# EU AI Act Article 50: do AI model images in fashion e-commerce need a label?

Article 50 of the EU AI Act applies from 2 August 2026. What it means for generated model images on product pages and in ads, and who labels what.

- URL: https://humblebob.com/journal/eu-ai-act-article-50-fashion-ecommerce
- Publisher: HUMBLEBOB (https://humblebob.com)
- Published: 2026-09-29
- Updated: 2026-09-29

---

Most likely, yes. Since 2 August 2026, Article 50(4) of the EU AI Act requires anyone who uses an AI system to create a "deep fake" in a professional context to disclose that the content is artificially generated or manipulated. A photorealistic generated model wearing your garment on a product page fits that definition as the European Commission's guidelines describe it, even though no official text uses a fashion example yet. Routine edits such as colour correction or an aesthetic background replacement are likely outside it. An image that makes the product look better than it is falls inside it. The disclosure has to be visible to the person looking at the image. Metadata in the file does not count on its own.

[Video: The short version: which images need a visible label, why metadata in the file does not count, and why the contract should say who labels.](https://humblebob-media.humblebob.workers.dev/humblebob-website/content/video/sj-ai-act-6ec70f7dc5.mp4) (Motion: HUMBLEBOB)

> Does a generated model need a label? A generated model wearing the product: yes, a visible one. Colour correction or a new background: likely not. Metadata in the file does not count as the label. Agree who labels, in the contract. Source: EU AI Act Article 50 and the Commission's guidelines, July 2026.
This reflects the rules and guidance as of 28 September 2026. It is not legal advice. For your own set-up, check with a lawyer.

## What Article 50 says

Article 50 of [Regulation (EU) 2024/1689](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202401689) has two duties that matter for fashion imagery, and they sit with different parties.

- **Article 50(2), for providers** of systems that generate images: outputs must be "marked in a machine-readable format and detectable as artificially generated or manipulated". This is the job of whoever builds or offers the generating system.
- **Article 50(4), for deployers**, meaning the people and companies using such a system professionally: deployers of a system that generates image content "constituting a deep fake, shall disclose that the content has been artificially generated or manipulated".

Article 3(60) defines a deep fake as AI-generated or manipulated content "that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful". Article 50(5) adds how the disclosure must be made: "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure".

## When Article 50 applies, and the one grace period

- **2 August 2026.** Article 50 applies.
- **2 December 2026.** The Digital Omnibus, [Regulation (EU) 2026/1744](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng), gives providers of generative systems that were already on the market before 2 August extra time for the machine-readable marking in 50(2). There is no extra time for the deployer's disclosure duty in 50(4).
- **Images made before 2 August 2026** do not need to be labelled retroactively, according to the [Commission's guidelines](https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems). For images, the generation date is what counts. The Commission encourages labelling older images anyway, without disproportionate effort.

![Timeline of six dates in 2026, from the Code of Practice in June to Zalando's metadata requirement in December.](https://humblebob-media.humblebob.workers.dev/humblebob-website/content/aiact-label-dates-1440-e8b36fd680.webp)

*The 2026 dates behind labelling fashion images, from the Code of Practice in June to Zalando's rule on AI data in files. (Diagram: HUMBLEBOB)*
## Is a generated model image a deep fake?

The Commission adopted [guidelines on Article 50](https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems) on 20 July 2026. They are not binding, but they are the best indication of how the rule will be applied. Three points from them matter for on-model images:

![Full-length product photo of a woman in a sleeveless linen shirt dress with a tie belt, on a plain studio backdrop.](https://humblebob-media.humblebob.workers.dev/humblebob-website/content/aiact-plausible-person-1024-abdc920614.webp)

*A generated model in an ordinary product shot: under the guidelines, a person who can plausibly exist counts as a person, even when no such person does. (AI-generated image)*
1. **A generated person counts as a person.** "Persons" include "realistic AI-generated human avatars or personas". The subject does not have to be a real, identifiable individual. It is enough that it "can plausibly exist".
2. **Generated essential elements tip the balance.** Using AI to generate "essential elements that impact audience perception", with "fully AI-generated actors" as the example, is likely to make content appear authentic. Photorealism makes it more likely, but is not decisive on its own.
3. **Intent does not matter.** The test "does not require the intention of the deployer to deceive".

