EU AI Act and AI-generated fashion imagery: a guide for brands

The EU AI Act's transparency rules have applied since 2 August 2026. For fashion teams using AI-generated on-model imagery, packshots, or campaign content, the practical question is now how to assess each use and build disclosure into the publishing workflow. Provider-side technical marking and brand-side disclosure are separate responsibilities. European Commission overview.

Updated 31 October 2026. This article provides general information for fashion marketing and production teams, not legal advice. Ask a qualified legal adviser to assess your specific imagery, roles, and distribution channels.

What has changed since the August 2026 deadline?

The Commission published its final Article 50 guidelines on 20 July 2026, and the AI Omnibus entered into force on 27 July 2026. The earlier draft guidance and provisional agreement should no longer be the starting point for your review. See the final guidelines and the Commission's Omnibus update.

• 2 August 2026: Article 50's transparency obligations began to apply, including the relevant duties to disclose deepfakes.
• 2 December 2026: providers of synthetic-content AI systems placed on the market before 2 August 2026 must meet Article 50(2)'s marking obligations by this date. This specific transition is now set out in Article 111(4).

The December transition is not a general extension for brands publishing AI imagery. It concerns the provider's marking obligation for qualifying existing systems; it does not postpone the deployer's disclosure duty.

Who is responsible: the AI provider or the fashion brand?

Article 50 separates two duties relevant to imagery:

• Providers: Article 50(2) requires machine-readable marking and detectability of synthetic outputs, subject to its scope and exceptions. These include systems performing an assistive role in standard editing or not substantially altering the supplied data or its meaning.
• Deployers: Article 50(4) requires disclosure when AI-generated or manipulated image, audio, or video content constitutes a deepfake.

A brand using an AI system under its authority will typically be a deployer. The exact role depends on the arrangement, including whether it has a system developed and puts it into service under its own name. The Article 3 definitions are the starting point.

For your production team, keep two separate checks: what technical marking the supplier provides, and what your audience needs to see when the imagery is published.

Does every AI-generated fashion image need a visible label?

Not automatically. The deployer's image-disclosure duty concerns content that meets the deepfake definition. Under Article 3(60), that involves AI-generated or manipulated media resembling existing people, objects, places, entities, or events and falsely appearing authentic or truthful.

The Commission's current guidance considers resemblance, whether the subject exists or could plausibly exist, and the false impression of authenticity together. Context, the message conveyed, and audience expectations matter. A fictional avatar is therefore not, by itself, enough to rule out the obligation.

For example, a realistic on-model image on a product detail page (PDP) should be assessed in the context of what shoppers might understand about the garment and how the image was made. A stylised campaign illustration may present a different case. These are assessment examples, not blanket classifications of either format.

Article 50 also provides a modified disclosure approach for deepfakes within evidently artistic, creative, satirical, fictional, or analogous works. That is a qualification of the duty, not a general exemption for all fashion campaigns.

What should disclosure look like in practice?

Where disclosure is required, people must receive it clearly and accessibly by their first exposure to the content. The Commission explains that they should not need a technical tool or a dedicated action to perceive it. Hidden metadata, a footer link, or a tooltip that must first be opened should not be treated as sufficient on its own. Commission guidance on disclosure.

A practical starting point is a visible label beside or on the relevant image, with wording that accurately explains whether it was generated or modified using AI. Review the actual placement on mobile, in image galleries, and when the content is shared or downloaded.

The EU has now released optional AI-content icons. Their use does not establish compliance by itself. The Commission's placement guidance also addresses overlays and resharing, so choosing an icon is only part of the implementation.

How the Code of Practice and C2PA fit in

The final Code of Practice on Transparency of AI-generated Content is now available and has been assessed as an adequate voluntary framework by the Commission and AI Board. It covers provider marking and detection, and deployer labelling. Organisations that sign it must implement its measures; other adequate routes to compliance remain possible. The underlying legal duties are not voluntary.

C2PA is an open technical standard for recording digital content's origin and editing history through Content Credentials. It can support a provenance workflow, but it is not synonymous with AI Act compliance. Ask your supplier which marking methods it uses and what happens when images are cropped, compressed, or exported.

As an operational precaution, retain provenance information and test the delivery chain. A supplier's technical marking should not be confused with the audience-facing disclosure your use may require.

A practical review checklist for fashion teams

Use this checklist with your production, e-commerce, and legal teams:

1. Inventory the uses. List AI-generated or modified PDP images, campaign visuals, lookbooks, and videos, together with their intended audiences and channels.
2. Record the assessment. Document the content, what it communicates, the audience's expectations, and the reasoning behind the disclosure decision.
3. Confirm responsibilities. Identify the provider and deployer in each workflow, including work involving agencies or other production partners.
4. Check the supplier's marking. Request documentation and test representative files through your normal editing and export process.
5. Review the published presentation. Check labels in desktop and mobile galleries, marketplace listings, and social formats. Confirm who handles each placement.
6. Keep a review record. Save the assessment, supplier information, approved label approach, and screenshots of the final presentation. Revisit the decision when the imagery or its context changes.

This is a suggested production workflow, not a guarantee of compliance. Product accuracy, rights to a person's likeness, and other applicable rules still need their own assessment.

Common questions

Does a designer's approval remove the image-disclosure duty?

No general exemption follows from human approval of an image. Article 50(4)'s human-review and editorial-control exception concerns certain public-interest text publications; it should not be transferred to deepfake imagery.

Can we wait until December to decide how to label images?

The transition in Article 111(4) is limited to Article 50(2) marking by providers of qualifying existing systems. It does not extend the brand's applicable disclosure deadline.

What are the potential penalties?

Article 99 sets a maximum for Article 50 infringements of €15 million or, for an undertaking, 3% of the preceding financial year's worldwide annual turnover, whichever is higher. The lower of the relevant percentage or fixed amount applies to SMEs, including start-ups, and to small mid-cap companies for these infringements. Actual penalties depend on the circumstances and applicable enforcement rules.

Watch the original Office Hours discussion

Graswald AI explored AI-generated fashion imagery and the EU AI Act in an Office Hours session moderated by Yannick Kwik, with Dr. Simon Hembt and Oliver Belitz from Bird & Bird. The recording provides context and discussion of the questions fashion teams raised.

This article has been updated against the official sources linked above. The recording reflects the guidance available when the session took place, so use the current sources when assessing dates and requirements.

Watch the Office Hours recording.