If you follow this from a distance, you probably came away thinking "the AI Act came into force in August 2026". That is half true, and the missing half is exactly the half that matters when you are buying software for a health service.
Two things happened days apart, and they pull in opposite directions.
The two dates
27 July 2026. The so-called Digital Omnibus on AI (Regulation (EU) 2026/1744) entered into force. Among other things it set fixed deadlines for the high-risk obligations that were previously tied to the pace of technical standardisation, and postponed them.
2 August 2026. The transparency obligations of Article 50 started to apply. These were not postponed: they are in force now.
The result is a three-step calendar:
| When | What |
|---|---|
| 2 Feb 2025 | Article 5 prohibitions and the AI literacy duty |
| 2 Aug 2025 | General-purpose AI model obligations |
| 2 Aug 2026 | Transparency (Article 50) |
| 2 Dec 2027 | Annex III high-risk (stand-alone systems: hiring, credit, education) |
| 2 Aug 2028 | Annex I high-risk: AI embedded in already-regulated products, including medical devices |
It is the last step that gets misread. An AI system inside a medical device does not have to meet high-risk requirements this month: it has until 2 August 2028. But two caveats are worth keeping in view. The deferral lapses if the system undergoes a significant design change. And none of this touched the MDR: the obligations of Regulation 2017/745 stayed exactly where they were.
What transparency actually requires
Article 50 is not about performance or clinical safety. It is about the person on the other side knowing what they are dealing with. It covers four situations:
- Direct interaction: if you are talking to an AI system, you must be told, unless it is obvious.
- AI-generated content: synthetic image, audio and video must be marked in a machine-readable way.
- Emotion recognition and biometric categorisation: exposed people must be informed.
- Deepfakes and text on public-interest matters: publishing AI-generated text to inform the public on matters of public interest must be disclosed.
Point 4 has an exemption that is almost never quoted: it does not apply where the text underwent meaningful human review and someone assumes editorial responsibility. In other words, an article written with AI assistance and reviewed by a person who answers for it is, to the letter, outside the duty.
The exemption is worth keeping in mind, because it changes the useful question to ask whoever is publishing. It is not "was this written with AI help?", the answer is increasingly yes, everywhere, and it will soon distinguish nobody. It is who takes editorial responsibility for it? On this blog the answer is the RVer team, for the same reason we publish what we cannot prove about return on investment.
Where this leaves RVer
RVer is a Class I medical device registered in EUDAMED and with INFARMED (Portugal). It has an AI-assisted feature, and that feature does one thing: it helps find a scenario in the library based on the kind of moment: relaxation, settling agitation, memory, distraction, sleep, stimulation, orientation. It ranks the content and explains why it suggested it.
What it does not do, and this is where the conversation becomes useful:
- it does not prescribe;
- it does not assess;
- it does not diagnose;
- it decides nothing on its own, and there is no way to make it decide.
The choice always belongs to the clinical team. This is not a position of convenience adopted because of the regulation: it has been on the site since before it, and we set it out in why AI is not ready to be the clinician.
The questions worth asking a supplier
If you are evaluating clinical software with AI, three questions quickly separate those who thought about this from those who did not:
- Does your AI decide or suggest? If it decides, and the product is a medical device, you are looking at Annex I and at 2028, and at a classification that may not be the one you were told.
- Where does the model run, and what leaves the institution? That is a GDPR question before it is an AI Act one, and the two intersect.
- What design changes are planned? Because a significant change makes the deferral lapse.
What we do not know
Worth saying plainly: this is not legal advice. We are a manufacturer, not a law firm. Parts of Article 50 are still settling in practice for small systems, and Commission guidance keeps arriving. If your institution needs a formal position, it has to come from someone who can sign it.
What we can state is what we did on our side, which is public and checkable.
In short
- 2 August 2026: Article 50 transparency applies.
- 2 August 2028: high-risk for AI inside medical devices: deferred, not cancelled, and the deferral lapses on significant design changes.
- The MDR did not change.
- In RVer the AI suggests and the team decides.
Frequently asked questions
Does the AI Act already apply to my clinical software?
It depends on the obligation. Article 50 transparency duties apply from 2 August 2026. High-risk obligations for AI embedded in already-regulated products, such as medical devices, only apply from 2 August 2028, following the Digital Omnibus that entered into force on 27 July 2026. Neither changed the MDR.
Does a medical device with AI automatically become high-risk?
Not automatically, and the answer depends on what the AI does and how the product is classified. What changes the conversation is AI deciding rather than suggesting. In RVer the AI-assisted feature ranks and explains scenario suggestions from the library; it does not prescribe, assess or diagnose, and the choice always belongs to the clinical team.
Does text written with AI help have to be labelled?
Article 50 requires disclosure of AI-generated text published to inform the public on matters of public interest, but exempts text that underwent meaningful human review with editorial responsibility assumed. The articles on this blog are reviewed by the RVer team, which assumes that editorial responsibility.