Two duties, two parties
Article 50(2) is a provider duty and it is about machine readability: the output has to carry a marking that lets a machine detect it was artificially generated or manipulated. Article 50(4) is a deployer duty and it is about people: deep fakes must be disclosed as such, and AI-generated or AI-manipulated text published to inform the public on matters of public interest must be disclosed too. Meeting one does not meet the other.
The one thing that was postponed — and what was not
Regulation (EU) 2026/1744 postponed the Article 50(2) machine-readable marking duty to 2 December 2026, and only for systems placed on the market before 2 August 2026. Everything else in Article 50 kept its original date. This is widely misread as 'Article 50 has been delayed'. It has not: the transparency duties are in force, and the deferral is a narrow grace period on one paragraph for existing systems.
The Code of Practice is voluntary, and useful anyway
The Commission published a Code of Practice on marking and labelling AI-generated content on 10 June 2026; by the end of July roughly 190 organisations had signed. It is voluntary and signing it is not compliance in itself, but it is the clearest available reference for how and where to display a label, including alternatives such as audio disclaimers where a visual label is impossible.
Where this bites in practice
Marketing images generated in-house, AI-assisted articles on public-interest topics, synthetic voice in customer service, and any content pipeline where a model drafts and a person lightly edits. The question to ask per pipeline is not 'did AI help' but 'is this output synthetic or manipulated, and who placed the system on the market'.
The Commission's own guidance
Alongside the Code of Practice, the Commission published guidelines on the Article 50 transparency obligations on 20 July 2026, applying from 2 August. They cover all four disclosure duties together — interaction, machine-readable marking, deep fakes and AI-generated public-interest content without human review or editorial control — which makes them the practical companion to the split described above.
What to do
Split your inventory in two: systems you provide (mark the output machine-readably) and content you publish using someone else's system (disclose where it is a deep fake or public-interest text). For systems already on the market before 2 August 2026, note 2 December 2026 as the marking date rather than assuming you have longer. Read the Code of Practice for the display conventions and record which of them you follow.
Sources
Last verified against the primary sources: 2026-08-04
Check your own systems
The free
Article 50 check maps one system to the transparency duties in about two minutes, quoting the official text per answer. No sign-up, and the full result before any contact details.
← EU AI Act overview