Home / Blog / AI Labels and Retouching
Retouch Atelier · 2026-09-23

Does retouched product photography need an AI label?

No — not ordinary retouching. The duty that applies to a brand or a studio is Article 50(4) of the EU AI Act, and it asks whether an image would falsely appear to a person to be authentic or truthful. Colour correction and cut-outs do not. But the line between that and generated content is narrower than most production teams assume, and since 2 August 2026 crossing it carries a legal consequence rather than an aesthetic one.

This is where we draw that line in our own studio, and why.

What changed on 2 August 2026

Article 50 of the EU AI Act sets transparency obligations for generative AI, and it applies from 2 August 2026. It splits the duty in two.

Providers of generative AI systems must mark their outputs in a machine-readable format, so the content can be detected as artificially generated or manipulated. Deployers — the businesses actually using those systems — must disclose content that is artificially generated or manipulated and constitutes a deepfake.

If you are a brand or a studio putting images on a product page, you are a deployer. The marking duty is not yours. The labelling duty is.

Enforcement sits with national market surveillance authorities, and with the AI Office for systems under its supervision.

The word that pulls product photography into this

Most people read "deepfake" and picture a fabricated video of a person. The definition is wider than that. It covers image, audio or video content which resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.

Objects. A product is an object.

Which means an image with no human being anywhere in the frame — a shoe on white, a bottle, a folded shirt — sits inside the definition if it is artificially generated or manipulated and would pass as a real photograph of the real thing. That is the sentence most product teams have not read yet.

The exception everyone quotes is the wrong one

The wording people reach for is that the duty does not apply where an AI system performs “an assistive function for standard editing” or does not “substantially alter the input data provided by the deployer or the semantics thereof”. That sentence is real, but it sits in Article 50(2) — the provider’s machine-marking duty. It is not the duty that lands on a brand or a studio, so it is not the reason the answer is no.

The duty that lands on you is Article 50(4), and it turns on a different question: would the image falsely appear to a person to be authentic or truthful? That is the test the answer at the top of this page rests on. Colour correction does not make a jacket appear to be a thing it is not. Cropping does not. Removing lint does not. The Commission makes the same point about technical pre- and post-processing in ordinary production work — it is not likely to make content falsely appear authentic.

But "standard editing" is a phrase, not a list. Someone has to decide which side of it a given operation falls on, and in practice that someone is the production team. So here is how we decide.

The line, drawn by a retoucher

Standard editing — we treat this as out of scope

Every operation on that list changes how the photograph looks. None of them changes what the photograph claims about the object.

Generated — mark it and label it

The grey zone, which is where studios actually get caught

Nobody sets out to fabricate a product image. The exposure arrives quietly. A studio turns on generative fill to save time on background extension, runs it across forty thousand images, and the status of a whole catalogue has changed without anyone making a decision.

The test we use

One question resolves almost every grey case.

Does the image still tell the truth about what arrives in the box?

That is close to what the Act is reaching for with "would falsely appear to be authentic or truthful". A background extended into empty white space does not change what arrives in the box. A fabric texture the model invented does. A colourway that was never made does.

This test is not a legal opinion and it will not survive every edge case. It is the rule we can actually apply at volume, on a shift, at two in the morning, by a retoucher who has thirty seconds to decide.

Where we stand

Retouch Atelier does not generate imagery. Every image is retouched by hand by a named retoucher, against a written standard, and every frame passes in-house QC before it is delivered. Client files are never used to train a model.

We do retouch frames that clients generated themselves. That is their disclosure decision to make, not ours, and we would rather say so plainly than pretend the distinction does not exist. We have written separately about what still has to be done by hand on generated model images.

We are not lawyers and this is not legal advice. It is how a studio shipping 40,000 images a month has decided to operate, published so that clients can check it against their own counsel's view.

Four things worth doing this month

  1. Ask your retouching supplier, in writing, whether generative fill or AI upscaling is enabled anywhere in their pipeline. Most brands have never asked. Some suppliers have it on by default.
  2. Write down which operations your team treats as standard editing. The exception only helps you if you can show where you drew the line and that you held it.
  3. Check whether your image metadata survives your CDN. If provenance information is stripped on upload, you cannot rely on it later.
  4. Decide who owns the decision. In most companies right now it belongs to nobody, which is how it ends up being made by whoever happened to tick a box in an editing tool.

Common questions

Does background removal need an AI label?

We treat background removal as standard editing. It changes the setting, not the object. The product in the frame is still the product that ships.

Does ghost mannequin retouching need an AI label?

Not when the neck join is built from a second real frame of the same garment, which is how we build it. A neck interior invented by a model rather than photographed is a different thing.

We use AI-generated models but photograph the real garment. Does that need a label?

The model is generated and the image resembles a person who does not exist, which is squarely the case the disclosure duty was written for. The garment being real does not remove the model from the frame.

Does this apply if our studio is outside Europe?

It applies to content placed on the European market. Where the studio sits is not the test; where the customers are is.

Is the Code of Practice compulsory?

No. The Code of Practice on Transparency of AI-generated Content, published on 10 June 2026, is voluntary, and around 190 organisations had signed it by the end of July 2026. The Article 50 obligations themselves are not voluntary. Signing the Code is one way to demonstrate compliance; not signing means demonstrating it another way.

Sources: the European Commission's pages on the Code of Practice on Transparency of AI-generated Content the Guidelines on transparency obligations, and the text of Regulation (EU) 2024/1689 itself. Article 50 has applied since 2 August 2026; Regulation (EU) 2026/1744 of 8 July 2026 deferred the separate high-risk regime, and did not move it. Retouch Atelier is a retouching studio, not a law firm. Check anything here that carries consequence with your own counsel.

Try before you buy

Send 5 images. Get them retouched free.

Photographed or generated — send the frames and the garment reference, and judge the match.

No payment, no commitment. We reply within one business day.