⚡ Quick Summary
Published by the European Commission, this annex sets out guidelines on the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689 (the AI Act). The document states that it is non-binding and that authoritative interpretation of the AI Act may ultimately be given only by the Court of Justice of the European Union; it is addressed to competent authorities, providers and deployers.
The guidelines examine the four transparency obligations in Article 50: informing natural persons that they are interacting with an AI system (50(1)); machine-readable marking and detection of AI-generated or manipulated audio, image, video or text outputs (50(2)); informing persons exposed to emotion recognition or biometric categorisation systems (50(3)); and disclosing deep fakes and AI-generated or manipulated text published on matters of public interest (50(4)). Section 7 sets out the horizontal requirements of Article 50(5) on clear, distinguishable and accessible information at the latest at first interaction or exposure.
Each obligation is broken down into cumulative conditions, exceptions and worked examples, with cross-references to the Digital Services Act, consumer and data protection law, and the European Media Freedom Act. Section 8 covers codes of practice, market surveillance, penalties of up to EUR 15 000 000 or 3% of worldwide turnover, and the 2 August 2026 application date, including a transitional period to 2 December 2026 for Article 50(2) marking.
🧩 What’s Covered
The document runs to nine numbered sections, moving from scope and concepts to enforcement.
- Background and objectives (Section 1): states that the AI Act entered into force on 1 August 2024, that the transparency obligations apply from 2 August 2026, and that the guidelines were informed by stakeholder consultation and input from Member States in the AI Board.
- Overview and horizontal topics (Section 2): tabulates the four Article 50 obligations and their exceptions, defines providers and deployers, and covers exclusions for purely personal non-professional activity, research and development, and free and open-source AI, plus interplay with prohibited practices, high-risk requirements, AI literacy and general-purpose AI models.
- Article 50(1) (Section 3): the four cumulative elements the provision requires (an AI system, intended to interact, directly, with natural persons), coverage of AI agents, disclosure formats and techniques, and the two exceptions for obvious interaction and law enforcement, each illustrated with examples.
- Article 50(2) (Section 4): the marking and detection obligation, the modalities in scope (text, images, audio, video), content falling outside it, the quality requirements of effectiveness, interoperability, robustness and reliability, and the standard editing, industrial or business-to-business and real-time ephemeral carve-outs.
- Article 50(3) (Section 5): definitions of emotion recognition and biometric categorisation systems, who must be informed, by what means and when, and the exception for law enforcement purposes.
- Article 50(4) (Section 6): the notion of deep fake with its four cumulative criteria, the disclosure duty, the attenuated regime for evidently artistic, creative, satirical or fictional works, and the rules for text publications including the human review and editorial responsibility exception.
- Horizontal requirements and enforcement (Sections 7 to 9): Article 50(5) clarity, first interaction or exposure and accessibility; the effects of an adequate code of practice, market surveillance authorities, penalties and entry into application; and review and update of the guidelines.
💡 Why it matters?
The Article 50 obligations apply from 2 August 2026 and engage different actors, so the guidelines matter most where responsibility is contested: whether a company is a provider or a deployer, whether a chatbot's artificial origin is obvious, whether an edit is standard, or whether content is a deep fake. The document supplies operational tests, lists disclosure techniques that are insufficient on their own, and works through examples across media, advertising, games, health and industrial settings. It also connects Article 50 to the DSA, consumer law, data protection and media law, which helps teams sequence overlapping duties instead of treating transparency as a standalone exercise.
❓ What’s Missing
The guidelines are expressly non-binding and leave the central judgements case-by-case: obviousness, substantial alteration, appreciable resemblance and the evidently artistic nature of a work. Harmonised standards and a provider-agnostic interoperable detection solution are described as not yet available, without a timeline for their emergence. Accessibility requirements are cross-referenced to Directives 2016/2102 and 2019/882 but not spelled out. There is no checklist, template or conformity-assessment method, and no quantified guidance on what counts as proportionate in complex content production and distribution chains.
👥 Best For
Serves compliance and legal teams mapping Article 50 duties, product and design teams specifying AI disclosure notices and labels, machine learning and content-provenance engineers implementing watermarking, metadata and detection, and market surveillance or policy staff assessing transparency claims across generative, agentic and biometric systems.
📄 Source Details
Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689 (the 'AI Act'), annex to the Communication to the Commission, C(2026) 5054 final, Brussels, 20 July 2026, 51 pages, in English. No individual authors are printed; the issuing body is the European Commission. The complete text extraction of all 51 pages was available, including the table of contents and footnotes.
