knowledge 8 June 2026 |

European Commission Publishes Draft Guidelines on Transparency Obligations under the EU AI Act

Rationale and Scope of Article 50

Article 50 of the AI Act includes four transparency obligations which apply to all AI systems (whether high-risk or otherwise) which are used in specific ways. Article 50 aims to reduce the risks of impersonation, deception, misinformation, manipulation and fraud when using an AI system.

• Article 50(1): Interactive AI Systems

Providers of AI systems intended to interact directly with natural persons must design those systems in such a way that individuals are informed they are interacting with an AI system.  Examples of AI systems which are intended to interact directly with natural persons include chatbots, voice assistants and bots on social media.

The Guidelines note that providers may use “any appropriate disclosure technique”, whether textual, auditory, visual or multimodal, provided that (i) the information is clear, distinguishable and delivered at the latest at the first interaction; and (ii) the characteristics of natural persons belonging to vulnerable groups due to their age or disability are taken into account to the extent that the AI system is intended to interact with those groups.  Equally, however, the Guidelines warn against overly intrusive disclosure techniques which may disrupt the user experience and lead to habituation effects. 

• Article 50(2): Marking and Detection of AI-Generated or Manipulated Content

Providers of AI systems that generate or manipulate synthetic audio, image, video or text content must ensure that their outputs are (i) marked in a machine-readable format and (ii) detectable as artificially generated or manipulated with technical solutions that are effective, interoperable, robust and reliable.  The Guidelines emphasise that both aspects of Article 50(2) must be satisfied. 

The Guidelines clarify that to comply with the marking element, providers may rely on a single marking technique or a combination of techniques.  Examples of technical solutions for marking outputs include watermarks, metadata identifications and cryptographic methods.  Providers may implement the marking solution at different stages of the value chain (e.g. at the level of the AI system or at the level of the underlying AI model) and rely on marking implemented by upstream model providers or third parties providing the solution (e.g. as an open standard or a specialised service).

Providers must also make detection tools available to enable exposed persons to identify AI-generated content.  Detection must be made available to people likely to encounter the content, interoperate with other solutions and produce clear, human-readable results at first exposure.  The Guidelines recognise that a provider may rely on a third-party or a publicly available detection solution so long as they ensure interoperability. 

The overall solution must be effective, interoperable, robust and reliable, assessed against the state-of-the-art and taking into account technical feasibility and costs.  The Guidelines acknowledge that while it is not currently possible, in the future the state-of-the-art may facilitate providers demonstrating compliance while meeting all four quality criteria via a single, technical solution.

• Article 50(3): Emotion Recognition and Biometric Categorisation

Deployers of emotion recognition systems and biometric categorisation systems must inform the natural persons exposed to those systems of their operation.  According to the Guidelines, this obligation applies both in real-time and where these systems are operated ex-post.  The method used to provide this information may depend on the context of deployment. 

• Article 50(4): Deep Fakes and AI-Generated Text Publications

Article 50(4) requires deployers of AI systems that generate or manipulate deep fakes (image, audio or video) to disclose that the content has been artificially generated or manipulated.  

Reduced disclosure obligations apply for evidently artistic, creative, satirical or fictional works, where disclosure must not hamper the work’s display or enjoyment.  In such circumstances, however, deployers must still ensure appropriate safeguards for third-party rights (e.g. intellectual property and data protection).  

A separate obligation applies to deployers of AI systems that generate or manipulate text published with the purpose of informing the public on matters of public interest.  This obligation, as detailed in the Guidelines, does not apply where the text has undergone substantive human review or editorial control, and a natural or legal person holds editorial responsibility for its publication.  Superficial or merely procedural checks do not satisfy this condition and the identity and contact details of the responsible person should be publicly available in an easily findable location. 

 

In relation to each of these transparency obligations, Article 50(5) mandates that all transparency information must be provided in a clear and distinguishable manner, at the latest at the time of the first interaction or exposure, and in conformity with applicable accessibility requirements.  This timing test applies to each natural person and each output. The Guidelines warn that disclosures hidden in manuals, buried under interface layers, or shown only once may be insufficient.  Sometimes, repeated disclosures will be needed (e.g. periodic reminders for interactive AI systems where users may form emotional attachments).  

What’s next?

While non-binding, the draft Guidelines aim to provide practical guidance to assist competent authorities, providers and deployers in ensuring consistent and effective application of Article 50. Once finalised, the Guidelines will complement the Code of Practice on marking and labelling of AI-generated content (which is also currently being finalised).

The Commission has indicated that it will review and, where necessary, amend these Guidelines in light of submissions by stakeholders made during the consultation process, practical experience, technological developments and any relevant CJEU decisions.