Does Virtual Try-On Need an AI-Generated Label? The EU AI Act, Explained

Most fashion stores file 'AI disclosure' under chatbots and voice assistants and move on. Article 50 covers a wider category than that. It applies to any synthetic image, audio, or video content that could pass for something real, and a virtual try-on result is built to do exactly that. The whole point of the technology is to look like an actual photo of an actual person in an actual garment.

What changed under the EU AI Act on August 2, 2026?

The EU AI Act was adopted in 2024, but its provisions came into force in stages over several years. Article 50, the transparency and disclosure section, became applicable on August 2, 2026. From that date, the law splits responsibility for AI-generated content between two different parties: the company that builds the AI system, and the business that uses it to reach its own customers.

That split matters for a fashion store more than it might first appear, because a Shopify store running virtual try-on is not the company that built the underlying AI model. It is the business showing that model's output to shoppers, which puts it on the receiving end of a separate legal duty, not just a beneficiary of someone else's compliance work.

Who has to label AI content: the provider or the store?

• Providers of the AI system: the company that builds and operates the tool, Corlen in this case, must mark its outputs in a machine-readable format so software can detect that the content was artificially generated.
• Deployers, meaning the business using the AI system: must separately disclose to the person viewing the content that it was artificially generated or manipulated, and do it in a way a shopper actually notices rather than a line buried in a policy page. For a fashion store running virtual try-on on its own product pages, the store is the deployer.
• Where the shopper is located, not where the company is registered: guidance on the rule describes the obligation applying whenever an AI system's output reaches a person in the EU, so a US-based Shopify store selling to European shoppers is in scope the same as an EU-based one.

This is a two-party structure by design. A store cannot assume that because its try-on vendor complies on the vendor's side, the store itself has nothing left to do. The deployer obligation sits with whoever puts the image in front of the shopper.

Does a virtual try-on image count as a deepfake under this rule?

Article 50 does not require a label on every image that used AI at some point. The dividing line is whether the output could pass as a real photograph of a real person, object, or scene, which is what the law means by a deepfake in this context.

An object-only edit, background removal, color correction, a garment cutout on a plain backdrop, generally does not trigger the deployer disclosure requirement on its own. A photorealistic image showing what looks like a real person wearing a garment is a different case. Coverage of the rule states plainly that even a fully synthetic model, one generated with no real photo behind it at all, can still trigger disclosure if the result looks like a plausible real person. A virtual try-on result, which starts from an actual shopper's own photo and renders a garment onto it, sits even more clearly inside that category than a synthetic model does.

None of this changes what Corlen actually generates. The render is still an approximation of fit and color, not a claim that the shopper already owns the garment. What changes is whose job it is to say so out loud, and how visibly.

What does a compliant AI-generated label actually look like?

Article 50 sets a standard rather than handing out a fixed template. Disclosure has to be clear, distinguishable from ordinary content, and delivered by the time a person first sees the image, not several clicks away in a terms page nobody opens.

• Visible at first exposure: the disclosure needs to sit on or right next to the image itself, not buried in a privacy policy the shopper never reads.
• Readable, not disguised: a faint watermark that only shows up under close inspection does not meet a clear and distinguishable bar.
• Not solved by the provider's metadata alone: even though Corlen and other AI providers embed a machine-readable marker in their outputs, guidance on the rule is explicit that a deployer cannot rely on that marker by itself to satisfy its own separate disclosure duty to shoppers.
• A ready-made reference point now exists: the European Commission published a free set of icons for labeling AI-generated content in June 2026, part of a broader code of practice on marking AI content, so a store does not need to design a label from scratch.

How does Corlen handle AI-generated disclosure today?

Corlen already treats disclosure as part of the product rather than something bolted on for one region's law. On the in-store kiosk, every result screen carries a visible AI GENERATED PREVIEW label directly on the image itself. Corlen's terms state plainly that every preview is an AI-generated approximation, not a guarantee of exact fit, color, or fabric appearance.

That covers Corlen's own surfaces. It does not automatically cover a merchant's product page, ad, or email built around a Corlen-generated image, since the deployer obligation under Article 50 sits with the business showing that image to its own customers, not with Corlen as the tool provider. This article is general information, not legal advice. A store selling into the EU should confirm its own labeling approach with qualified counsel rather than assume a vendor's badge alone covers its own duty.

On a Shopify store, virtual try-on runs directly on the product page through Corlen's Shopify app, with the same visible disclosure already built into every result. A team building on WordPress or a custom stack can use the same generation and disclosure behavior through the developer API instead.

For the related question of what actually happens to a shopper's photo, rather than how the finished result gets labeled, see is virtual try-on safe.

See it directly: try Corlen on your own photo, no account needed for a first result.