The EU's AI Labelling Rules Are Live — and Widely Misread

Eleven days ago the EU began requiring AI-generated content to be labelled. Most of the coverage since has repeated the same wrong thing: that everything an AI produces now needs a disclaimer. That is not what Article 50 says — and the gap between the two readings is exactly where the compliance work sits.

Two duties, not one

Article 50 of the EU AI Act (Regulation (EU) 2024/1689) applied from 2 August 2026. It creates two separate obligations that get collapsed into one in most summaries.

The first is machine-readable marking, and it binds providers — whoever builds and ships the system. Under Article 50(2), an AI system generating synthetic audio, image, video or text must mark its output in a machine-readable format so it is detectable as artificially generated or manipulated. This is a technical signal — watermarking, embedded metadata, provenance data of the C2PA variety — not a caption. A reader may never see it. For generative systems this duty is close to universal.

The second is visible disclosure, and it binds deployers — whoever publishes. Under Article 50(4), a deployer publishing a deepfake must disclose that the content is artificially generated or manipulated, and a deployer publishing AI-generated text to inform the public on matters of public interest must say so. This duty is narrow and targeted.

Two further rules sit alongside them. Article 50(1) requires providers to build systems that interact directly with people — chatbots, voice agents — so users are told they are dealing with an AI, unless that is obvious to a reasonably well-informed, observant and circumspect person. Article 50(3) requires deployers of emotion recognition or biometric categorisation systems to inform the people exposed to them.

So the shape is: near-universal invisible marking, plus targeted visible labelling. "Put a disclaimer on everything we generate" over-complies on one axis while quietly under-complying on the other — because a visible caption is not a machine-readable mark.

The exemptions are doing real work

• Assistive editing. Systems performing only assistive editing, or not substantially altering the input data or its semantics, fall outside the 50(2) marking duty. Spellcheck and grammar correction do not trigger it. Nor do short sequences, source code, or machine-to-machine output.
• Human review and editorial responsibility. AI-generated text is outside 50(4) where it has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publication. The carve-out asks for substantive review, not a rubber stamp.
• Artistic and satirical work. Where a deepfake forms part of an evidently creative, satirical or fictional work, disclosure is made in an appropriate manner that does not hamper the display or enjoyment of the work — a corner credit, not a watermark across the frame.
• Law enforcement. Systems lawfully authorised to detect, prevent, investigate or prosecute criminal offences are carved out of each limb.

Who is affected

Anyone placing a generative AI system on the EU market, regardless of where they are established. On the deployer side: newsrooms, agencies, platforms and public bodies publishing AI-drafted text on matters of public interest. Note the trigger there is the purpose of publication — informing the public — not the subject matter, so a corporate blog explaining a policy change can land inside it while an AI-written product description does not.

Timing splits by when the system reached the market, and this is where the Digital Omnibus matters. The Omnibus is best known for pushing the high-risk (Annex III) deadlines out into 2027 and 2028, and it left Article 50's substance and its 2 August application date alone. But it did grant a four-month transitional period on one specific limb: generative systems placed on the EU market before 2 August 2026 have until 2 December 2026 to meet the 50(2) marking and detection duty. Anything placed on the market on or after 2 August 2026 got no transition at all.

Note how narrow that relief is. It softens the provider marking duty and nothing else — Article 50(4) deployer disclosure took effect on 2 August 2026 for everyone, with no transitional period for anybody. So the two duties do not merely differ in kind; they started on different dates. Anyone who read "the Omnibus delayed the AI Act" and assumed it bought them time to label their published output has been non-compliant since the second of the month. Breaches carry fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher, with proportionality for smaller firms.

Three things to do this week

1. Split the inventory by role, not by system. Marking duties follow the builder; disclosure duties follow the publisher. The same company is frequently both, for different products, with the obligations landing on different teams.
2. Verify the mark survives your pipeline. Check that your generation stack actually emits a machine-readable mark — then check it is still there after re-encoding, resizing, cropping, screenshotting and CMS ingestion. A watermark that dies in post-processing is not compliance, and this is where most implementations quietly fail.
3. Settle your Article 50(4) editorial position. If you publish AI-drafted text on public-interest matters, either label it or document the human review and named editorial responsibility that exempts it. Decide once, in writing, rather than per article.

Related context

The Code of Practice on Transparency of AI-generated Content, published on 10 June 2026, is the practical route through this. The Commission and the AI Board concluded on 8 and 9 July 2026 that it is adequate to support implementation of Article 50(2), (4) and (5), and the first signatories were listed ahead of 2 August. Signing is voluntary, and not signing is not itself non-compliance — but signatories can rely on adherence to demonstrate compliance, while everyone else has to build that evidence from scratch.

The parent instrument is the EU AI Act itself, whose GPAI transparency duties under Article 53 run on a separate track and are not satisfied by Article 50 compliance.

Note: this article was drafted by AI (Claude, Opus 5) and reviewed by a human editor before publication — which is the Article 50(4) question in miniature. AI-generated text published to inform the public on a matter of public interest is disclosable unless a person takes editorial responsibility for it. We disclose either way.