EU AI Act and AI-Generated Fashion Images: Visible Labels and Provenance

AI fashion imagery is moving from experiments into production: product pages, lookbooks, ads, virtual try-on, marketplace feeds, and campaign refreshes. That changes the compliance question from "can we make this image?" to "can we prove what this image is?"

In the EU, the important upcoming date is August 2, 2026. The AI Act is already in force, and the Article 50 transparency rules for AI-generated content become applicable on that date. The European Commission's final July 2026 guidance confirms two separate duties: providers must technically mark in-scope outputs, while professional deployers must give people a clear, perceivable disclosure when they publish content that constitutes a deepfake.

What Article 50 Actually Requires

The official AI Act text is Regulation (EU) 2024/1689. For fashion teams, the relevant section is Article 50, "Transparency obligations for providers and deployers of certain AI systems."

The Commission published the final Article 50 guidelines in July 2026 and the final Code of Practice on Transparency of AI-Generated Content in June 2026. The code is voluntary, but Article 50 is not. Providers and deployers that do not sign the code must demonstrate compliance through other adequate means.

Why Fashion Brands Should Care

The AI Act definition of "deep fake" is broader than face swaps or impersonations of known people. It covers 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.

The final guidelines say "existing" includes subjects that exist, can plausibly exist, or could have plausibly existed. They also say a person includes a realistic AI-generated human avatar or persona. A synthetic model does not have to resemble a known real individual. A photorealistic AI model wearing a garment in an ordinary ecommerce or advertising context is therefore likely to qualify when viewers could believe that a real person participated in a real photoshoot. The assessment remains case-specific and considers the image, its message, placement, foreseeable audience, and audience expectations.

Commercial fashion imagery should not assume the lighter disclosure rule for evidently artistic, creative, satirical, or fictional works. The guidelines exclude content whose nature is recognisably commercial from that lighter regime in many cases, and specifically list realistic synthetic humans demonstrating or advertising products as examples that do not qualify. Even an eligible creative work still requires an appropriate disclosure.

Not every AI edit needs a visible deepfake label. The guidelines say a real product shown against an AI-generated background can fall outside the deepfake definition when the ad is not likely to mislead people about the product's appearance, characteristics, or use. Insignificant edits such as ordinary colour correction, lighting adjustments, background cleanup, resizing, and compression may also fall outside. A fully synthetic photorealistic model is materially different because it can create the false impression that a real person took part in a real shoot.

Machine-Readable Marking Is Not the Same as Visible Disclosure

Fashion teams should separate two jobs:

When an image-generating company is both provider and professional publisher, both duties can apply to it. When a fashion brand uses the system under its authority and publishes an in-scope deepfake, the provider's technical mark does not discharge the brand's visible disclosure duty. Roles and control over the AI system should be assessed for each workflow.

What a Visible Disclosure Should Look Like

• -Show it at first exposure. A shopper should perceive the disclosure when first encountering the image, without opening a menu, reading terms, or using a verifier.
• -An interface label can work. The law does not require words to be permanently burned into the pixels. The final code permits an equivalent interface overlay when it appears on the content and meets the placement rules.
• -Plan for exported assets. The code's placement approach aims for the disclosure to accompany content through distribution. The Commission's icon guidance says the label must remain visible when content is reshared or downloaded.
• -Use plain language. "AI-generated model image" is a practical label for synthetic fashion models because it identifies what is artificial without suggesting that the garment design is fictional.

The Commission's AI-generated and AI-modified icons are optional. A company may use its own icon or wording, but the disclosure still has to meet Article 50's clear, distinguishable, timely, and accessible standard.

What a Good AI Image Provenance Stack Looks Like

A practical fashion AI workflow should not rely on a filename, folder name, or internal Slack message to prove an asset was generated. Those are easy to lose. A real provenance stack should survive normal production use and provide a public way to verify the file.

The final Code of Practice uses a multilayer provider approach for signatories: digitally signed metadata and imperceptible watermarking, supported by detection mechanisms. Fingerprinting and logging are optional supplementary measures. The code separately sets design and placement rules for deployer labels.

What Uwear Does

Uwear's image provenance layer is designed around Uwear-generated final assets. Instead of depending only on whatever marker an upstream image model may or may not provide, Uwear applies its own provenance signals at the delivery layer. Uwear also gives teams separate controls for invisible provenance and a visible AI disclosure, so both Article 50 jobs can be handled in the same workflow.

Anyone can use the public checker at uwear.ai/verify-ai-image. The endpoint accepts an uploaded image and returns one of four statuses: `verified`, `not_verified`, `tampered`, or `unsupported`.

The verifier checks the invisible provider-side layer. For an EU publishing workflow, teams can also turn on Uwear's visible label for any output classified as a deepfake. The two controls are complementary: keep the invisible provenance enabled for technical marking and detection, then enable and retain the visible disclosure whenever Article 50 requires people to see it.

curl -F "file=@image.png;type=image/png" \
  https://api.uwear.ai/public/provenance/verify

A Simple Checklist for Fashion Brands

The limited grace period until December 2, 2026 applies only to the provider-side Article 50(2) marking and detection duty for systems placed on the market before August 2, 2026. It does not postpone the deployer-side visible deepfake disclosure duty. Content generated before August 2, 2026 does not have to be labelled retroactively, although the Commission encourages voluntary labelling where possible.

This is not legal advice. The final guidelines require a case-specific assessment, so providers and brands should confirm their roles, content classification, and disclosure design with EU counsel.