Lost in the noise around the Digital Omnibus deferral is a date that did not move: 2 August 2026, when the transparency obligations of Article 50 of the EU AI Act become applicable.

These obligations were never part of the high-risk regime that got pushed to December 2027. They sit in their own category, they apply broadly, and — unlike much of the AI Act — they land on ordinary product teams shipping ordinary features. In June 2026 the European Commission detailed how it expects them to work in practice. If you generate content, run a chatbot, or publish AI-assisted material, this is the obligation most likely to apply to you, and the one you have the least time to meet.

What Article 50 actually requires

Four duties matter for most companies:

  • Interacting with an AI system. If a person is talking to an AI — a support bot, a voice agent, a chat interface — they must be told, clearly, unless it is obvious from context. Disclosure has to reach the user at the moment they encounter it, not buried in terms and conditions or a secondary settings panel.
  • AI-generated media (deepfakes). Deployers who use AI to create or manipulate image, audio, or video content that resembles real people, places, or events — content that would falsely appear authentic — must disclose that it is artificial.
  • AI-generated text published to inform the public. This must be disclosed unless it has undergone genuine human review and a named person or organisation takes editorial responsibility for it.
  • Marking at source. Providers of generative systems must mark outputs as artificially generated in a machine-readable format — detectable by other systems, not just visible to a human.

The details that trip teams up

The Commission’s guidance is more specific than most teams expect, and the specifics are where compliance is won or lost.

Disclosure has to be legible, not technically-present. A watermark no one can see, or a disclaimer three clicks deep, does not discharge the obligation. The guidance points toward a uniform, recognisable “AI” visual cue and clear placement — the user should understand what they are looking at without hunting for it.

Human review is a real carve-out — if it is real. AI-drafted text escapes the disclosure duty only where a natural or legal person has genuinely reviewed it and assumes editorial responsibility. A rubber-stamp does not count. This is an accountability test, not a checkbox.

Creative and satirical work is treated proportionately. Where content is evidently artistic, fictional, or satirical, the obligation narrows to disclosing its existence in a way that does not spoil the work — you do not have to stamp a film frame-by-frame.

The high-risk regime is about systems that decide about people. Article 50 is about honesty — telling people when they are looking at, or talking to, a machine. It is narrower, simpler, and due first.

Transparency obligations fail in the same place every time: they are treated as a legal review at the end of the pipeline instead of a design decision at the start. By the time a launch is in front of counsel, the disclosure surface, the labelling format, and the machine-readable marking are already baked — or already missing.

The teams that will be ready in August are the ones treating Article 50 as three product requirements:

  1. A disclosure layer in every AI-mediated interaction, surfaced at first contact.
  2. A labelling standard for generated media — consistent placement, a recognisable cue, and provenance metadata that travels with the asset.
  3. An editorial-ownership process for any AI-assisted text you publish, so the human-review carve-out is genuinely available to you.

None of this is heavy. But all of it has to exist before the feature ships, which means the decision to build it belongs in the sprint, not in the pre-launch legal pass.

The deferral bought most companies time on the hard, expensive, high-risk work. It bought no time at all on this. August is close, and transparency is the obligation you cannot retrofit quietly.

If you are shipping AI features and are not sure which of them trip Article 50, that is exactly the kind of scoping we do in a first call.