable line near the image, not a watermark over the product and not a note buried in the spec table.
| Weak |
Why it fails |
Better |
| "AI" badge over the product |
Obscures the thing being sold, says nothing about what is generated |
"Room setting generated; product and dimensions as photographed." |
| "AI-generated image" in the alt text only |
Most buyers never see alt text |
Same line in the visible caption, with a descriptive alt text |
| "Illustrative purposes only" |
Vague, does not identify what is synthetic |
"Model image generated. Garment measurements taken from the physical sample." |
| No disclosure, with a real-looking person |
Hits the deepfake criteria directly |
"Model image generated with AI." near the image |
Keep it in the market's language, keep it short, and keep it off the product's critical details. Done well, the disclosure costs nothing at the point of sale because it answers a question the buyer already has.
FAQ
Do AI-generated product photos need a label under the EU AI Act?
Not automatically. The AI tool provider must mark synthetic output in a machine-readable way, and you must add a visible disclosure only when the image is a deepfake. A generic generated backdrop that no buyer would mistake for a real place does not need a visible "AI" label under Article 50, though marketplace rules and accuracy rules still apply.
Is a generated background a deepfake?
Usually not. A deepfake needs all three: resemblance to an existing person, object, place, entity or event; something that exists or could plausibly exist; and a false appearance of authenticity. The Commission's guidance says background scenes and standard production effects are not likely to make content falsely appear authentic. A scene that recreates a specific real store, or adds a real-looking person, is where the test starts to bite.
Does Article 50 apply to suppliers outside the EU?
Yes, when the output is used in the EU. Providers outside the EU are covered if their output is used in the EU, and a supplier publishing that output in an EU-facing listing is a deployer. The market the image is shown in decides it, not your address.
What is the penalty for not labelling a deepfake?
Fines can reach EUR 15 million or 3% of worldwide annual turnover, whichever is higher, with proportionality for SMEs. National market surveillance authorities enforce it, so the practical risk starts with a complaint or a market check.
Do old catalog images need to be relabelled?
No, not for content generated before 2 August 2026; the Commission says it does not need retroactive labelling. A new generation or a new variant is new output, and accurate representation of the product still applies to old files.
Sources & References
