A newsroom's survival guide to the EU AI Act's Article 50 transparency rules

Regulation explainers might not be your idea of a fun read, so I’ll try to keep this one short. Fines for getting it wrong run up to €15 million, or 3% of worldwide annual turnover, so it should be worth ten minutes of your time.

Full disclosure: as a consumer, I’m rather thrilled about this one. Knowing what I’m reading was actually written by a person feels like a small win, and I’d guess some of your readers feel the same way.

Two duties, split between us and you

Good news first: Article 50 splits the work in two, and one half was never yours to carry. Whoever builds the AI tool handles one duty, whoever publishes with it handles the other.

Machine-readable marking is the AI provider's duty. When you use AI writing tools, the output has to be marked as AI-generated in a way computers can detect. That’s a provider obligation under Article 50(2), and we carry it. It isn’t optional on our end either: the marking ships enabled and stays on.

Visible labels are the publisher’s duty. Article 50(4) puts the publisher (or the “deployer” in the regulation’s rather bureaucratic vocabulary) in charge of what readers actually see. Two things trigger it: deepfakes, and AI-generated text “published with the purpose of informing the public on matters of public interest” — also known as “News”.

Know what actually needs a label

We went through the likely newsroom scenarios one by one. The pattern that came out the other end is pretty consistent:

Label required

• AI-written text on public-interest matters, published without human review

• Photorealistic AI images that could pass as real

• Evidently satirical or artistic AI work (discreetly; a caption or credits note, not a banner)

No label needed

• An AI draft substantively edited by a journalist

• AI titles and summaries edited by a journalist

• Spellcheck, light retouching (assistive tooling, not generation)

• Stylised or abstract AI illustration (it’s not a deepfake)

Edit like a journalist, and you’re mostly covered

If a journalist genuinely reviews and edits AI-drafted text, and an editor holds editorial responsibility for the piece, no reader-facing label is required. “Genuinely” is doing some work in that sentence: the European Commission’s guidance is explicit that spellchecking and a quick skim don’t count.

Sounds a lot like normal editing, doesn’t it? That’s because it basically is. The law isn’t asking you to change how you work, just to be able to show that a person actually did the work.

At the end of the day, publishers who keep a human in the loop should feel very little of Article 50. Publishers who auto-publish AI text carry all of it.

One catch worth flagging now, while we’re on the subject: you may someday need to prove the human review happened, not just that it could have. Which is exactly what we’re building next.

Coming soon to Labrador

Rolling out ahead of the marking deadline (2 December 2026):

• Machine-readable marking on AI-generated text and images across published pages: summaries, captions, alt text, meta tags, all of it. On by default, and we manage it, so this one’s off your plate entirely.

• An “AI image” flag on every image. It feeds the invisible marking, and, if you want it, a visible reader label you can style yourself.

• Reader disclosures for AI summaries and AI-assisted articles. Set a site-wide policy, hand editors per-article control, or make it automatic whenever AI touched the piece. You write the copy, we handle where and when it shows up.

• Review logging. When a journalist edits AI-drafted text, we log it quietly in the background. That’s the audit trail that backs up your editorial exemption if a regulator ever comes asking.

• AI transparency settings per site, so publishers outside the EU can turn reader-facing labels off entirely. No need to carry rules that were never written for you.

Quick answers

Does an AI draft that a human rewrote need a label?

No, that’s the editorial exemption, as long as the review was substantive and someone holds editorial responsibility.

Do AI illustrations need a label?

Only if they could pass as real people, places, or events. A clearly stylised graphic isn’t a deepfake and needs no visible label. The invisible marking still applies either way, quietly, in the background.

Do AI-translated articles need their own review?

Yes. Reviewing the original doesn't cover the translation, they're treated as separate content. If you publish in multiple languages, each translated version needs its own substantive human check to qualify for the editorial exemption.

What about satire?

Still disclosed, but discreetly, in a way that doesn't ruin the joke - yes, really! The regulators qualify it as "an appropriate manner that does not hamper the display or enjoyment of the work.".

We publish outside the EU. Can we turn this off?

Yes, reader-facing labels are a per-site setting you control. Machine-readable marking keeps running quietly in the background wherever you publish, so it's one less thing you need to manage.

 ­

This is an overview, not legal advice. Primary sources: Article 50, Regulation (EU) 2024/1689, the European Commission’s FAQ on Article 50, and the Commission’s Guidelines on transparency obligations of 20 July 2026.

Questions about your setup? Talk to your customer success manager, or book a 20-minute walkthrough.