Do AI-generated fashion images need a label? The AI Act transparency rules

Not every AI-assisted fashion image needs a visible label, but many do. Under Article 50 of the AI Act, applicable since 2 August 2026, brands must tell customers when they are talking to a chatbot and must disclose deepfakes, meaning AI-generated or manipulated images, video or audio that resemble real people, places or events and could appear authentic. A photorealistic AI model wearing a real product on a product page is a typical case to assess.

What does Article 50 require, and from whom?

Article 50 splits duties between providers, who build or sell AI systems, and deployers, who use them. A fashion brand running a campaign with an image generator is normally a deployer; the tool vendor is the provider. A brand that builds its own customer chatbot may be both.

The information must be clear and distinguishable, given at the latest at the first interaction or exposure, and must meet accessibility requirements. Breaches of Article 50 fall under the fine tier of up to EUR 15 million or 3% of worldwide turnover.

When is an AI fashion image a deepfake?

The Act defines a deepfake by two features: the content resembles existing persons, objects, places, entities or events, and it would falsely appear to a person to be authentic or truthful. That definition was written with politics in mind, but it fits several common fashion practices:

• Digital twins of real models placed in scenes or outfits they never wore.
• Photorealistic synthetic models on product detail pages that a shopper would take for a photograph.
• Virtual try-on or background swaps that place a real product in an invented real-looking location.
• Synthetic voice or video of a recognisable designer or ambassador in social content.

Less likely to qualify: clearly illustrative or stylised artwork, retouching that does not substantially alter the image, and visibly fantastical campaign imagery. Article 50 exempts providers' marking duties where a system only performs assistive standard editing or does not substantially alter the input. For evidently artistic, creative, satirical or fictional work, the deployer's disclosure is limited to noting that generated content exists, in a way that does not spoil the work.

Read also

How are fashion brands using AI in marketing, from content to measurement?

What is the Code of Practice on AI-generated content?

The AI Office facilitated a Code of Practice on transparency of AI-generated content, drawn up by independent experts. According to the Commission, the final code was scheduled for publication on 10 June 2026, around 190 organisations had signed by the end of July 2026, and the Commission and AI Board confirmed it as an adequate voluntary tool for demonstrating compliance. The Commission also published draft guidelines on the scope of Article 50 on 8 May 2026. Signing is voluntary; the Article 50 obligations are not.

How should a fashion brand label AI content in practice?

Because the law sets outcomes rather than wording, brands need a house rule. A workable approach combines a policy, a workflow and a vendor check:

1. Classify assets at creation. Tag every image, video and text in the DAM as human, AI-assisted or AI-generated, and record the tool used.
2. Apply a deepfake test. Ask whether a reasonable shopper could take the asset for a real photograph of a real person or place. If yes, disclose.
3. Use consistent wording and placement. A short caption such as "AI-generated image" near the asset, in alt text and in social captions, avoids hidden disclosures.
4. Keep provider marking intact. Do not strip metadata or watermarks when exporting; check that the CDN and image compression preserve them.
5. Disclose chatbots on first contact. Name the assistant as AI in the opening message, not only in the privacy policy.
6. Secure consent and rights. Digital twins need contractual consent from the person depicted; the AI Act does not replace image rights or GDPR.

Read also

AI-generated models in fashion: what are the ethics and disclosure rules?

What are the limits and risks?

Machine-readable marking is still maturing, and the Act itself qualifies the duty as "as far as technically feasible". Metadata can be lost in editing, and visible labels can be cropped by marketplaces or resellers. Over-labelling carries a commercial cost too: a disclaimer on every lightly retouched image dilutes the signal. The defensible position is a documented policy that applies the deepfake test consistently, plus vendor contracts that confirm marking capabilities.

Agencies and freelancers add a further layer. Much fashion imagery is produced by external studios, retouchers and content agencies using their own tools. The brand that publishes the content remains the party customers see, so briefs and contracts should require suppliers to declare which assets were generated or substantially altered with AI, to keep provider marking intact and to hand over the information needed to apply the house labelling rule. The same applies to wholesale partners who receive product imagery for their own webshops: if an image needs a disclosure on the brand site, the retailer needs to know that too, ideally through a field in the product data feed rather than a separate email.

Transparency rules also sit alongside consumer law. Misleading a shopper about how a garment fits or looks remains an unfair commercial practice regardless of how the image was made, so product imagery that is AI-generated should still represent the real item faithfully.