Labelling AI images in your online store under the EU AI Act

Where the label has to go

Under Article 50(5) of the EU AI Act, the information must be provided at the latest at the time of first exposure. The guidelines apply this to every person, not just the first, and to each output at any moment someone is reasonably likely to see it (para. 143). In an online store, that is rarely the product page first; more often it is a product card in a collection, a search result, a recommendation or a banner on the home page.

The sizes are our recommendation; the AI Act and the Code of Practice set no pixel values. The Code asks for a label that stays clear and distinguishable at any size (Measure 1.1 c).

There is no exception for scaled-down images. The guidelines only exclude icon-scale outputs from the marking duty under paragraph 2 (para. 68), meaning tiny graphics generated by AI, not a small rendering of a deep fake.

Hover images and variants

If a product card shows a second image on hover, or the variant changes, the label has to follow the image currently on screen. If only the second image is AI-generated, it needs its own label as soon as it appears.

Checkout, order confirmation and emails

Many e-commerce platforms do not allow overlays in checkout or in order emails. Anyone who reaches checkout has almost always seen the image with its label before; in our view, that covers first exposure. For direct links into checkout, such as “buy again” buttons in emails, an “AI” burned into the image file reduces the residual risk (see AI label design).

What does not count

• A notice image in the gallery. A first gallery slide saying “Some images are AI-generated” labels none of the AI images, does not appear on product cards, in search or in zoom, and can be skipped (para. 143).
• A line in the product description. It sits beside or below the image, often below the fold, and is missing from cards altogether. Yet the information has to stand apart from other information and must not be easy to overlook (para. 142).
• Terms, FAQ, legal notice or an information page. The guidelines name terms of use, manuals and hidden menus as unsuitable (para. 142). An explanatory page can complement the label.
• A blanket “we use AI”. It does not say which image is affected. For chatbots, the guidelines call such general statements insufficient (para. 38), and the reasoning carries over to images.
• Metadata in the file. Nobody perceives machine-readable marking without tools (para. 117).
• A label that only appears on hover. The “AI” has to be visible without any interaction (Code, Sub-measure 1.2.1 a). Only additional detail may sit behind a click.

Channels outside your store

An overlay only works inside your own store. Wherever you reuse a deep fake yourself, you label it there too. In multi-step distribution chains, deployers should take proportionate steps, such as contractual terms and interface settings, so that their label reaches the audience (para. 12).

• Google Shopping and Merchant Center: Google requires the IPTC value “trainedAlgorithmicMedia” in AI image files. Since July 2026 it expressly allows your own AI labels in the image but advises against the outermost corners, which may be cropped. The general image specification still lists overlays as a reason for disapproval. More in Google Merchant Center.
• Meta (Facebook, Instagram): Ads only get an automatic AI notice when Meta's own AI tools were used. Images from other tools are yours to label, for example with a burned-in version.
• TikTok: Under its ad policy, AI-generated ad content needs the AIGC label or a clear disclaimer of your own. TikTok's exemptions for edits do not replace the Article 50 assessment.
• Amazon: Images with photorealistic, fully AI-generated people should carry the keyword “contains-synthetic-performer” in their metadata.
• Newsletters, marketplaces and your own social posts: Where no overlay is possible, use a version with the label burned in.

Within a very large online platform, deployers may rely on the platform's labelling tools if they allow a clear and distinguishable disclosure, but only there (para. 126). Machine-readable entries are explained in AI metadata.

Supplier images, stock photos and agency work

Your role for someone else's image depends on who decided to use AI.

If you edit someone else's image with AI, you are the deployer for that edit. Never remove an existing label or provenance data; the guidelines encourage keeping them (paras. 16 and 98). Under consumer protection law you are responsible for your advertising, borrowed images included. When in doubt, ask your supplier.

Documentation: process, record, corrections

Signatories of the Code of Practice set up internal processes and documentation describing how they label, in proportion to their size and resources (Commitment 2, Measure 2.1). Non-signatories are expected to show by other means how they meet the obligation, for example through a gap analysis against the Code, and may face more detailed requests for information (para. 148).

1. A short procedure. It sets out who decides on labelling, which AI tools the team uses and who checks new images before publication. Those who create AI images regularly should also check that labels display correctly (Measure 2.1).
2. A record per image. The image, when and with which tool it was made or changed, the decision with reasons and source, label on or off, and who decided. For images demonstrably created before 2 August 2026, a retroactive label is not required under the guidelines (para. 154); the record provides the evidence.
3. A correction route. Review reports of missing or wrong labels and fix them without undue delay (Code, Measure 2.3). The Code encourages a reporting channel for customers but does not require one.

Staff and contractors involved in labelling should know about the obligation (Code, Measure 2.2). More in Code of Practice.

What a label does not replace

Labelling meets the transparency duty under Article 50. Recital 137 of the AI Act states that compliance with the transparency obligations should not be read as meaning that the use of the content is lawful. The guidelines name misleading advertising as an example of deep fakes that remain unlawful despite a label (para. 129).

If an AI image shows a product in a colour, size, quality or effect it does not have, that can be misleading under consumer protection law, with or without a label.

Real people need to consent

If you use AI to change an image of a real person, such as a model's face, figure or skin, data protection and personality rights apply (paras. 127, 129); as a rule you need that person's consent. Under the Code, labelling expressly does not exempt you from it (Section 2, Commitment 1). Check your model contracts before editing shoots with AI.

Typical cases

• AI model wearing the product
• Real product in front of an AI background
• Background removal with AI
• Colour variant generated with AI

All cases are collected under Cases. For Shopify there is an app that adds the label across all store surfaces and records each decision.