Must AI Be Disclosed in the Physician Letter? Article 50 of the EU AI Act, Read Soberly

Friday afternoon, the signature folder. This time the draft of the discharge letter comes from an AI system; the resident has checked it against the record, corrected two doses, rearranged a paragraph. While signing, she asks: "Do we actually have to write in that an AI helped write this?"

The question deserves a more precise answer than most fact sheets provide. The legal situation consists of one part that is settled, one part that is interpretation — and a gap that nobody has closed so far.

What applies from 2 August 2026

The transparency obligations of Article 50 of the EU AI Act apply from 2 August 2026. The Digital Omnibus package has not changed that: what was postponed were the obligations for high-risk systems (to December 2027 and August 2028 respectively for AI embedded in products); Article 50 was left untouched [1]. Anyone in the hospital who was waiting for a postponement was waiting for the wrong provision. The European Parliament approved the Omnibus package on 16 June 2026, the Council on 29 June; publication in the Official Journal was expected for July 2026 [1].

What Article 50 requires — and of whom

The article addresses providers and deployers differently, and this distinction carries half the answer:

• Paragraph 1 obliges providers: AI systems that interact directly with people must be designed so that those affected learn they are interacting with an AI — unless it is obvious to a reasonably informed person [2]. A patient chatbot on the hospital website or in the portal clearly falls under this.
• Paragraph 2 likewise obliges providers: outputs of generative systems must be recognizable as artificially generated in machine-readable form [2]. This obligation falls on the manufacturer of the tool; for the hospital as deployer, it is above all a test question for vendor meetings.
• Paragraph 4 obliges deployers — but in a narrow field of application: what must be disclosed is AI-generated or AI-manipulated text that is published in order to inform the public about matters of public interest. Even there, the obligation lapses if the content underwent human review and a natural or legal person carries editorial responsibility [2].

The physician letter against the wording

Holding the text of the norm up to it, much suggests that the discharge letter to the colleague continuing treatment is not even covered by the elements of the provision: it is not published, and it informs not the public but a specific addressee about a specific patient. And even someone who saw that differently would run into the exception: a letter that a physician has reviewed and stands behind with her signature is the textbook example of human review with an assumption of responsibility.

That is an interpretation of the norm's wording — the obvious one, but still an interpretation. No official guideline confirming it exists so far; nor one refuting it [2]. What can seriously be said at present is only this: an EU-law obligation to label AI involvement in physician letters cannot, as things stand, be derived from Article 50 — with the residual uncertainty that every unsettled legal question carries. Vendors or consultants selling certainty in either direction are selling too much.

The German gap

What is remarkable is everything that does not exist in Germany: no rule in the Musterberufsordnung — the model professional code for physicians —, no requirement from the Bundesärztekammer (the German Medical Association) or any state chamber of physicians that demands labeling of AI support in physician letters — or prohibits it [3]. After repeated checking, this is a genuine regulatory gap.

What does exist draws the line of responsibility elsewhere. The Scientific Advisory Board of the Bundesärztekammer records that responsibility for diagnosis, indication and therapy remains medical and cannot be transferred to AI; AI results are not to be adopted uncritically [3]. The 129th Deutscher Ärztetag — the German Medical Assembly — in 2025 reaffirmed physicians' ultimate responsibility and expressly urged clarification of the liability questions [4]. The Deutsche Krankenhausgesellschaft, the German Hospital Federation, sees large language models as possible support for discharge letters too, but considers human oversight and review "still indispensable" — and names as an open question where the manufacturer's responsibility ends and the hospital's liability begins [5].

One obligation has long applied all the same: since 2 February 2025, Article 4 of the AI Act requires deployers to ensure the AI literacy of their staff; the DKG specifically points to possible personal liability of decision-makers [5].

What you can decide today

1. Formulate a house position: is AI support documented internally — who generated, who reviewed, who signed off — even if the letter itself carries no label? A documented position is worth more in any later discussion than silence.
2. Make the review provable. The exception in Article 50 (4) presupposes an actual human review. A sign-off that is traceable in the system — who, when, with which changes — carries further than one merely asserted. What should happen between draft and signature is described in the article on the defensible physician letter pipeline (in German).
3. Treat patient-facing AI separately: for chatbots and portals, paragraph 1 is unambiguous. There is no interpretation question here, only an implementation question.
4. Ask the vendor question: how does the manufacturer fulfill the machine-readable labeling under paragraph 2? The question belongs in every contract review — alongside the ten points of the data processing agreement checklist for hospital AI (in German).

The interpretation of Article 50 for clinical documentation remains in motion; we follow it continuously in our weekly briefing Visite (German; English edition Grand Rounds is in preparation). If you want to formulate a house position for your hospital, write to us.

Sources

1. Gibson Dunn: EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes (2026); Sidley Austin, Data Matters: EU AI Act Transparency Obligations — Preparing for Compliance by 2 August 2026 (24.06.2026). Kanzleianalysen. https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/ ; https://datamatters.sidley.com/2026/06/24/eu-ai-act-transparency-obligations-preparing-for-compliance-by-2-august-2026/
2. Verordnung (EU) 2024/1689 (KI-Verordnung), Art. 50. https://artificialintelligenceact.eu/article/50/
3. Bundesärztekammer, Stellungnahme des Wissenschaftlichen Beirats „Künstliche Intelligenz in der Medizin“, 14.01.2025. https://www.bundesaerztekammer.de/fileadmin/user_upload/wissenschaftlicher-beirat/Veroeffentlichungen/KI_in_der_Medizin_SN_neu.pdf
4. 129. Deutscher Ärztetag, Leipzig, 27.–30.05.2025 — Beschlussprotokoll. https://www.bundesaerztekammer.de/fileadmin/user_upload/BAEK/Aerztetag/129.DAET/2025-05-30_Beschlussprotokoll_129._DAET.pdf
5. Deutsche Krankenhausgesellschaft, Positionspapier „Nutzung von Künstlicher Intelligenz (KI) im Krankenhaus“, 23.10.2025. https://www.dkgev.de/fileadmin/default/Mediapool/1_DKG/1.3_Politik/Positionen/2025-10-23_DKG-Positionspapier_KI_im_Krankenhaus.pdf

Sources accessed in July 2026.