You have until 2 December 2026. The Omnibus grants a transitional period of four months, not six, and the difference is the entire margin most organisations think they have.
Article 50 of the EU AI Act applied from 2 August 2026. Its disclosure duties — telling a person they are dealing with an AI — took effect that day with no relief of any kind.
The marking duty in Article 50(2) is different. Generative systems already on the market before 2 August were given a transitional period by the Digital Omnibus, Regulation (EU) 2026/1744. The regulation states the length of that period rather than a date:
“To allow sufficient time for providers of generative AI systems subject to the marking obligations laid down in Article 50(2) of Regulation (EU) 2024/1689 to adapt their practices within a reasonable time without disrupting the market, it is appropriate to introduce a transitional period of four months for providers who have already placed their systems on the market before the 2 August 2026.”
Regulation (EU) 2026/1744, recital 38
Four months from 2 August is 2 December. Not February.
Why February is in circulation. There is a real 2 February 2027 date in this area: the Code of Practice sets it for watermark-detection interoperability. Different instrument, different obligation, and voluntary. The Commission's original proposal also floated six months before the co-legislators settled on four. Two nearby dates and a superseded draft is enough for the wrong one to spread.
We know, because we published the wrong one ourselves and had to correct it. That correction is on our evidence page with the other three.
Article 50(2) requires that generated output be marked in a machine-readable form, so that a system downstream can detect it as artificial. A visible “made with AI” label in your interface is a disclosure, and it is a good thing, but it is not this. The obligation is about the artefact, not the interface.
In practice that means provenance metadata travelling with the file — C2PA Content Credentials being the standard most of the industry has converged on — and, for formats where metadata is easily stripped, an embedded signal in the content itself.
We scanned the public pages of twenty European generative-AI companies in August 2026 — homepage, pricing, legal, privacy, terms, trust and security. None of them published any mention of C2PA, Content Credentials or content provenance on those pages. That is a statement about what is documented publicly, not about what is implemented: several of them may well have it in the pipeline and simply not written about it. But a buyer or a regulator reading those pages cannot tell, and increasingly both will look.
Three things, depending on where you actually are. Prices and terms below come from one table and are not typed into this page — the last figure this site typed by hand was wrong for two days.
Something was postponed and something was not, and the two get quoted as one thing constantly. The Omnibus deferred the high-risk regimes by sixteen months and left Article 50 exactly where it was. If the advice you were given said the Act was pushed back, it was describing a different part of it. That is what the Second Opinion is for, and we would rather tell you the advice was right than sell you a correction you do not need.
Four corrections are published on our evidence page, three of them found by something other than our own review. If an entry here does not follow from the source it cites, that is a defect and we would rather be told than found out.
Contact