Who enforces the EU AI image labelling rule in your store?

Since 2 August 2026, Art. 50(4) of the EU AI Act (Regulation (EU) 2024/1689) applies to stores that show AI images that look like real photos. The next question: who checks, and what happens when a label is missing? The short answer for Germany, as of 14 September 2026: the competent authority is the Federal Network Agency (Bundesnetzagentur), the practical risk is a warning letter (Abmahnung) from a competitor or an association, and no documented case on product images exists so far. That is neither cause for alarm nor an all-clear. The guide on what Article 50 requires covers the obligation itself.

Who enforces the labelling obligation in Germany?

Germany set up enforcement in the AI Market Surveillance and Innovation Promotion Act (KI-MIG, signed 22 July 2026, BGBl. 2026 I No. 223). Under Section 2(1) KI-MIG, the Federal Network Agency (Bundesnetzagentur) is the market surveillance authority for the AI Act; under Section 8 KI-MIG it is also the central complaints body for competitors and consumers.

The fine ranges are set in Art. 99(3) to (5) AI Act; Section 17(1) KI-MIG makes the market surveillance authority the fining authority. Two points:

• For SMEs and start-ups, Art. 99(6) AI Act applies the lower of the two amounts in the range.
• Warnings and non-monetary measures as milder means, and regard for the economic survival of SMEs, were already part of Art. 99(1) AI Act in its 2024 version. The Digital Omnibus Regulation (EU) 2026/1744, in force since 27 July 2026, recast the paragraph: fines are now named explicitly, for “any infringement”, and the regard now also covers small mid-cap companies, for which the new Art. 99(6a) likewise applies the lower amount. Which authority is competent in which EU Member State, and which cases are known, is in the enforcement tracker.

The Federal Network Agency supervises the store’s deployer obligation under Art. 50(4) (visible disclosure). The machine-readable marking under Art. 50(2) is owed by the providers of the image generators, not by the store; providers of systems on the market before 2 August 2026 have until 2 December 2026 (Art. 111(4) AI Act as amended by Regulation (EU) 2026/1744). Deployers have no such transition period.

Why is a warning letter the more realistic risk than a fine?

Art. 50(6) AI Act leaves other transparency obligations unaffected. In Germany, that opens the route through the Act Against Unfair Competition (UWG), where competitors and associations, not authorities, drive enforcement. Two independent hooks:

The second hook predates the AI Act: an AI image that shows a product differently from what is delivered (color, details, features that do not exist) is misleading, with or without a label. A label does not cure a misleading image, and a lawful image can still need a label under Art. 50.

The Centre for Protection against Unfair Competition (Wettbewerbszentrale) writes in its AI labelling guide (version 2, 29 July 2026, accessed 31 August 2026) that AI Act violations can, in its view, also constitute unfair competition, so competitors and associations can bring cease-and-desist claims. It reads the rule broadly: an abstract resemblance is enough if the image as a whole looks realistic, and it recommends labelling AI images when in doubt until the courts settle the question.

Collective enforcement is open: Art. 110 AI Act adds the regulation to the annex of the Representative Actions Directive (EU) 2020/1828, so injunctive actions by qualified entities are possible; redress actions remain untested.

Can a competitor report me without sending a warning letter?

Yes. On 28 July 2026, the Wettbewerbszentrale set up an online form for AI complaints. The form lists three subject categories, including “insufficient labelling of AI content/deepfakes”; the other two cover misleading statements by an AI and advertising with AI features that do not exist. A competitor can thus report a store at no cost, without sending a warning letter of its own. The Wettbewerbszentrale publishes no case numbers.

Which cases are known so far?

Review of the Wettbewerbszentrale news page and public sources, as of 14 September 2026:

Claims of existing rulings or a wave of warning letters over AI product images have no verifiable source. Yet the complaints channel has existed since late July 2026, and one of the bodies that sends warning letters in Germany reads the rule broadly: the risk is real but has not materialized yet.

