Do you have to label AI-generated fashion photos in the EU?
By The AURA Studio team. Published .
If you publish AI-generated or AI-manipulated images to an audience in the EU, Article 50 of the AI Act has applied since 2 August 2026 and the disclosure obligation falls on you as the deployer, not on the tool that made the image. Not every AI-touched photo needs a label. The test is whether the AI changes how your audience perceives the authenticity of what is being shown, and an AI model wearing your real garment sits squarely inside that question.
What changed on 2 August 2026
Article 50 of the EU AI Act sets transparency obligations for AI-generated content, and its application date was 2 August 2026. The part that matters to a shop is Article 50(4), which requires deployers to disclose when content is a deepfake, defined in the Act as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful.
The word deepfake does a lot of misleading work here. In ordinary use it means a fabricated video of a real person saying something they never said. In the Act it is a much wider category defined by the authenticity test, and a photorealistic image of a person wearing a garment is not obviously outside it. That is the whole difficulty, and any page that tells you the answer is simple in either direction is not reading the definition.
This guide is a plain-language summary written by a product team, not legal advice. It is accurate to the sources cited at the foot of the page as of the date shown. If AI imagery is central to how you sell, take advice on your own assets.
The obligation follows the publisher, not the tool
This is the single most useful thing to understand, and it is the thing most tool vendors would rather not say. Article 50(4) binds deployers: the businesses and people who take AI output and publish it to an audience. If you generate an on-model photo and post it, you are the deployer. Switching tools does not move the obligation, and neither does using a general model like ChatGPT or Gemini instead of a fashion-specific one.
It also applies on the basis of who sees the content rather than where you are incorporated, so a shop outside the EU whose posts and listings reach EU customers is inside the scope. If you have EU buyers, assume you are covered.
- You generate and post the image: you are the deployer.
- An agency generates it and you publish it: both of you may be deployers, so agree in writing who applies the label.
- The tool that generated the image has its own separate obligations as a provider, including machine-readable marking. Those do not discharge yours.
The test that decides each asset
Published guidance for the advertising supply chain converges on a workable question: does the AI change how the audience perceives the authenticity of the actual product, person or event being advertised? Routine production work generally does not. Rendering the thing being sold generally does.
| What the AI did | Where it tends to fall |
|---|---|
| Colour correction, background extension, rescaling a real photo | Minor impact on perceived authenticity. Generally no label needed. |
| Replacing or extending the scene around a real photographed garment | Usually treated as scene dressing rather than a claim about the product. |
| A fully AI-generated model wearing your real garment | The difficult case. A synthetic person presented as a real photograph goes to authenticity, and labelling is the cautious reading. |
| AI making the product look better, bigger, or different from what ships | Clearly inside the rule, and a consumer-protection problem quite apart from the AI Act. |
| Obviously fantastical imagery that nobody would read as a photograph | Outside the deepfake definition, which excludes clearly impossible content. |
For most boutiques the honest conclusion is that on-model AI imagery is the row where the cautious answer and the comfortable answer differ, and that the cautious answer costs almost nothing. A short line of text is cheaper than an argument.
What a compliant label looks like
The requirement is that disclosure is clear, distinguishable and understandable to the audience, and present at the latest at the first exposure to the content. Metadata alone is not enough, because a person scrolling a feed never sees it. There is no mandated wording or icon.
- Put it where the image is seenOn the image itself, in the first line of the caption, or in the alt text plus a visible caption line. Not in a policy page three clicks away.
- Say what was generatedWording like "AI-generated image. The garment shown is the item sold." is short, honest, and answers the customer's real question rather than only satisfying a rule.
- Keep the metadata tooMachine-readable marking is the tool provider's obligation, but leaving it intact costs nothing and supports your position if the question is ever asked.
- Write it into your processDecide once, per asset type, rather than per post. A rule like "every on-model generated image carries the line, real photography does not" is one decision instead of hundreds.
The part that is not about compliance
Shops resist labelling because they assume it costs conversion. The available evidence for that is weak, and the failure mode on the other side is well documented: a customer who discovers the model was generated after the fact treats it as a lie about the product, even when the garment was rendered faithfully. Disclosure converts the discovery from a betrayal into a fact you already told them.
There is also a claim worth making explicitly if it is true of your workflow: the model is generated, the garment is not. If your tool reproduces the piece from a photograph of the actual item rather than reimagining it, that is the sentence a customer cares about, and it is a stronger message than silence.
Questions
- Does the EU AI Act apply to my shop if I am not in the EU?
- The transparency rules follow the audience. If your posts, listings or ads reach people in the EU, assume the obligations apply to you regardless of where the business is registered.
- Does using a fashion-specific tool instead of ChatGPT change anything?
- No. The disclosure obligation under Article 50(4) sits with the deployer, meaning whoever publishes the content. The tool has its own separate provider obligations, including machine-readable marking, but those do not discharge yours.
- Is a photo of an AI model wearing my real garment a deepfake?
- It is the genuinely difficult case. The Act's definition turns on whether content resembling real persons or objects would falsely appear authentic. A synthetic person presented as a photograph goes to that question, so labelling is the cautious reading, and the cost of labelling is very low.
- What happens if I do not label?
- The transparency obligations are enforceable with significant penalties at the top of the range, which are aimed at serious and systematic breaches rather than at a boutique's Instagram post. The realistic risks for a small shop are platform-level enforcement and customer trust, both of which a one-line disclosure avoids.
- Do I need to label images where AI only changed the background?
- Generally no. Guidance treats scene dressing around a genuinely photographed product as having minor impact on perceived authenticity. The line is crossed when the AI renders the product itself, or the person wearing it, in a way that changes what the audience believes they are seeing.