EU AI Act Article 50: What Actually Changes on 2 August 2026
From 2 August 2026, the EU AI Act starts requiring AI-generated content to carry a machine-readable mark. A lot of vendors are about to tell you this means you need their product. It mostly does not. Here is what Article 50 obliges, who it obliges, and where detection genuinely still matters.
Last updated July 27, 2026
What Article 50 requires
Article 50 of Regulation (EU) 2024/1689 — the AI Act — sets transparency obligations for certain AI systems. It applies from 2 August 2026, and breaches carry fines of up to EUR 15 million or 3% of total worldwide annual turnover.
Four duties become enforceable on that date. Providers of AI systems that interact directly with people must make it clear the user is talking to an AI, unless that is already obvious. Providers of generative AI must mark synthetic audio, image, video and text output in a machine-readable format so it is detectable as artificially generated. Deployers of emotion recognition or biometric categorisation systems must tell the people exposed to them. And deployers must disclose deepfakes, plus AI-generated text published to inform the public on matters of public interest.
The marking obligation in Article 50(2) is the one that matters most for written content, and it is the one most often described inaccurately.
Who the obligation actually falls on
The AI Act separates providers from deployers, and this distinction decides whether Article 50 applies to you at all.
A provider develops an AI system, or has one developed, and places it on the EU market under its own name or trademark. For generative text, that means the companies building the models — OpenAI, Anthropic, Google, Mistral and their peers. Article 50(2) puts the machine-readable marking duty on them, not on the people who use their output.
A deployer uses an AI system under its own authority in a professional capacity. If you are a company, the organisation is the deployer, not the individual employee. Deployers have a narrower duty: disclose deepfakes, and disclose AI-generated text published to inform the public on matters of public interest.
A student submitting an essay is neither. A teacher marking one is neither. Neither of them acquires an Article 50 obligation on 2 August 2026.
The exemption that removes most written content
The deployer disclosure duty for text carries a significant carve-out: it does not apply where the content has undergone human review or editorial control, and where a natural or legal person holds editorial responsibility for its publication.
The Commission has been explicit that this means substantive review. Running a spell-checker over AI output does not qualify. An editor who reads, revises and takes responsibility for the piece does.
Several other things fall outside the marking obligation entirely: short sequences of numbers, symbols or letters; source code; machine-to-machine output; output from standard assistive editing functions that do not substantially alter the input; and certain closed-loop industrial uses. Assistive editing was carved out deliberately — using AI to tighten a sentence you wrote is not the same as generating the sentence.
What the Digital Omnibus changed — and what it did not
The AI Omnibus package created real confusion in mid-2026, because it deferred a lot of the AI Act. It did not defer Article 50.
Annex III high-risk obligations moved to 2 December 2027. Article 50 stayed on 2 August 2026, with full enforcement powers activating on the same day.
There is one narrow transitional allowance: generative AI systems already placed on the market before 2 August 2026 get until 2 December 2026 to meet the Article 50(2) machine-readable marking requirement specifically. Systems launched on or after 2 August 2026 must mark from day one, with no transitional relief. Content generated before 2 August 2026 does not have to be labelled retroactively.
What this does not require you to do
Article 50 does not require you to run an AI detector. It does not require you to certify content as human-written. It does not create a compliance badge, and no third party can make you compliant on your own behalf — the obligations sit with providers and deployers of AI systems, defined by the roles above.
We want to be direct about this because we sell a detection product and it would be easy to imply otherwise. If a vendor tells you that Article 50 means you need to buy AI detection, ask them which paragraph. There is not one.
It is also worth saying plainly: marking under Article 50 and detection are different mechanisms. Marking is provenance applied at the source by the model provider. Detection is statistical inference applied afterwards by a third party. The regulation is building the first. It says nothing about mandating the second.
Where detection still genuinely matters
Article 50 improves provenance for content produced by compliant providers, going forward. It does nothing about three situations that are already the ones causing people real problems.
First, unmarked output. Models outside EU jurisdiction, self-hosted open-weight models, and anything generated before 2 August 2026 carry no mark. Marking is also required only as far as it is technically feasible, effective and robust — and text watermarking is the least robust modality, because paraphrasing degrades it.
Second, the accusation problem. The situation that actually harms people is not undetected AI text — it is a human writer being told their own work is AI. A machine-readable mark cannot help there, because genuinely human text has no mark to show. The absence of a mark proves nothing.
Third, disagreement between detectors. Detectors contradict each other on a meaningful share of texts, and their errors fall hardest on non-native English writers. That is the specific gap we built for: running several independent engines and reporting both the consensus and the disagreement, so an ambiguous text looks ambiguous instead of looking certain.