Written by: Haim Ravia, Dotan Hammer
The European Commission’s AI Office has opened a public stakeholder consultation on Draft Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689 (the “AI Act”). Published on 8 May 2026 by the Directorate-General for Communications Networks, Content and Technology, the draft is intended to assist competent authorities, providers and deployers of AI systems in ensuring “consistent, effective and uniform” compliance ahead of the entry into application of the Article 50 transparency regime on 2 August 2026.
According to the draft, Article 50 sets out four cumulative transparency obligations covering distinct categories of AI systems and outputs—interactive systems, synthetic content (audio, image, video, and text), emotion recognition and biometric categorization systems, and deepfakes and AI-generated text published to inform the public on matters of public interest. The Commission notes that the same AI system may trigger several obligations simultaneously, engaging the responsibility of both providers and deployers, and that the regime applies extraterritorially: third-country providers and deployers fall within scope whenever the system’s output is used in the Union. The Guidelines are themselves non-binding; only the Court of Justice of the European Union may give authoritative interpretations of the AI Act.
Provider obligations (Articles 50(1) and (2)). Providers must design interactive AI systems—chatbots, voice assistants, AI companions, agentic systems, and humanoid robots—so that natural persons are informed of the AI origin no later than at first contact through multimodal, accessibility-adapted disclosures. Providers of generative AI systems must additionally embed machine-readable marks (watermarks, metadata, cryptographic provenance, and fingerprints) and ensure that the detection of synthetic audio, image, video, or text content meets combined effectiveness, robustness, reliability, and interoperability standards.
Deployer obligations (Articles 50(3) and (4)). Deployers must inform natural persons exposed to emotion recognition or biometric categorization systems, regardless of whether operation is real-time or ex post, and must clearly and distinguishably label AI-generated or manipulated deepfake image, audio, or video content, as well as AI-generated text published to inform the public on matters of public interest. Attenuated disclosures apply to evidently artistic, creative, satirical, fictional, or analogous works, and substantive human review, coupled with assumed editorial responsibility, exempts text publications.
Enforcement and timing. The transparency regime applies from 2 August 2026, with a targeted grandfathering rule for pre-existing generative AI systems envisaged under the AI Omnibus proposal currently before the EU co-legislators. Enforcement rests with national market surveillance authorities, the AI Office, and the European Data Protection Supervisor, and infringements are punishable by fines of up to EUR 15,000,000 or 3% of worldwide annual turnover, with up to EUR 750,000 for EU institutions, bodies, and agencies.
The Commission frames the strategic objective as reducing the risks of “impersonation, deception, misinformation, manipulation at scale, and fraud” and mitigating adverse impacts on democratic processes and societal trust. Adherence to the parallel Code of Practice on transparency of AI-generated content—currently being finalized—will be treated as a straightforward way to demonstrate compliance, although providers and deployers may also rely on alternative means and may then face more detailed information requests from competent authorities.
Click here to read the draft guidelines on AI transparency.
Click here to access the European Commission’s consultation page.
