AI transparency: the delay does not cover Article 50

Regulation (EU) 2026/1744, known as the "Digital Omnibus", was adopted on 8 July 2026, published on 24 July and entered into force on 27 July. It postpones the obligations for high-risk AI systems listed in Annex III — recruitment, credit scoring and others — from 2 August 2026 to 2 December 2027, and those in Annex I, tied to product safety, to 2 August 2028.

Many companies read that delay as a general reprieve. It is not: according to analyses of the text, the Omnibus does not amend Article 50 of the AI Act. The transparency obligations have applied since 2 August 2026, and the deadline on marking generated content, set out in paragraph 2 of that article, falls on 2 December 2026.

What Article 50 asks, in plain terms

This is not a heavy compliance machine, but four situations in which you owe the person in front of you the truth:

• A system that interacts with a person must make clear that it is an AI, unless this is obvious to a reasonably well-informed user.
• Synthetic content — audio, image, video, text — must be marked in a machine-readable way and detectable as artificially generated.
• Deepfakes and manipulated content must be disclosed as such.
• Text published to inform the public on matters of public interest, where AI-generated, must be disclosed — unless a human has taken editorial responsibility for reviewing it.

The exact scope of each case, and in particular the split between provider and deployer, depends on the role the entity plays: legal counsel is the right interlocutor for settling a specific situation.

Why this belongs at group level

Because these obligations do not live in the technical department. They live in the website, customer support, campaigns, the documents sent to clients — that is, everywhere a subsidiary speaks in its own name. A holding company has no business steering its entities' tools; it does have an interest in obtaining comparable answers, as we wrote in our article on AI at board level.

Five questions to put to each entity

• Where does an outside person talk to a system — form, assistant, phone, messaging — and do they know it?
• Which published content is generated or heavily assisted, and under what human review?
• Do images and videos produced by generative tools keep their marking as they pass through the publishing chain?
• Who is accountable for these disclosures — a named person, not a diffuse team?
• What would we do if a client asked tomorrow how a document was produced for them?

Marking survives the publishing chain badly

This is the most common practical difficulty, and it is rarely anticipated. Provenance metadata attached when an image is generated often disappears on a crop, a recompression or a trip through a layout tool. Checking that the marking survives the real publishing path takes a test, not a statement of principle.

Our reading

A calendar delay does not change the substance: an organization that cannot say where it uses AI will be no better able to say so in December 2027. The group's entities keep a short register — uses, owners, data concerned, risk level — and it is from that register that answers to the questions above come, not from a twenty-page policy written for the occasion.

Sources

• Usercentrics — The Digital Omnibus, the high-risk delay and Article 50
• Dastra — What the Omnibus deal changes in the AI Act
• AI Act Article 50 — text