EU opens consultation on Article 50 guidance

The European Commission published draft guidance on the transparency obligations in Article 50 of the EU AI Act and opened a targeted consultation on May 8, 2026, that runs through June 3, 2026.

Article 50 sets four distinct transparency duties—disclosure for interactive systems, machine‑readable marking and detectability for synthetic content, notifications for emotion recognition and biometric categorisation, and labelling of deepfakes or public‑interest text—and the obligations take effect on August 2, 2026.

The draft clarifies Article 50(1), which requires providers of AI systems that directly interact with people to inform users they are dealing with AI unless the artificial nature of the interaction is obvious or an exception applies. The guidance stresses in‑context notices, not buried terms and conditions.

Article 50(2) is the most technical: providers must mark outputs in a machine‑readable way and make outputs detectable as AI‑generated or manipulated. The draft lists candidate techniques including watermarks, metadata identifiers, cryptographic provenance proofs, logging and fingerprinting, and allows combinations of methods.

The draft emphasises that marking alone is not enough: providers must also provide means of detection and ensure interoperability so that detection tools produce human‑readable results at first exposure. The guidance allows providers to rely on third‑party or upstream marking so long as responsibility for compliance is demonstrable.

For Article 50(3), deployers of emotion‑recognition and biometric categorisation systems must inform people exposed to those systems, whether in real time or after the fact. The Commission ties those notifications to the broader high‑risk regime that governs emotion recognition.

Article 50(4) covers deepfakes and AI‑produced text on matters of public interest and requires deployers to disclose manipulation or synthetic origin. The draft narrows and explains exceptions for clearly artistic, satirical or editorially controlled content but keeps the obligation broad for material that could mislead vulnerable audiences.

The guidance is non‑binding but practical: the Commission expects market‑surveillance authorities and the AI Office to use it when monitoring compliance. The draft also flags penalties for noncompliance, up to €15 million or 3% of worldwide annual turnover, whichever is higher.

The consultation is explicitly targeted at providers, deployers, SMEs, large companies, public authorities, academia and civil society and asks technical and operational questions about traceability, labelling, provenance chains and practical detection methods. The goal is to collect feedback on how marking, detection and provenance should work in real‑world systems.

For vendors and deployers serving EU customers, the draft guidance and the consultation are immediate compliance‑relevant items. The guidance reiterates that deployers have obligations that are separate from, and in some cases additional to, provider duties, and legacy generative systems may get a limited additional transition period in specific cases.

The draft also points to a voluntary Code of Practice on transparency of AI‑generated content that is being finalised in parallel; adherence to an adequate code may be taken into account as evidence of compliance, though it is not the sole route to meeting Article 50. The Commission warns that firms deviating from the guidance without documented justification face regulatory risk.

Practically, the draft and consultation push providers and deployers to map systems against Article 50 categories, pilot interoperable marking and detection mechanisms, document provenance choices and be ready to supply detection tools or interfaces to competent authorities. The Commission is inviting concrete technical proposals in the questionnaire until June 3, 2026.