The guidelines contain no fashion or on-model example. The conclusion that a photorealistic generated model on a product page is a deep fake is therefore an inference, but a strong one. The nearest thing to a fashion example comes from Germany's Wettbewerbszentrale, a self-regulatory body for fair competition. [Its guide](https://www.wettbewerbszentrale.de/wp-content/uploads/2026/02/2026_2_Leitfaden_KI_generierte_inhalte_1-1.pdf) says advertising with AI models should disclose the AI origin once the regulation applies, and that a realistic AI image used to advertise swimwear on social media carries a considerable risk of being taken as real.

## Which product image edits need a label?

The guidelines give product advertising examples on both sides of the line.

![Decision chain: an image that improves the product needs fixing, a generated person needs a label, routine edits likely not.](https://humblebob-media.humblebob.workers.dev/humblebob-website/content/aiact-label-decision-1440-126ebffa97.webp)

*The line the guidelines draw for product images, as two questions to ask of every generated or AI-edited image. (Diagram: HUMBLEBOB)*
| Likely not a deep fake | A deep fake |
|---|---|
| AI colour correction, background extensions, aesthetic background replacements, compositions of existing products, re-scaling | An AI-generated product image that makes the product "appear not identical to the real product, more appealing or with improved quality than in real life" |
| A real product shown against an AI-generated background, as long as the ad does not mislead about the product | An AI-generated depiction of a celebrity influencer in an advertising context |

For fashion that means a real packshot or a real on-model photo, cleaned up or placed on a generated background, is likely fine without a label. A generated model wearing the product, or a garment that looks better than the one in the box, is where the duty starts.

## Does the creative exception cover advertising?

Article 50(4) has a lighter regime for "evidently artistic, creative, satirical, fictional or analogous" works. The guidelines read "evidently" strictly. Content that is exclusively informative or commercial and recognisable as such is excluded. Advertisements "might be regarded as evidently creative or fictional works in certain, specific situations, but not in others", and when content is both informative and creative, "the informative character should always prevail". Works "mainly motivated by functional or technical considerations cannot be regarded as creative". A product page image is functional, so do not count on the exception there. A clearly fictional campaign film is a different case, and even there the exception only softens the duty: the AI content still has to be disclosed, in a way that does not spoil the work.

## Who has to label: the brand, the agency or the vendor?

The disclosure duty sits with the **deployer**: the person or company "using an AI system under its authority". The guidelines explain that authority means deciding to use the system and how, including its outputs. Employees and contractors working under your control do not become separate deployers. Two passages matter for the chain between brand, agency and image vendor:

- "A company that merely commissions an advertising agency to produce an advertisement, without taking decisions and exercising control over whether and how the advertising agency uses AI in the production process, is not a deployer."
- One company can be provider and deployer at the same time. Platforms that only host other people's content are not deployers, so a marketplace showing your images does not take the duty off your hands.

In practice: if your brand decides that generated models are used and approves the images, you are most likely a deployer. If you only buy finished images and the agency decides the method, the agency probably is. Put it in the contract either way. The guidelines themselves suggest securing correct display at first exposure through contractual conditions with distributing partners.

## How to label

The disclosure has to be clear and distinguishable at the first exposure, for every output and every person who sees it. The guidelines say it fails when it can "be easily overlooked or missed", for example in a manual, in menus, or in terms of use.

- **Metadata is not enough.** Deployers "cannot rely on the machine-readable marking" embedded by the provider, because it is not visible to people. Content Credentials are still worth having for other reasons, covered in [our article on C2PA](https://humblebob.com/journal/c2pa-content-credentials-product-images).
- **The Code of Practice sets a concrete format.** The [Code of Practice on Transparency of AI-generated Content](https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content), final since 10 June 2026 and signed by around 190 companies, with Bulgari and Getty Images among those signing as deployers, asks signatories to use an icon with "AI" as its main element, embedded in the image or overlaid on it, for example in the top right corner. The Commission publishes [free EU icons](https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content) for this. Using them is optional; labelling is not.
- **National guidance points the same way.** The Dutch advertising code body [accepts a disclosure "in or directly next to the image"](https://www.reclamecode.nl/nieuws-agenda/nieuws/nieuwe-europese-regels-voor-ai-content-in-reclame/). The Wettbewerbszentrale advises German-language labels for the German market.

