How to label AI content under Article 50, and which part of it is not your job

Start by separating two roles, because almost every confused conversation about this rule is two people using one word for both. The AI Act calls the party that develops a system and puts it on the market a provider, and the party that uses that system in its own work a deployer. If you build a site that calls a model over an API, you are a deployer, and the heaviest technical duty in Article 50 is not aimed at you. If you train or wrap a generative system and put it out under your own name, you have become a provider, and it is.

Step one: work out which paragraph you are in

Article 50 is four obligations, not one, and they land on different parties. Paragraph 1 binds providers of systems that interact with people: tell the person they are dealing with an AI system. Paragraph 2 binds providers of systems that generate synthetic audio, image, video or text: mark the output in a machine-readable format so it can be detected as artificially generated. Paragraph 3 binds deployers of emotion recognition and biometric categorization: inform the people exposed to it. Paragraph 4 binds deployers who publish: disclose deepfakes, and disclose AI-generated text published to inform the public on matters of public interest.

None of this depends on the system being high-risk. Article 50 applies to any AI system used in one of those situations, and it applies to anyone serving people in the Union regardless of where the company sits.

For the ordinary case — a site with an assistant and some published text, running on a model somebody else provides — you are in paragraphs 1, 4 and 5, and paragraph 2 is your vendor's problem. Check that your vendor actually does it, because you are choosing a supplier whose compliance becomes visible on your pages, but do not build a marking pipeline you were never asked for.

Step two: make the assistant say what it is, at first contact

If a person on your site is talking to a model, they have to be told, and paragraph 5 fixes the moment: at the latest at the time of the first interaction. Not in the privacy policy, not on hover, not after the third message.

There is one exemption and it is narrower than people hope. The duty falls away only where it is obvious to a reasonably well-informed and observant person, from the circumstances and the context, that they are dealing with an AI system. A widget in the corner of a shop page is not obvious. A page titled AI assistant, with a first message that says so in words, is.

The Commission's guidance is explicit that the disclosure has to be clear and distinguishable, which is where the sparkle icon dies. An icon is a decoration in a design system, it carries no text for a screen reader, and it means four different things across four sites your visitor used this week. Write the sentence.

Step three: for published text, choose a name or a label

Paragraph 4 covers text that is published to inform the public on matters of public interest — health, safety, environment, economy, finance, politics, science, culture — and was generated or manipulated by AI. For that text you disclose, unless the content went through human review or editorial control and a natural or legal person holds editorial responsibility for it.

Read that exemption as the instruction it is. The regulation is not asking whether a model touched the text. It is asking whether a person stands behind it. So there are exactly two compliant answers: put a named human in the loop and record that they took responsibility, or say plainly that the text is machine-written. What is not available is publishing it unsigned and unlabeled and hoping the topic counts as trivial.

This site takes the first road and shows its work. Every piece stores who wrote it, which model, which pass, and whether an editor has reviewed it yet, and the byline is a person, not a brand. The point is not that our arrangement is the only correct one. The point is that the answer has to be a record you can produce, not a claim you make afterwards.

Two things are outside this. Content that forms part of an evidently artistic, creative, satirical or fictional work is exempt, disclosed instead in a way that does not spoil the work. And a system performing an assistive function for standard editing, which does not substantially alter the input data or its meaning, is outside paragraph 2 altogether — the guidance treats spellcheck, grammar, formatting, cropping and color correction as that kind of function.

Step four: a deepfake gets disclosed, and that is the whole rule

If you publish image, audio or video that has been generated or manipulated to resemble a real person, object, place or event, and would falsely appear authentic, you say so. The test is resemblance to something real plus the appearance of authenticity, not whether a model was involved: a rendered illustration of nothing in particular is not a deepfake, and a photograph of a real politician with the words changed is.

The artistic, creative, satirical and fictional carve-out applies here too, and it changes how you disclose rather than whether you do.

Step five: build a label that survives being reshared

A label that only exists as a line of body text stops existing the moment someone downloads the image or reshares the card. The guidance asks for disclosure that stays attached to the content and reaches assistive technology, which in practice means three layers rather than one.

In the page, put the disclosure in text near the byline, inside the article markup, where a screen reader reaches it in reading order. In the data, keep the fact as a field — who wrote it, which model, who reviewed it — so it can be rendered anywhere, exported, and answered for later. In the file, keep whatever provenance metadata your generator writes and do not strip it in your image pipeline, which is the single most common way a compliant asset becomes a non-compliant one: a resize step that drops metadata undoes the provider's paragraph 2 work on your server.

Then check the boring things. The label must be readable in both themes, must not be hidden behind a hover state, and must survive the card that a social platform renders from your page.

Step six: use the Code of Practice as a ready-made checklist

The Commission published the final Code of Practice on Transparency of AI-generated Content on June 10, 2026, after a drafting process that started in November 2025. It has one section for providers, on marking and detection, and one for deployers, on labeling deepfakes and certain AI-generated text. It covers Article 50(2), (4) and (5). By the end of July 2026 roughly 190 companies and organizations had signed it.

Signing is voluntary and the Commission and the AI Board have confirmed the code is an adequate voluntary tool for demonstrating compliance. Read the practical consequence carefully: the obligations are law whether or not you sign, and the code is the cheapest way to show that what you built satisfies them, with the same answer accepted across all Member States. If you are looking for the shortest path from this article to an implementation, that document is it.

Step seven: the dates, and the one grace period

Article 50 has applied since August 2, 2026. Today is not a deadline you are preparing for; it passed 19 days ago, and the obligations in steps two through five are live.

There is one extension and it is narrow. Under the AI Omnibus provisional agreement of May 2026, generative AI systems already on the market before August 2, 2026 have until December 2, 2026 to meet the machine-readable marking requirement in paragraph 2. That helps providers of existing systems. It does not extend the deployer duties in paragraph 4, and a provisional agreement is not a final text, so treat it as a reason to ask your vendor where they are rather than as your own extra quarter.

What this guide does not cover

It is not legal advice and does not price the risk. Penalties, national enforcement bodies and how each Member State handles a complaint are outside it, and the carve-outs for law enforcement — present in every paragraph — are outside it too.

It also does not tell you whether your text is on a matter of public interest. That judgment is yours, it is the one place where the rule has real edges, and the honest way to handle a borderline case is to label it and move on. A label on a text that did not need one costs a line. A missing label on one that did is the failure the rule exists to catch.