Article 50 Applies Today, and the Disclosure I Didn't Ship

The EU AI Act’s Article 50 transparency obligations apply from today. This post covers what the text actually separates out, why I decided the exemption covered this blog rather than putting a banner on it, and where our own coverage stops.

Article 50 is not one obligation

Article 50 gets discussed as a single popup. The text splits it into duties with different triggers and different natural homes.

Art 50(1), interaction disclosure. Where an AI system is intended to interact directly with people, those people are to be informed they are interacting with an AI system, unless that is already obvious to a reasonably well-informed person. The duty attaches to a moment, and it is discharged by showing a notice at the point of interaction.

Art 50(2), synthetic content marking. AI-generated or manipulated content is to be marked in a machine-readable format so that it can be detected as artificially generated. This duty attaches to an artefact, and it is discharged by the system that produced the artefact.

Art 50(4), published text and deep fakes. Where AI-generated text is published to inform the public on matters of public interest, the fact that it was artificially generated is to be disclosed. It carries an exemption where the content has been through human review or editorial control and a natural or legal person holds editorial responsibility for its publication.

Three duties, discharged at three different points in a system. A sentence like “AI transparency is the content team’s problem” can be true of one and false of another. That is how a site ends up disclosing on surfaces the text never reached, and missing the one where it did.

The exemption I applied here

This blog is written with AI assistance. There is no disclosure banner on it.

My reading was that the 50(4) editorial-control exemption covers it. These posts are edited and published under a named author, and the byline is doing the work the exemption describes. That was a call I made about my own site rather than a rule anyone should copy off a blog, and where a position actually turns on it the answer comes from a lawyer.

The exemption is easy to get almost right, so it’s worth being precise. It asks for editorial responsibility held by a natural or legal person. A named author satisfies that. A house byline, whether that’s a team name or a product name, is thinner, and on a one-person studio it isn’t even accurate. If you’re going to lean on the exemption, the byline has to name someone.

Over-disclosure carries its own cost. An affirmative claim about what a system does is a claim you then have to keep being right about, which can be harder to hold than a position that stayed inside an exemption in the first place. Working out where a disclosure is not required takes as much reading as working out where it is.

Content versus mechanism

When a disclosure is required, the usual approach is to write the sentence into the page. That puts words in front of someone. It leaves nothing behind.

A disclosure that holds up under review behaves more like a gate. It fires before the interaction, records that it was shown, and records which language it was shown in. Months later there is then an answer to what was disclosed, to whom, and when.

That’s a mechanism problem rather than a content problem, and it’s why 50(1) sits differently from 50(2). Detecting a visitor’s first relevant interaction, suppressing a notice already shown in this session, localising it, targeting only the paths where the AI system actually runs, and logging the impression: none of that list is specific to AI disclosure. It describes a consent layer. The full argument is set out separately in Content vs Mechanism: Why Article 50(1) Belongs in the Consent Layer.

What shipped, and what didn’t

The AI disclosure gate in Nuvo Consent covers Article 50(1) only. It renders the notice in either a passive or an acknowledge mode, records each impression against the gate and domain with a timestamp, exports that log per domain, and surfaces the 50(1) exemption test rather than resolving it for you.

It does not do 50(2) machine-readable content marking, and it does not cover the 50(4) duties this post opened with. That’s a scope decision rather than a gap: both attach to artefacts at the point of production, and a page-level script is the wrong place to discharge them.

It supports the obligation. Discharging it stays with you.

→ Scan your site: consent.nuvocode.com

Still open

50(1) turns on whether the AI interaction is obvious to a reasonably well-informed person. That threshold has no test I can implement. A chat widget on a support page reads as obvious to the people who build them and not always to the people using them, and nothing in the text tells you where the line sits.

So the gate records what you decided. It can’t decide it, and I’d rather say that than ship something implying otherwise.