Who Must Label and What Risks Are at Stake

On 1 August 2024, the EU Artificial Intelligence Act (AI Act/KI-VO) formally entered into force. However, a lengthy transitional period followed, as the Regulation provides for a phased application. The first two stages have already been reached. The third stage will follow on 2 August 2026, at which point the relevant parts of the Regulation will become directly applicable for the first time. The AI Act aims to regulate the use of AI systems in a reliable, transparent, and comprehensive manner. Structurally, the AI Act is based on a risk-based regulatory approach. Accordingly, AI systems are classified into different categories depending on their risk potential. The Regulation distinguishes in particular between prohibited AI practices, high-risk AI systems, AI systems with limited risk, and certain AI applications for which only voluntary codes of conduct are envisaged. The higher the risk associated with an AI system is assessed to be, the stricter the regulatory requirements. For what is currently the most relevant form of generative AI – which creates images, texts, and videos – a range of transparency obligations will apply from 2 August 2026 onwards. Among other things, operators of AI systems will be subject to labelling obligations with regard to AI-generated content. Particularly in connection with deepfakes, numerous legal and practical questions are currently arising, which will be examined in the following.

I. What Are Deepfakes?

The fundamental question is first what is to be understood by a “deepfake” within the meaning of the AI Act. The AI Act defines deepfakes as image, audio, or video content generated or manipulated by artificial intelligence that resembles existing persons, objects, places, or other situations and is thereby capable of creating the impression – particularly among technically unsophisticated persons – that the content is authentic or real. This captures, in particular, content that appears deceptively genuine and is therefore capable of concealing its artificial origin. At first glance, the definition seems comparatively clear: deepfakes are AI-generated content intended to appear real.

Upon closer examination, however, it quickly becomes apparent that the distinction between merely artificially generated content and content that actually triggers labelling obligations as a deepfake raises considerable uncertainties. In particular, it remains somewhat unclear how strictly the requirements of “similarity” to real persons and situations are to be understood. Additional uncertainty arises from the fact that it is already becoming apparent that the European Commission may interpret the concept of deepfakes rather broadly. This suggests that not only near-photorealistic content may be captured, but possibly also depictions that merely contain certain realistic elements or that could evoke an authentic impression in individual viewers. The boundary between clearly artificial fantasy depictions and labelling-obligatory deepfakes is thus increasingly blurred.

Furthermore, additional problems may arise in the practical classification of deepfakes. Whether a depiction qualifies as a deepfake is unlikely to be assessable in isolation based on image content alone. Rather, the specific context of use could play a decisive role. What could be particularly relevant is the purpose for which a depiction is used and what expectations are created among the relevant audience.

This becomes particularly apparent in the case of artistic or stylised depictions. For example, an AI-generated image of an oil painting with a portrait of a woman might, at first glance, not convey a realistic impression in the sense of an actual photograph. Compared to a photographic depiction, many viewers would recognise that it is an artificially generated or at least artistically altered representation. However, if such an image is used by a provider of oil paintings or art prints for advertising purposes, the perspective shifts. For technically and artistically unsophisticated viewers, the impression could well arise in this context that it is a depiction of an actually existing painting or a real work of art. The artificial generation of the depiction would then, from the perspective of the relevant public, potentially recede into the background. It is precisely this shift in perspective that illustrates the considerable legal uncertainty in interpreting the concept of deepfakes.

II. Who Can Be Subject to the New Obligations?

The determination of the nature and extent of transparency obligations will depend primarily on the distinction between the provider and the deployer (operator) of an AI system. Under the AI Act, a distinction must be drawn between providers of an AI system and deployers. A provider is assumed where the undertaking was actively involved in the development of an AI system or integrates an AI system into an already existing programme. By contrast, a deployer within the meaning of the AI Act is anyone who uses an AI system for their own purposes without the AI being further developed or offered as a separate product. As deployers, not only established companies are to be classified, but also sole traders, influencers, or small businesses can qualify as deployers with corresponding obligations. In a purely private context, however, no deployer status is assumed.

With regard to labelling obligations, it is primarily deployers who are affected: pursuant to Art. 50(4) AI Act (Art. 50 Abs. 4 KI-VO), they are obliged to disclose, in relation to deepfakes created by AI systems, that the content was artificially generated or manipulated.

III. Risk of Cease-and-Desist Actions: What Are the Consequences of Non-Compliance with Labelling Obligations?

While the AI Act itself already provides for severe sanctions for infringements of labelling obligations – under Art. 99(4)(g) AI Act (Art. 99 Abs. 4 lit. g KI-VO), fines of up to EUR 15 million or up to 3 % of total worldwide annual turnover may be imposed – it is particularly the competition-law risk that is likely to be of considerable practical relevance for deployers.

Of central importance here is the competition-law classification of the labelling obligations under Art. 50 AI Act. There is much to suggest that these provisions qualify as market conduct rules within the meaning of § 3a UWG (German Act Against Unfair Competition). Under § 3a UWG, anyone who contravenes a statutory provision that is intended, at least also, to regulate market conduct in the interest of market participants acts unfairly, provided the contravention is capable of appreciably impairing the interests of consumers, other market participants, or competitors. What matters, therefore, is not only whether a norm has a regulatory character, but also whether it simultaneously serves a competition-related protective function. This is precisely what is likely to be the case with the labelling obligations under Art. 50 AI Act.

The provisions evidently pursue the objective of ensuring transparency in the market and protecting consumers from being deceived about the origin or authenticity of content. Users should be able to recognise whether they are interacting with AI-generated or manipulated content or whether it is authentic human communication or genuine image, audio, or video content. Art. 50 AI Act thus does not merely aim at the abstract regulation of AI systems, but specifically at steering the market conduct of undertakings in competition. The labelling obligations directly influence the manner in which undertakings may present their products, services, and communication measures on the market.

In addition, the rules also serve the protection of fair competitive conditions. Undertakings that properly label AI-generated content could be at a competitive disadvantage vis-à-vis market participants who deliberately forgo labelling in order to make content appear more authentic, credible, or emotionally impactful. Particularly in the areas of advertising, social media communication, and audiovisual content, the failure to disclose AI-generated content can confer significant competitive advantages.

Against this backdrop, it is reasonable to regard Art. 50 AI Act as a provision that not only protects individual interests but also aims to ensure a level playing field. If Art. 50 AI Act is accordingly classified as a market conduct rule, this would have significant practical consequences. Infringements of the labelling obligations could then simultaneously constitute unfair commercial practices under § 3a UWG and thus be enforced under private law. Competition-law claims can be asserted not only by supervisory authorities and consumer protection associations, but in particular also by trade associations and competitors. It is especially the latter that carries considerable escalation potential, since infringements of labelling obligations are comparatively easy to identify and thus attractive targets for competition-law enforcement measures. Conceivable measures include, in particular, injunctive relief, cease-and-desist letters (Abmahnungen), and applications for interim injunctions.

The Wettbewerbszentrale (Centre for Protection Against Unfair Competition) stated in its guidance published in February 2026 that infringements of the AI Act’s labelling obligations can simultaneously constitute unfair commercial practices within the meaning of the UWG. At the same time, it draws attention to the specific risk that competitors and associations will assert injunctive relief claims where the use of unlabelled or insufficiently labelled AI content impairs fair competition. Against this background, there is a considerable likelihood that, in particular, unlabelled deepfakes will increasingly become the subject of competition-law cease-and-desist actions and court proceedings in the future.