The Council of the EU gave final green light to the Omnibus on 29 June 2026, following the European Parliament’s formal endorsement on 16 June. The adopted text does two things to Article 50 that last week’s edition did not distinguish. It postpones the provider obligation under Article 50(2), the requirement to watermark AI-generated content in a machine-readable format, from 2 August 2026 to 2 December 2026. It leaves the deployer obligations under Article 50(1), 50(3), and 50(4) exactly where they were: disclosing to a natural person that they are interacting with an AI system, labelling deepfakes, and flagging AI-generated content on matters of public interest all remain due on 2 August 2026.

Last week’s piece said Article 50 was not moved. For the deployer obligations, that is still true. For the provider watermarking obligation, it was wrong the day it published, because the final text had already settled the point ten days earlier.

Why this happened

The error traces to a distinction this newsletter argues every week and, on this occasion, failed to apply to itself. Reporting on the provisional political agreement reached on 7 May treated Article 50 as broadly stable while the high-risk Annex III timeline moved. That was an accurate read of the provisional text. It stopped being accurate the moment the Council adopted the final version on 29 June, and nobody checked the newsletter’s Article 50 claim against that later, authoritative text before publication on 5 July.

That is a small failure by the standard of most governance commentary. It is a large one by the standard this newsletter sets for everyone else. The argument made here every week is that a compliance record is only as good as its contemporaneity, whether it was checked and dated at the point that mattered, not reconstructed afterward from whatever was true when someone last looked. Last week’s Article 50 claim was reconstructed from a source that was accurate when written and stale by the time it was published. No one checked it against the adopted text before it went out under this byline.

What actually applies now

Two obligations, two dates, and they should not be tracked as one line item:

Provider watermarking, Article 50(2): AI-generated content must carry a machine-readable marking identifying it as artificially generated or manipulated. Due 2 December 2026, moved back four months from the original date.

Deployer disclosure, Articles 50(1), 50(3), 50(4): informing a person they are interacting with an AI system, labelling synthetic audio, image, video, or text content that constitutes a deepfake, and disclosing that text published on matters of public interest was AI-generated. All three remain due 2 August 2026, unchanged.

An organisation that builds AI-generated content in-house is a provider for that content and has until December. The same organisation, deploying a chatbot that interacts with customers, is a deployer for that system and has three weeks.

The governance point underneath the correction

If a newsletter whose entire purpose is tracking exactly this kind of distinction can publish an inaccurate compliance date, the question worth asking is not how the mistake happened. It is what confidence a board should have in whichever internal function is tracking the same obligation without a newsletter’s dedicated attention to it.

The standard this newsletter has argued for elsewhere applies here without exception. A correction is only as good as the record behind it: who checked the adopted text against the previous claim, when they checked it, and whether that check happened before or after the gap was found. This piece is that record. It is dated, it is attributed, and it corrects against the primary legislative text rather than secondary commentary, including this newsletter’s own commentary from a week ago.

The wider context, briefly

The correction lands the same week the UN’s first independent scientific panel called for shared global rules on AI to prevent catastrophic harm, the UN and ITU launched an AI for Good Global Commission in Geneva, and a UK parliamentary committee warned that the country lacks sovereign control over its own AI infrastructure. All three are real and worth boards’ attention. None of them changes what a specific board needs to know about a specific compliance date inside a regulation that has already been adopted. Global governance debates move in years. The gap between a provisional agreement and a final text moved a compliance date by four months in the space of six weeks, and most commentary, including this newsletter, did not catch it until now.

Regulatory references: EU AI Act Articles 50(1), 50(2), 50(3), 50(4), 65; Companies Act 2006 section 174; Council of the EU final adoption, 29 June 2026.

The Roche-Review is the weekly publication¹ of Dr Ivan Roche FRSSy FRSA MInstP, Founder of Otopoetic Limited. Subscribe at roche-review.com.