Since 2 August 2026, the transparency rules in Article 50 of the EU AI Act have applied across the Union. Slovenia has now passed its own implementing law naming the authorities that will supervise them, which is what put the subject back in the news. If you sell into the EU and your product images are made or edited with AI, the obvious question is whether every photo now needs a label.
The short answer: probably not the way you fear. But the rules are worth ten minutes of your time, because the part that does apply is easy to get right and the fines are not small.
There are two obligations, and only one of them is yours
The AI Act splits the duty between the company that builds the tool and the company that uses it.
Providers — the people who make the generator — must ensure that outputs are marked in a machine-readable format and detectable as artificially generated or manipulated (Article 50(2)). That is an invisible, technical mark carried in the file. It is the tool vendor’s job, not the seller’s. Ask your image tools whether they do it; you cannot do it for them.
Deployers — that is you, the business publishing the image — must disclose that the content has been artificially generated or manipulated when it is a deepfake (Article 50(4)). Everything then turns on that word.
What the Act actually means by deepfake
The definition is broader than the popular use of the word. A deepfake 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.
Note objects. The definition was not written with catalogue photography in mind, but product images are images of existing objects, and a generated scene is one that never happened in front of a camera. That is why this deserves a careful read rather than a shrug.
The exemption that covers most product editing
Article 50(2) carries a carve-out that matters enormously for e-commerce. The marking obligation does not apply
to the extent the AI systems perform an assistive function for standard editing or do not substantially alter the input data.
Removing a background, straightening, cropping, resizing, correcting white balance, cleaning dust off a jar — these are standard editing. Retouching has been part of commercial photography since long before generative models, and the Act does not treat a background eraser as a deepfake engine.
Generating a full scene around a product is a different matter. The product is still real, but the room, the light and the table are not. That is closer to substantial alteration, and it is where a considered approach pays off.
Where it is genuinely unsettled
We would rather say this plainly than sell you false certainty: whether an AI-generated product scene counts as a deepfake has not been tested. A reasonable lawyer can argue it either way. The image resembles an existing object and does look authentic — but so does every studio photograph ever taken with a rented sofa and a rented lamp, and nobody labels those.
Two things tilt the practical answer. First, the Act is aimed at content that misleads people about reality, not at commercial presentation that has always been staged. Second, if the content forms part of an evidently artistic, creative, satirical or fictional work, the obligation shrinks to disclosure in a manner that does not hamper the display or enjoyment of the work.
What to actually do
- Do not strip metadata on export. If your generator writes a provenance mark, keep it. Many resizing and compression pipelines quietly remove it, which undoes the provider’s compliance and yours.
- Know which images are generated. Keep the distinction in your asset records. If a regulator or a marketplace ever asks, the answer should take a minute, not a week.
- Show the product truthfully. Colour, size, finish, what is included in the box. This is not new: EU consumer protection law has forbidden misleading images for years and it is enforced far more often than the AI Act will be.
- Check your marketplace rules. Amazon, Shopify and the large European platforms each have their own policies on AI imagery, and they move faster than legislation.
- Label when in doubt on people. If a generated model, face or body is involved, disclosure is the sensible default. That is the heartland of the rule.
The numbers
Breaching Article 50 falls under Article 99(4): fines up to €15 million or 3% of total worldwide annual turnover, whichever is higher. For SMEs and startups the ceiling is the lower of the two, which is a deliberate softening for smaller sellers.
The dates worth keeping: the transparency rules have applied since 2 August 2026; systems already on the market have until 2 December 2026 to adapt; the high-risk regime follows in December 2027 and August 2028.
What this does not change
Good product photography was already held to a standard stricter than the AI Act: the picture has to match what arrives in the box. Returns, reviews and chargebacks punish a misleading image long before a regulator does. If your generated images show your real product in a plausible setting, the new rules ask very little of you. If they show a product that does not exist, you had a problem before August 2026 as well.
This article is general information, not legal advice. The AI Act is new, national implementation differs, and if your use is unusual — generated people, medical or safety claims, regulated goods — talk to a lawyer who knows your market.
