Transparency obligations under Art. 50 AI Act for AI systems

From 2 August 2026, transparency and marking obligations under Art. 50 of the AI Act will apply to providers and deployers of certain AI systems. While not every AI-generated output will have to be labelled in the future, these rules nonetheless create obligations for a large number of companies that use AI applications to generate content (e.g. in marketing) or to directly interact with individuals (e.g. chatbots in customer service). Such companies will be required to label specific AI systems and AI-generated content.

1. Purpose and scope of transparency obligations under Art. 50 AI Act

Chatbots that respond to customer queries. Videos featuring a familiar voice that appears to convey a message it never actually uttered. Images that look deceptively real yet are entirely machine-generated. The use of certain AI systems is associated with risks of deception, identity fraud or, more generally, the lack of distinguishability between real and artificially generated content and the resulting risk of confusion for individuals.

Art. 50 AI Act addresses these risks by imposing transparency and marking obligations for certain AI systems, irrespective of their risk classification otherwise. In principle, the following AI systems are covered:

• AI systems intended for direct interaction with natural persons (e.g. chatbots, social bots), Art. 50(1) AI Act,
• AI systems (including general-purpose AI systems) that generate synthetic content (audio, image, video, text), Art. 50(2) AI Act,
• AI systems designed to detect or infer emotions (emotion recognition systems) or to categorise individuals based on biometric data (biometric categorisation systems), Art. 50(3) AI Act, and
• AI systems that generate or manipulate image, audio or video content constituting a deep fake, Art. 50(4) AI Act.

2. Providers and deployers of AI systems

The transparency and marking obligations under Art. 50 AI Act apply to both providers (Art. 50(1), (2) AI Act) and deployers (Art. 50(3), (4) AI Act) of certain AI systems. Providers are, for example, companies that develop (or have developed) an AI system and place it on the market or put it into service under their own name or trade mark. By contrast, companies that use an AI system on their own responsibility for business purposes by organising and systematically controlling its operation will typically qualify as deployers.

Companies regularly use commercially acquired or licensed AI systems that they have not developed themselves. In such cases, they will usually qualify as deployers. However, the role of a company as an actor under the AI Act must always be assessed on a case-by-case basis. Extensive (pre-)configuration or fine-tuning of a generative AI model may suffice for a company to qualify as a provider under the AI Act with respect to the AI system built on that model.

3. Information in cases of direct interaction with AI systems (Art. 50(1) AI Act)

Art. 50(1) AI Act imposes obligations on providers of AI systems that are intended for direct interaction with individuals. Direct interaction means the immediate effect of, or communication by, the AI system with a human being. Where they qualify as AI systems, this covers, for instance, chatbots, social bots primarily used in social networks, or voice assistants and voice-controlled systems. Pure machine-to-machine interaction (M2M), where no human is involved, does not fall within this category.

Where direct interaction is present, the provider must design the AI system in such a way that the affected individuals can be informed that they are communicating or interacting with an AI system.

Art. 50(1) AI Act provides for exceptions where the use of an AI system is obvious from the per-spective of an average addressee or in the field of law enforcement, provided the system is not made available to the public for the purpose of reporting a criminal offence.

4. Marking of synthetically generated AI content (Art. 50(2) AI Act)

Art. 50(2) AI Act likewise addresses providers of AI systems, including general-purpose AI systems. Where an AI system generates synthetic content, the provider must ensure that the output of the AI system is marked in a machine-readable format and is recognisable as artificially generated or manipulated. The technical solutions must, inter alia, be effective, interoperable, robust and reliable.

Synthetic content may include text, image, audio or video content. This particularly covers generative AI systems but can also encompass chatbots or AI-based search engines. The AI Act provides for exceptions to the marking obligation, for instance where the AI systems perform a merely auxiliary function in standard processing or do not substantially alter the input data or their semantics as provided by the deployer. This may be the case for AI-based image editing tools or text checking applications.

5. Transparency in emotion recognition and biometric categorisation (Art. 50(3) AI Act)

Under Art. 50(3) AI Act, deployers of an emotion recognition system or a biometric categorisation system must inform the individuals affected by the operation of such systems. An exception applies where these AI systems are lawfully authorised for law enforcement purposes.

