AI Labels on Product Imagery: What the EU AI Act Now Requires, and How adidas Is Already Responding

AI-generated imagery shown to EU customers now needs a clear label under the EU AI Act, and adidas is already doing it. How fashion brands can comply on Shopify with a single metafield.

News Jason West18 August 20266 min read

Key takeaways

From 2 August 2026, AI-generated or AI-manipulated imagery shown to EU customers must carry a clear label at first exposure under Article 50 of the EU AI Act, with penalties up to 15 million euros.
The duty sits with the brand publishing the content, and it captures UK and GCC brands whenever their imagery reaches EU users.
adidas has already started labelling AI content on its website, an early signal that visible disclosure is becoming the premium retail norm.
On Shopify, a single Yes/No metafield and a small theme block are enough to render compliant labels across product pages, no app or replatform needed.

Browse adidas.com this month and you may notice something new: content carrying a small label marking it as AI-generated. It is easy to miss, and that is the point. The label is quiet, factual and entirely deliberate.

It is also a preview of where every fashion and footwear brand selling into Europe is heading. Since 2 August 2026, Article 50 of the EU AI Act has required AI-generated and AI-manipulated content shown to EU users to be clearly disclosed.

The question for brand owners is no longer whether AI imagery needs labelling. It is how to do it without rebuilding the website, and what it signals to customers when you do it well.

What does the EU AI Act require from 2 August 2026?

Article 50 is the transparency chapter of the EU AI Act, and its obligations took effect on 2 August 2026. For imagery, it creates two duties that matter to ecommerce teams.

First, visible disclosure. AI-generated or AI-manipulated content that could pass as authentic must be clearly labelled, and the disclosure has to reach the viewer no later than their first exposure to the content. A note buried in the footer does not meet that bar.

Second, machine-readable marking. The content itself should carry embedded signals that identify it as AI-generated, so platforms and detection tools can read its provenance automatically.

The European Commission has published an official set of labelling icons, released in June 2026 alongside its Code of Practice on Transparency of AI-Generated Content, giving brands a recognised visual standard rather than leaving every marketing team to invent its own. Penalties for non-compliance run up to 15 million euros, which moves this from a nice-to-have to a board-level item.

Does this apply to your brand?

Almost certainly, if AI touches your imagery and your imagery reaches EU customers. The rules bind deployers as well as providers: a brand publishing AI-generated campaign or product imagery is a deployer, regardless of which tool generated the asset.

Geography does not offer an exit. A UK menswear brand shipping DTC to France, or a GCC-based label with EU traffic, is in scope the moment the content reaches users in the EU. The UK has no equivalent domestic rule yet, but for most UK brands the EU reach makes that distinction academic.

Two boundaries are worth knowing. Content published before 2 August 2026 does not need retroactive labels. And providers of generative AI systems already on the market have a transitional window until 2 December 2026 for the technical marking layer. The brand-facing duty to disclose visibly, however, is live now.

What counts as AI-generated imagery?

Fully generated images are the obvious case: an AI model wearing your product, a generated backdrop, a campaign visual that never involved a camera. All of that needs a label when it appears authentic.

The harder territory is manipulation. The Act targets content that has been meaningfully altered so that it appears genuine when it is not. Generating a model who does not exist, swapping a shot location, or building virtual try-on imagery sits inside the rule. Routine colour correction and standard retouching of a real photograph does not.

For anything in between, a useful test: would a reasonable customer assume this was photographed? If yes, and AI did the work, label it. Over-disclosure carries almost no cost. Under-disclosure now carries a legal one.

adidas is already labelling, and that matters

adidas has begun surfacing AI labels on content across its website, one of the first major sportswear names to make the disclosure visible in the customer journey rather than burying it in legal copy.

When a brand of that size normalises the label, the competitive frame flips. Unlabelled AI imagery stops being the industry default and starts looking like the outlier. Premium and heritage brands, whose value rests on authenticity, have the most to lose from being caught undisclosed and the most to gain from disclosing cleanly.

How do you build AI labels on Shopify?

This is where the compliance conversation gets mercifully simple. Shopify's native metafields can carry the disclosure logic without an app, a replatform or a development project.

The core pattern is one field: a true/false metafield such as custom.is_image_ai, set to Yes on any product or image where AI generated or meaningfully altered the visual. A small theme block then checks the field and conditionally renders the label, either the Commission's official icon or a plain "AI-generated" badge, over the PDP gallery or alongside the image.

In practice the rollout looks like this. Audit your imagery and agree internally what counts as AI-generated under the definitions above. Create the metafield definition, with storefront access enabled so the theme can actually read it, a step that is easy to miss and silently breaks the display when skipped. Add the conditional block to your product media snippet. Then set the flag per product or per image as part of your normal upload workflow, so new assets are classified at the point they enter the store.

Brands with more complex needs can step up to a metaobject: a small "AI disclosure" object holding the label type, the generation tool, and per-market display rules, referenced from products. That suits larger catalogues mixing photographed, retouched and generated assets, and it future-proofs the setup as disclosure standards evolve. This is the same native-first approach we take across client builds: use Shopify's own data structures before reaching for an app.

The machine-readable layer

The visible badge is half the obligation. The Act also expects AI content to be detectable by machines, which in practice means provenance metadata such as C2PA content credentials embedded in the file itself.

Most major generation tools now embed these credentials on export. The practical risk is stripping them accidentally: re-exporting through an editor, compressing through a third-party optimiser, or flattening files in a way that discards metadata. Keep the original credentialed file in your asset pipeline and upload that version to Shopify.

The Commission's voluntary Code of Practice offers a recognised route to demonstrating compliance on this layer, and aligning with it early is considerably cheaper than retrofitting provenance across a live catalogue later.

Labels as a trust signal, not a tax

It is tempting to treat all this as regulatory friction. The stronger read is that disclosure is becoming brand language.

Fashion customers are increasingly alert to imagery that looks synthetic, and suspicion is corrosive precisely because it is silent: shoppers do not complain about an uncanny campaign image, they just trust the product page a little less. A clear, confident label resolves the doubt before it forms, and it lets brands draw a valuable line: our hero product photography is real, and where we use AI, we say so.

That positioning compounds with where discovery is heading. As we covered in our guide to getting products found in ChatGPT and AI shopping agents, AI systems increasingly read structured signals from your store, and clean provenance data is exactly the kind of signal they reward. Well-structured disclosure serves the regulator, the customer and the recommendation engine at once.

Where this goes next

The December 2026 deadline for the technical marking layer will pull every major generation tool into line, and the Commission's icons will start appearing across European retail the way cookie notices once did. Within a season or two, the label will read as normal.

The brands that move now get to define what their disclosure looks like while it is still a choice rather than an enforcement response. A single metafield, a small theme block and a clear internal policy on what gets flagged: that is the whole build. The brands that wait will make the same change later, under deadline pressure, with a regulator watching.