For images created before 2 August 2026, the legal text has no grandfathering clause. The European Commission’s FAQ says content created before the cutoff date need not be labelled retroactively but expressly recommends that deployers label it. The German law firm IT-Recht Kanzlei reads it the same way (article of 31 July 2026): no retroactive labelling for older images, unless re-published. That is an authority position and a law firm opinion, not a legal provision; if you rely on it, be able to show when an image was created.

What should you document as a merchant?

Para. 148 of the Commission guidelines C(2026) 5054 final of 20 July 2026 says that deployers who have not signed the code of practice must demonstrate to the market surveillance authorities how their measures implement Art. 50. What a comparison with Section 2 of the Code of Practice looks like for a store is covered in The EU Code of Practice on AI labelling. Three layers of documentation:

1. Decisions per image or image group

• Which images are AI-generated or AI-edited, with which tool, and when?
• Which case group the image falls into (fully generated product image, AI person, lifestyle scene, real photo on an AI background, plain standard editing).
• How and why you decided; for borderline cases, a short sentence recording the reasoning helps.
• With agencies: the contract should state whether AI may be used, which images are AI-generated, and who applies the label. Under para. 14 of the guidelines, the deployer is whoever decides whether and how AI is used; a company that only orders a result without making that decision is not, and employees working under instructions are not separate deployers. More: Creating AI images for store clients.

2. Evidence on the original file

Many generators write a provenance marker into the file on save (C2PA/Content Credentials, IPTC DigitalSourceType). A marker found is evidence of AI origin; a missing marker is evidence of nothing. In a measurement series of 1 September 2026 on the Easy AI Act Image Labels test bench, covering 47 image models from one generator catalog, 13 wrote no marker, and screenshots or right-click saving strip every marker. Check the original file: in a measurement of 10 September 2026, copies served by the Shopify CDN carried no marker for 3 of 3 positive controls although the originals had one; the original remains available under Content > Files in the Shopify admin.

3. Visibility of the label

Art. 50(5) AI Act requires clear and distinguishable information, at the latest at first exposure. Paras. 142 and 143 of the guidelines specify: per image, to every viewer, and not where it is easily overlooked (expressly named: deep menu levels and terms of use; a footer is ruled out by the same logic). Under para. 117, deployers may not rely on the provider’s metadata, which is not visible to humans. For your documentation:

• A dated record that the label is visible on the image, for example a product page screenshot.
• A fresh check after every theme change, because theme app embeds are turned on per theme.
• The alt text as a supplement for assistive technologies, not as a replacement for the visible label.

An app such as Easy AI Act Image Labels reads provenance markers on the original file and records your decisions per image. It cannot tell you whether an image without a marker is an AI image; only your own knowledge of what you created with AI closes that gap.

How high is the risk of a warning letter, really?

The assessment this guide is based on (Germany, as of 1 September 2026):

Whether costs and a contractual penalty are actually owed depends on the sender: from a competitor they may be excluded under Section 13(4) No. 1 and Section 13a(2) UWG, from an association they are not; whether this applies to missing AI labels has not been decided. Details and what to do right away: Warning letter over an unlabelled AI image.

The competitor, not the authority, is the party you are more likely to face. And the obligation has applied only since 2 August 2026; the Commission guidelines are expressly non-binding (para. 5) and will be updated (para. 155).

Austria and Switzerland differ: Austria has not yet designated its competent authority, and the Austrian Protection Association Against Unfair Competition (Schutzverband gegen unlauteren Wettbewerb) refers to the German Wettbewerbszentrale’s guide (Austria in detail). In Switzerland, according to the Federal Council’s legal baseline analysis of 12 February 2025, national law generally has no labelling obligation; the AI Act can apply where the images are used in the EU, for example when selling to EU customers, and under the Commission guidelines (para. 13) possibly already with a website accessible worldwide (Switzerland guide, as of 1 October 2026).

For you as a merchant: label photorealistic AI images when in doubt, because the label costs little and a dispute costs a lot. Replace images that show the product differently from what is delivered instead of labelling them. And for both, record what you decided and when.