What Article 50 actually says

This guide is not legal advice. It is a clause-by-clause walk through the actual text of Article 50 of the EU's AI Act, translated into the language of people who run websites. Where we quote, we quote the Act. Where we interpret, we say so.

Why read the source at all

Most of what circulates about the AI disclosure rules is paraphrase of paraphrase, and each layer sands off a nuance that mattered. Article 50 is seven paragraphs long. You can read the important parts in ten minutes, and afterwards you will know which of the confident claims in your feed are true, which are half-true, and which are somebody's newsletter trying to frighten you. Here is the whole article, one paragraph at a time.

Paragraph 1: tell people when they're talking to a machine

The Act requires that AI systems "intended to interact directly with natural persons" are built so that those people "are informed that they are interacting with an AI system, unless this is obvious".

Translation: chatbots must say they are chatbots. This is a duty on the provider of the system, so your chat widget vendor carries the design obligation, but the honest reading for a website owner is simpler: check that your chatbot introduces itself as one. The "unless this is obvious" carve-out covers the robot that is plainly a robot; it does not cover the support widget with a human name and a stock photograph.

Paragraph 2: the machine-readable watermark

Providers of AI systems that generate "synthetic audio, image, video or text content" must ensure outputs "are marked in a machine-readable format and detectable as artificially generated or manipulated".

Translation: the AI companies must watermark what their models produce, at the file level, for software to detect. Note who this binds: the provider, not you. And note what it is: machine-readable. This paragraph is why metadata credentials exist, and it is also why metadata alone never satisfies the duties later in the article, which are about what humans see. Two different layers, two different audiences.

Paragraph 3: emotion recognition and biometric systems

Deployers of emotion recognition or biometric categorisation systems must inform the people exposed to them. If your website does not analyse anyone's face or infer anyone's feelings, this paragraph is not your problem, and we mention it only so you know nothing relevant is being skipped.

Paragraph 4, first limb: deep fakes

Here is the paragraph that matters most for imagery. Deployers of an AI system "that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated".

A deep fake, in the Act's usage, is AI content resembling real persons, objects, places or events that would falsely appear authentic. That definition comfortably covers the generated team photo and the rendered premises shot, not just the celebrity face-swap. The duty sits on the deployer, which for content on your website means you.

There is one softening: where the content is "evidently artistic, creative, satirical, fictional" or analogous, the duty narrows to disclosing the generated content's existence "in an appropriate manner that does not hamper the display or enjoyment" of the work. Honesty without a warning sticker. The operative word is "evidently", and it is judged from the visitor's chair, not yours.

What you will not find anywhere in this limb: an exception for human review. There is none. Images cannot be reviewed out of the duty. Our images guide exists because of this sentence.

Paragraph 4, second limb: AI text that informs the public

Deployers of an AI system "that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated".

Two boundaries are built into that sentence. The text must be published to inform the public, and on matters of public interest. Your AI-drafted guide to tenants' rights is inside both. Your own company describing itself is at the outer edge, which is why our own advice is that self-description carries lower risk but a one-line disclosure still buys quiet.

Then comes the most practical exception in the whole Act. The duty does not apply "where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication".

Read the load-bearing words. "Has undergone a process": something that actually happened, with steps, not a vibe of having been glanced at. "Human review or editorial control": a person engaged with the content. "Holds editorial responsibility": someone identifiable stands behind it. Meet all three and your reviewed AI text needs no label. The unstated fourth requirement is the practical one: when anyone asks whether the process happened, you need something better than your word. A review with no record is, evidentially speaking, no review. Readers of the rest of this site will recognise that sentence as our entire business model, so weigh our enthusiasm accordingly, and then read the paragraph yourself and see if we have it wrong.

Paragraph 5: how and when the disclosure is made

All the information required by paragraphs 1 to 4 "shall be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure", and must meet accessibility requirements.

Translation: the disclosure has to be somewhere the person actually encounters, by the time they encounter the content. For a web page, that is on the page, with the content, visible without archaeology. This single sentence is what disqualifies the footer disclaimer, the hidden policy page and the metadata-only approach in one pass.

Paragraphs 6 and 7: the plumbing

Paragraph 6 says these transparency duties sit alongside, and do not replace, other obligations in the Act and in other Union or national law. Paragraph 7 tasks the AI Office with encouraging codes of practice for detecting and labelling generated content, which the Commission can approve or, failing that, replace with implementing rules. That machinery is already turning: the Commission published final guidelines on these obligations in July 2026, non-binding but the reference document the enforcement authorities work from, and they settle, among other things, that content published before 2 August 2026 needs no retroactive labelling. Translation: the detail of how labelling is done in practice will keep being refined at EU level, and the sensible posture is to follow the codes and guidance as they emerge rather than invent private theology.

What to take away

Article 50 is shorter and more reasonable than its reputation. Chatbots must say what they are. AI companies must watermark at the file level. Realistic AI imagery must be visibly disclosed, with no review escape. AI text informing the public must be disclosed, unless a real person really reviewed it and takes responsibility, provably. And every disclosure must be somewhere humans actually look, by the time they look.

If you want that checked against your own site rather than in the abstract, the scan is free and takes about a minute: swornmark.com. And the fines for getting it wrong, since you will ask, are in Article 99: up to €15m or 3% of worldwide turnover, with smaller companies fined at the lower of the two.

Questions people actually ask

Is there an official EU label or wording we have to use for AI content? No. Article 50 asks for disclosure in a clear and distinguishable manner and mandates no formula, badge or logo. Anyone selling you the officially approved sentence is improvising, because there is none. Codes of practice may standardise the mechanics over time, which is a reason to write plainly now and follow the guidance as it firms up.

Which parts of Article 50 actually apply to a normal business website? Usually three. Paragraph 1 if you run a chatbot, which must say it is one. Paragraph 4 for content: realistic AI imagery needs a visible disclosure, and AI text that informs the public needs one unless a real person reviewed it and takes editorial responsibility. And paragraph 5 governs placement for all of it: the disclosure has to be met by the time the content is. Emotion recognition and biometrics, in paragraph 3, are not most websites' problem.

What counts as a deep fake under the AI Act? AI-generated or manipulated content resembling real persons, objects, places or events that would falsely appear authentic. That comfortably covers the generated team photo and the rendered premises shot, not just the celebrity face swap. Content that is evidently artistic or satirical gets a softer duty, but "evidently" is judged from the visitor's chair, not yours.

Does human review exempt us from labelling AI content? For text, yes, and it is the most practical exception in the Act: a process of human review or editorial control, with a natural or legal person holding editorial responsibility. The words doing the work are "has undergone a process", which is something that happened and left a record, not a vibe of having been glanced at. For images there is no review exception at all.

Is Article 50 actually in force, and does it cover our old content? The duties have applied directly since 2 August 2026, whether or not your country has finished appointing its enforcement authority. The Commission's guidance says content published before that date needs no retroactive labelling, but the exemption only protects what you leave alone: update or republish an old page and it comes into scope as a fresh publication.

---

This guide is not legal advice, and where our reading and the Act's text disagree, the Act wins. It was drafted with AI assistance and reviewed before publication. Swornmark holds editorial responsibility for it.