European Commission publishes draft Guidelines on Article 50 IA Transparency Obligations of the RIA
The European Commission published on 8 May 2026 draft Guidelines to clarify and harmonise the application of the transparency obligations of Article 50 of Regulation (EU) 2024/1689 ("RIA"). The document is intended to serve as a practical guide for providers and deployers of IA systems to apply these obligations in a consistent manner across the European Union, in line with Article 50 of the RIA itself.
Taking an operational approach, the draft delimits the scope of each obligation and is aligned with the Commission's forthcoming Code of Best Practice on Transparency in AI-generated content. This code will not be mandatory, but aims to provide a common framework for complying with the requirements of Article 50, in particular with regard to the marking and labelling of AI-generated content.
The Guidelines recall that these obligations will apply from 2 August 2026 and that they mainly impact those AI systems that interact with people, generate or manipulate content, employ emotion recognition or biometric categorisation, or produce deepfakes and certain public interest texts.
Key points:
- Clear and accessible transparency (Art. 50.1 RIA): in systems that interact with people, the provider must design the interface to inform, from the first interaction, that it is an AI. The notice must be visible and understandable on the relevant channel (screen, voice or signage), accessible to people with disabilities and maintained throughout the experience. It is not sufficient to hide the information in manuals, terms and conditions or sections that are difficult to access.
- Marking and detection of AI-generated content (Art. 50.2 RIA): the provider must comply with a double requirement: content created or modified with AI must include machine-readable marking and have mechanisms that allow third parties to detect its artificial character. Combined, interoperable, robust and reliable solutions (such as watermarking, metadata or cryptographic techniques), proportionate to the type of content and respectful of data protection regulations are encouraged.
- Texts to inform on matters of general interest (Art. 50.4 RIA): the deployer must explicitly disclose that the text has been generated or manipulated by AI when its purpose is to inform the public. Except where there is human review or effective editorial control and a person or entity with editorial responsibility. Superficial or automated verification is not sufficient.
- Labelling of Deepfakes (art. 50.4 RIA): when an image, audio or video generated or manipulated by AI may appear to be authentic, it is up to the person responsible for the deployment to clearly identify it as such. In personal or domestic uses, the obligation of the controller may not apply; in creative, satirical, artistic or fictional works, transparency must be made in an appropriate manner that does not disrupt their enjoyment.
- Monitoring and sanctions: Adherence to an appropriate code of good practice may help to demonstrate compliance, and competent authorities will monitor these obligations. 15 million or up to 3% of annual worldwide turnover, whichever is higher.
The draft is subject to public consultation until 3 June 2026. The comments received will serve as the basis for the final version of the Guidelines which, although non-binding, are expected to be a key practical reference for interpreting the transparency obligations of Article 50 of the RIA.
Article written by the TMT area of ECIJA Madrid.
Links:
