Article 50 of the EU AI Act: from transparency obligations to implementation questions

Written by Zoi Roupakia

The EU AI Office published its draft guidelines on Article 50 transparency obligations last month. Noetic AI served on the working groups for the Code of Practice on Transparency of AI-Generated Content, which supports implementation of the Article 50 obligations on marking and detection of AI-generated content and the labelling of deep fakes and certain AI-generated text. This note sets out our reading of the draft guidelines and the implementation questions they raise.

Article 50 of the AI Act covers four transparency obligations: informing users when they are interacting with an artificial intelligence (AI) system; machine-readable marking and detection of AI-generated content; disclosure for emotion recognition and biometric categorisation systems; and labelling of deep fakes and AI-generated text published in the public interest. The Article 50 obligations apply from 2 August 2026; the guidelines are intended to support their implementation.

The document is detailed and useful on the mechanics of implementation. Its treatment of agentic AI systems is notable, especially in recognising that providers may not always be able to identify every instance in which an AI agent interacts with a natural person. In those cases, the guidelines state that agents should be instructed to disclose their nature whenever such interaction is likely. The worked examples throughout the document are helpful, and the opening summary table clarifies the structure of the Article 50 obligations.

Three implementation questions, however, deserve further attention before the consultation closes.

Marking and detection: technically demanding, but not yet operationally proportionate

Article 50(2) requires machine-readable marking and detection of AI-generated content using solutions that are effective, interoperable, robust, and reliable. The guidelines acknowledge (paragraph 78) that no single technique currently meets all four criteria simultaneously, meaning providers must combine methods. The guidelines also refer to technical feasibility and implementation costs, but they do not translate these considerations into a clear, operational proportionality framework for startups and small and medium-sized enterprises. Smaller providers may therefore face uncertainty about what level of marking and detection capability is sufficient, particularly where state-of-the-art technical solutions are still evolving.

The obviousness exception in a changing technical landscape

The Article 50(1) exception for interaction that is obvious to a reasonably well-informed, observant and circumspect person borrows from European Union consumer law. The exception is defensible in principle, but it also raises a temporal problem. What appears “obviously” artificial today may not remain obvious as AI systems become more human-like, multimodal and context-aware. A voice assistant that was recognisably non-human two years ago may not be today. The draft guidelines recognise the need to assess context, audience and AI literacy, but they do not yet sufficiently operationalise how providers should reassess obviousness over time as capabilities and user expectations change.

Industrial and B2B exemptions need clearer boundaries to support proportionate implementation

Paragraphs 81 and 82 take a sensible approach by recognising that some industrial and business-to-business AI applications should not be subject to marking requirements where outputs are strictly technical in nature and accessed only by a limited, pre-defined number of natural persons acting in a professional capacity. This flexibility is important for innovation, especially in technical and internal business settings where the risks of public deception, impersonation or information manipulation are low.

However, both conditions may generate interpretive divergence across Member States, as national market surveillance authorities may define “strictly technical” differently. This matters because some technical outputs may still influence human judgement, worker safety or organisational accountability, for example AI-generated instructions, risk assessments, workflow recommendations or quality-control reports. The guidelines should therefore preserve the proportionality of the exemption while setting clearer boundaries for when technical outputs remain sufficiently internal and low-risk to justify lighter treatment.

The direction of the guidelines is welcome. The remaining questions are practical ones: how to apply the rules consistently, proportionately, and in ways that improve user understanding rather than simply adding procedural disclosures. The consultation is open because the Commission is seeking input before these guidelines are finalised. Stakeholders that build or deploy AI systems in Europe can provide feedback through the consultation portal until 3 June.

Consultation portal: https://digital-strategy.ec.europa.eu/en/consultations/consultation-draft-guidelines-transparency-obligations-under-ai-act