2026-07-30 — Your code is exempt. Your UI isn't.

The Commission adopted its Article 50 guidelines on 20 July, thirteen days before the rules apply. Source code is carved out of the marking duty; the two-way chat surface around your agent is not — and Article 50 was never part of the Omnibus deferral. Classify surfaces, not models.

Post

Your coding agent's output is exempt on Sunday. The chat box around it is not.

On 2 August the EU AI Act's Article 50 transparency rules start applying. The Commission adopted the implementing guidelines on 20 July — thirteen days of lead time. There is also a Code of Practice on marking AI-generated content, published 10 June. Adherence is voluntary, and per the Commission's own page it is still undergoing adequacy assessment: the presumption of conformity arrives only if that assessment lands positively. Signing it today is a bet, not a shield.

Most teams filed this under "deepfakes" and moved on. Read the exemptions and it gets more specific than that.

Excluded from the machine-readable marking duty: source code. Short sequences of numbers, symbols or letters. Machine-to-machine output with no human exposure. Closed-loop industrial and product development environments. Assistive functions for standard editing.

Not excluded: the interface. Article 50(1) binds the provider of any system that interacts directly with a person — a genuine two-way exchange — to make it known that the person is talking to an AI. The "it's obviously an AI" escape hatch exists, and the Commission's own FAQ says to read it restrictively, judged against a reasonably well-informed, circumspect and observant average person.

And Article 50 was not deferred. The Digital Omnibus pushed stand-alone high-risk obligations to 2 December 2027 and regulated products to 2 August 2028. It left Article 50's scope alone. The one concession is a grandfathering rule: generative systems already on the market before 2 August have until 2 December 2026 — for the marking duty only, not for disclosure.

What I'd do Monday:

  1. Enumerate every user-facing surface and map each to 50(1), 50(2), 50(3) or 50(4). A coding agent usually has both a 50(1) chat surface and 50(2) non-code output.
  2. Emit provenance metadata at generation time, not at export. A mark added downstream is not attached to the artefact that actually left your system.
  3. Log the disclosure event as a first-class record. "Banner rendered" is not evidence. "Disclosure D shown to user U at time T" is.

Article 99(4) sets the ceiling at €15,000,000 or 3% of worldwide annual turnover. For SMEs and start-ups, paragraph 6 flips it to whichever is lower — a real difference for a small shop, and one almost nobody quotes.

Which of your surfaces did you forget to classify?

#EUAIAct #AgenticAI #AIGovernance #AIEngineering #Compliance

First comment

Sources, primary first:

The Commission's Article 50 guidelines and the Code of Practice, adopted 20 July 2026: https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems

The Commission's own FAQ — this is where the source-code and machine-to-machine exemptions are spelled out, and where the "obvious" exception is told to be read restrictively: https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act

The announcement, with the full list of documents published on 20 July: https://digital-strategy.ec.europa.eu/en/news/commission-publishes-guidelines-transparency-obligations-providers-and-deployers-certain-ai-systems

Article 99, for the fine ceilings and the SME rule in paragraph 6: https://artificialintelligenceact.eu/article/99/

Latham & Watkins on the Omnibus — confirms Article 50's scope was left alone, and sets out the 2 December 2026 grandfathering for generative systems already on the market: https://www.lw.com/en/insights/ai-act-update-eu-resolves-to-change-rules-and-extend-deadlines

The Commission's page on the Code of Practice — published 10 June, voluntary, still under adequacy assessment by the Commission and the AI Board. It covers Article 50(2), (4) and (5): https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content

Stibbe, on the provider/deployer split. Worth reading, but note it describes signatories as receiving a degree of presumption of conformity, which reads stronger than the Commission's own conditional wording — I went with the Commission: https://www.stibbe.com/publications-and-insights/the-ai-acts-transparency-obligations-rules-scope-and-timeline

One point I cut for length: the FAQ is explicit that a deployer cannot discharge its own 50(4) duty by pointing at the provider's machine-readable mark. For deepfakes and AI-generated text published on matters of public interest, the label has to be perceivable by a human without any specific technical tools. If you are both provider and deployer — which most agent shops are — that is two separate artefacts, not one.

Claims

Sources