6. Disclosure of deep fakes (Art. 50(4) AI Act)

In particular in the context of using AI in marketing and advertising, the obligation for deployers to disclose artificially generated or manipulated deep fakes under Art. 50(4) AI Act will become relevant. A deep fake is understood as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. Distinguishing deep fakes from other AI-generated content is often complex.

For “resemblance” to be present, it is sufficient in a broad understanding that the AI-generated content exhibits an abstract (photorealistic) similarity to fictitious persons, places or objects, irrespective of whether they actually exist or are known to the targeted audience of the content. Decisive is the perception of the content as falsely genuine or truthful. Thus, AI-generated depictions of well-known personalities, audiovisually realistic AI-generated persons, or real persons in fictitious, AI-generated yet authentic-looking environments may all constitute deep fakes. This also applies to AI-generated product prototypes where they convey a realistic appearance.

Clearly unrealistic depictions do not fall within the concept of “deep fake”, for example where they contravene laws of nature or basic physical or biological principles. This may include, for instance, talking objects in an advertising video. Likewise, ordinary AI-supported standard processing of content (e.g. changes to light, colour, contrast, audio parameters or minor details in the background) is not covered, provided such alterations are of minor relevance for assessing the authenticity of the content.

The obligation under Art. 50(4) AI Act encompasses the disclosure that the deep fake content is AI-generated. Where the content consists of text, it only has to be marked as AI-generated if it is published in order to inform the public about matters of public interest. For works that are clearly recognisable as artistic, satirical or fictional, a marking that does not interfere with the effect of the work will suffice. Further exceptions apply where the AI content is lawfully used for law enforcement purposes and where AI-generated content is subject to genuine editorial control and a responsible natural or legal person assumes responsibility for the publication.

7. Practical implementation of transparency and marking obligations

Under Art. 50(5) AI Act, the decisive point in time for implementing the marking obligation is the first interaction with the AI system: at the latest when the AI system encounters the individual, the transparency and marking obligations must be fulfilled. Information must be provided in an ac-cessible manner and in a clear and unambiguous way; it may not be hidden, barely legible or placed at a remote location on a website.

Possible implementation measures (depending on the specific AI system used) include, for example:

• Textually highlighted notices or banners upon first interaction with the AI system or on a permanent basis during use, for instance in close proximity to the input and output fields (e.g. “You are interacting with an AI system”).
• Visual notices or symbols: permanently visible icons or AI markings or watermarks.
• In voice-based deployment contexts (e.g. telephone customer service): an explicit spoken notice at the beginning of the call (e.g. “You are speaking with an AI assistant”).
• Multimodal combinations of textual, visual and auditory notice elements.

Not sufficient to meet the marking obligation are notices that appear solely in terms and conditions, URLs or other documentation, or machine-readable information that is not perceivable for users of the AI system at the time of interaction with the system, such as metadata. A mere reference to the underlying technology (e.g. “This system uses a large language model”) likewise does not ensure adequate transparency.

8. Sanction risks in case of non-compliance

Non-compliance with the transparency and marking obligations under Art. 50 AI Act may trigger severe sanctions under the AI Act: administrative fines of up to EUR 15 million or, in the case of companies, up to 3 percent of the total worldwide annual turnover of the preceding financial year may be imposed. From the perspective of the German Wettbewerbszentrale, infringements of the AI Act may also be classified as unfair commercial practices under the Act against Unfair Competition (UWG), enabling competitors and associations to assert injunctive relief claims.

9. Support in implementing individual obligations under Art. 50 AI Act

Whether companies will have to label their AI-generated content under Art. 50 AI Act in the future and, if so, which requirements they must comply with, will primarily depend on the role they assume under the AI Act in the specific case and on which AI systems they use for which purposes. To support the interpretation and application of these obligations, the European Commis-sion has already issued guidelines on transparency obligations for providers and deployers of AI systems as well as a draft code of conduct on transparency of AI-generated content under Art. 50 AI Act. Although both documents provide valuable guidance, they do not relieve companies from assessing their specific obligations on a case-by-case basis. We provide comprehensive support in the legally compliant design and implementation of AI projects and in meeting national and European regulatory requirements in AI law.