Most people in this sector read one headline in June and stopped. The AI Act was delayed. The August cliff moved. Back to work.
Half of that is right.
The high-risk regime under Annex III, the one covering education and the evaluation of learning outcomes, has been deferred to 2 December 2027. As of yesterday that is settled law rather than a political agreement: the Digital Omnibus was published in the Official Journal on 24 July as Regulation (EU) 2026/1744 and entered into force on 27 July. High-risk systems embedded in regulated products move to 2 August 2028.
Both are now unconditional calendar dates. That is worth knowing, because the Commission's original proposal tied those deadlines to an assessment of whether the supporting standards were ready, and commentary written against that earlier draft is still circulating. The adopted text does not work that way. The dates are fixed.
That is genuine relief, and it is the part everybody read.
The part almost nobody read is that Article 50 did not move. It applies from Sunday.
Article 50 is the transparency layer. It has nothing to do with risk classification or conformity assessment. It is about telling people when they are dealing with a machine, or looking at something a machine made. And while this sector spent eighteen months watching the high-risk deadline, it also spent those eighteen months filling programmes with synthetic presenters, generated voiceover, AI illustration and the occasional tutor bot.
Those are precisely the things Article 50 is about.
The finding that makes this concrete
On 20 July the Commission adopted its final guidelines on Article 50. They are eight days old. They are interpretive rather than binding, but in practice they are how this will be read by everyone who matters.
The guidelines state expressly that "persons" in the deepfake definition covers digital replicas of real people, realistic AI-generated avatars and personas, and personal characteristics including a person's image, voice, behaviour and performances.
And, more pointedly: a realistic synthetic depiction of a fictitious but natural-looking person constitutes a deepfake within the meaning of Article 3(60), even where no identifiable rights-holder is involved. Content must be labelled even where no deception was intended and even where no real individual is depicted.
Read that against your own library.
The presenter you generated last spring, the one who does not exist, who was invented precisely so that nobody's likeness was used, is a deepfake in the Act's terms. Not as an accusation. As a definition.
Inventing a person who does not exist was the careful thing to do. It is also the thing that brings you inside the definition.
Clearly unrealistic material stays outside it. Fantasy scenes, things that defy biology, illustration nobody would take for a photograph. The line is realism, not intent.
What you might be holding, and how each part is treated
A synthetic video or audio presenter. This is the deployer disclosure duty under Article 50(4). You must disclose that the content is artificially generated, to the person exposed to it, at first exposure at the latest.
An AI tutor, coach or support chatbot inside the programme. Article 50(1). The system has to make clear the learner is talking to a machine, at the point of interaction. The guidelines are explicit that a disclosure buried in terms and conditions does not discharge this, and that helpdesk-style chatbots often will not clear the "obvious to a reasonably well-informed person" exemption. Agentic systems are in scope too.
AI-generated images and illustration. The same realism test applies. A photorealistic scene of people who do not exist sits differently from an obviously stylised graphic.
AI-drafted written content. Mostly not caught. The text obligation bites on text published to inform the public on matters of public interest. Course copy, assessment items and learner guidance are generally not that. Worth saying plainly, because the useful half of any compliance question is knowing what you can stop worrying about.
Machine-readable marking of the file itself. Not your duty. That sits with the provider of the generative tool under Article 50(2), and the adopted text gives those providers a split deadline: tools already on the market before 2 August 2026 have until 2 December 2026, while anything placed on the market from 2 August onwards must comply straight away. Two things follow. Ask your vendors where they are with it. And do not assume it covers you, because the guidance is explicit that a deployer cannot rely on the provider's embedded marking to satisfy their own disclosure obligation. Platform-applied labels do not discharge it either.
The date that matters is generation, not publication
This is the most practically useful thing in the final guidelines, and it changed between draft and final.
For image, audio and video, the trigger is the date the content was generated. Material generated before 2 August 2026 does not have to be labelled retroactively. The Commission encourages labelling it, but explicitly does not expect disproportionate effort, and names auditing content databases as an example of what is not expected.
For text on matters of public interest the trigger is different: the date of publication. So text generated in July and published in September is caught.
The practical shape of that is simple. Your back catalogue is largely fine. Your next render is not. If you have a production run scheduled for August, that is the thing to look at this week. The archive can wait.
The argument you could make, and why I would not lean on it
The final guidelines soften the fourth limb of the deepfake test. Whether content "would falsely appear to be authentic" now requires a holistic assessment: the level of resemblance, the substantive message, the deployment context, the environment in which it appears, and the composition and expectations of the foreseeable audience. Where an audience does not expect content to be authentic in a given context, it may fall outside the definition.
A learner who has logged into a CPD module arguably does not expect the presenter to be a real colleague. That is a genuine argument and it may well be right.
It is also an argument you would be making after the fact, about a judgement call, to a market surveillance authority, with penalties reaching fifteen million euro or three per cent of worldwide turnover. The artistic and fictional carve-out does not rescue you either. It limits the manner of disclosure so it does not spoil the work. It does not remove the duty.
Set against that, the cost of compliance is one line of text on screen before the presenter opens their mouth.
One more thing worth knowing if you sit outside the EU. These duties follow the audience, not the establishment. A UK training provider serving learners in Ireland or elsewhere in the Union is in scope.
What adequate disclosure looks like
Clear and distinguishable. Noticeable, easy to understand, and easy to identify as a disclosure. Perceivable without special tools, which means visible or audible, not metadata. Available at first exposure at the latest. Accessible, which the guidelines treat as part of clarity rather than as an extra.
For a module, that is a sentence on the opening screen. Across a library it is a morning's work.
The part that is not about compliance
There is a version of this where a certification body adds a line of disclosure text because a regulation says it must, resents the intrusion, and makes the line as small as the rules allow.
That wastes the moment.
If what you sell is assurance, then how the programme was made is not an embarrassment to be minimised. It is evidence that you know what you are doing. A body that states plainly which parts were machine-generated and which parts were checked by a named human is making a claim about its own standards. A body that stays quiet is asking learners to trust a credential while declining to say how it was built.
The programmes that age well out of this period will be the ones that treated disclosure as part of the product, rather than a footnote bolted on the Friday before a deadline.
Informed commentary rather than legal advice. Article 50 duties are role-specific, and the Commission's guidelines are interpretive rather than binding, so read your own position against the text of the Regulation and take advice where the stakes warrant it.
Not sure what sits inside your own programmes, or how much of it was generated? The Programme Design Diagnostic is a fixed-fee, independent read on a single programme: a board-ready diagnosis and a costed build scope, in two to three weeks.
For the enforcement timeline itself, what arrives on 2 August and what was deferred, the companion piece is what actually happens on 2 August 2026. For the separate literacy obligation under Article 4, see AI literacy in plain English. See more insights from LearnFrame.