For your own product pages, the simplest pattern that follows the Code is a small "AI" icon on each generated image in the gallery and in the zoom view.

## Marketplaces add their own rules

Marketplaces set rules on top of the law, and they can be stricter. Zalando's [image guidelines](https://partner.zalando.com/university/article/zalando-image-guidelines), updated 31 August 2026, do not allow visible labels on images at all, "e.g. text overlays or watermarks". They also ask partners not to provide content that requires visible labelling under applicable laws, and they name deep fakes explicitly. From December 2026, Zalando also requires identifying data embedded in the metadata or as an invisible watermark in AI-generated files.

![A blazer on a ghost-mannequin form in front of a paper backdrop, lit by two softboxes, with a camera on a tripod in front.](https://humblebob-media.humblebob.workers.dev/humblebob-website/content/aiact-packshot-studio-1536-32dfb915dd.webp)

*A real garment photographed on a form: where a marketplace allows no visible labels, images that need no label are the ones you can publish there. (AI-generated image)*
Read together, a generated model image that counts as a deep fake may not be usable on Zalando at all, whatever label your own site uses. Yet the same guidelines set quality rules for AI-generated model images, so Zalando does accept some, and it does not say which ones it thinks need a label. Ask in writing before you plan a season around generated models there. Check every channel's current rules before you plan a production around generated models, and agree in writing with your agency how disclosure is handled per channel.

## A label does not fix a misleading image

The guidelines are explicit that Article 50 does not make otherwise unlawful content lawful, and they name misleading advertising as an example. The Dutch advertising code body says the same: a mention like "generated with AI" does not automatically make an ad acceptable. If a generated image shows a garment in a better colour, a better fit or a better finish than the real product, the label does not help. That is a consumer protection problem as well as an AI Act problem.

![Hands lifting a linen shirt dress out of an opened parcel, beside a printed photo of a model wearing the same dress.](https://humblebob-media.humblebob.workers.dev/humblebob-website/content/aiact-parcel-check-1536-086c9c2e9b.webp)

*The dress from the parcel next to its product image: if the image shows a better colour, fit or finish than the real garment, a label does not make it acceptable. (AI-generated image)*
## Enforcement and fines

Breaches of Article 50 can be fined up to €15 million or 3% of worldwide annual turnover, whichever is higher, under Article 99. For small and medium-sized companies, and since the Omnibus also small mid-caps, the lower of the two applies. Enforcement is national. In Germany the Bundesnetzagentur has been the [market surveillance authority](https://www.gesetze-im-internet.de/ki-mig/BJNR0DF0B0026.html) since 29 July 2026. In the Netherlands the Autoriteit Persoonsgegevens is [preparing for its supervisory role](https://www.autoriteitpersoonsgegevens.nl/actueel/transparantie-eisen-ai-gelden-vanaf-2-augustus-ap-adviseert-praktijkcode-te-ondertekenen) and advises companies to read and sign the Code of Practice. For a brand selling in Germany, a competitor may be the quicker risk: the Wettbewerbszentrale notes that an AI Act breach can also count as unfair competition, which competitors and trade associations can stop with an injunction.

## A checklist for fashion brands

1. **List where generated or AI-edited images appear**: your own site, marketplaces, ads, social, newsletters.
2. **Sort them** into routine edits (likely no label) and generated people or improved products (label).
3. **Decide who is the deployer** for each production route, and write it into agency and vendor contracts.
4. **Choose one visible label** and put it on the image, at first exposure, on every channel. Where a channel does not allow visible labels, do not publish images that need one there.
5. **Check each marketplace's rules**, because they differ, and some do not accept images that need a visible label at all.
6. **Keep the images honest.** A label does not rescue a garment that looks better on screen than in the parcel.
7. **Keep machine-readable provenance** in your master files as well. It does not replace the label, but platforms and marketplaces read it.

Ask your image vendor which provenance data its files carry (see [Content Credentials](https://humblebob.com/journal/c2pa-content-credentials-product-images)), and agree with it who labels what, per channel, before production starts. For the wider picture of how brands have handled generated models in public, see [AI fashion campaigns so far](https://humblebob.com/journal/ai-fashion-campaigns-what-brands-learned).

---

About the publisher: HUMBLEBOB is a production partner for fashion imagery. We read your brand, sort your briefing, make the assets, deliver them ready for your channels and put what we learn into the next drop.