EU AI Act Text Labeling Rules: Who Article 50 Actually Covers

Somewhere in the last month you probably saw a claim that the EU now requires every AI-assisted email to carry a disclosure label. It does not, and the actual rule is both narrower and more interesting than that panic suggests.

Article 50 of the EU AI Act sets out transparency obligations that apply from August 2, 2026, and the European Commission finalized its guidelines on how to read them on July 20, 2026, less than two weeks before the deadline. The rules do real work. They just do it in a much smaller space than the rumor mill suggests.

What Article 50 Actually Requires

Article 50 has several distinct duties bundled under one article, and mixing them up is where most of the confusion starts.

Providers of generative AI systems must mark the content their system generates so it is technically detectable as AI-made, a machine-readable watermark or metadata tag rather than a visible label.

Deployers of deepfake systems must clearly disclose that content has been artificially generated or manipulated.

Deployers of generative AI must clearly label AI-generated or manipulated text when it is published to inform the public on matters of public interest.

Providers of chatbots and similar interactive systems must disclose, at the first interaction, that a person is talking to an AI rather than a human.

Content generated before August 2, 2026 does not need to be retroactively labeled. The Commission has also published optional standardized icons providers and deployers can use, though using them is not itself mandatory.

The Phrase That Decides Whether This Touches You

Read the text labeling clause again, slowly: “text which is published with the purpose of informing the public on matters of public interest.”

That is a narrow, specific category. It is aimed at content that functions like journalism or public communication, published outward, addressed to a general audience, on a matter the public has an interest in understanding accurately. Think a news-style article, a public policy explainer, an official public-facing statement.

A routine work email is not that. A Slack message to a colleague is not that. An internal memo, a client proposal, a performance review draft, a sales follow-up: none of these are text published to inform the public on matters of public interest, even when you used an AI tool to help write them. They are private, addressed communication, not public information content. The labeling duty for text, specifically, was built for a different target than ordinary business writing, and the guidelines finalized in July 2026 did not expand that target; they clarified it.

This is the single most useful thing to internalize from Article 50 if you are not a publisher, a newsroom, or a public communications office: the duty almost certainly does not reach your inbox.

Who Is Actually Caught

The people who need to build a labeling process are narrower than the coverage implies:

News and media organizations publishing AI-assisted articles to inform readers on public-interest topics.

Public sector bodies and institutions issuing public-facing AI-generated communications.

Organizations running deepfake or synthetic-media content meant for public consumption.

Providers building generative AI systems, who carry the machine-readable marking duty regardless of what the content is used for.

Companies operating a public-facing chatbot, who owe the first-interaction disclosure regardless of subject matter.

If none of those describe what you do, Article 50’s text-labeling clause is background information, not a compliance task. It is still worth understanding, since the broader landscape of who has to disclose what is expanding fast; see AI transparency and disclosure obligations in 2026 for the fuller map across jurisdictions.

What To Actually Do With A Rewritten Draft

Even where the law does not require a label, plenty of workplaces have their own norms about saying an email was AI-assisted, and that is a judgment call worth making deliberately rather than by accident. We covered the reasoning in whether you should disclose you used AI to write an email.

Before:

This message was generated using artificial intelligence in accordance with EU transparency requirements.

After:

Quick note before the meeting: I’ve moved the budget review to Thursday at 2pm so finance has the updated numbers in hand. Let me know if that conflicts with anything on your end.

The first version drags a public-interest-content disclosure into a routine internal message where it does not apply and adds nothing useful. The second just says what needs saying. If your workplace does want AI-use disclosed internally, do it as a plain statement in your own words, not as a legal notice borrowed from a law aimed at a different kind of content.

A Wrivio Context for checking whether something needs an Article 50 label could say:

Before rewriting, tell me plainly whether this text is being published to inform the public on a matter of public interest, versus being sent privately to a named recipient. Keep every name, date, figure, and commitment exactly as written. Do not add any AI-disclosure language unless I explicitly ask for it.

Press Ctrl+Shift+Space, paste the draft, and check the diff. What you are watching for is a model that adds a disclosure line on its own initiative because the topic sounds regulatory; that instinct is usually wrong for private, addressed correspondence.

Common Questions

Do I have to label an email I wrote with AI help under the EU AI Act?

No, in almost every case. Article 50’s text-labeling duty applies to AI-generated text published to inform the public on matters of public interest, and a private email to a named recipient does not meet that description.

When did Article 50 start applying?

The transparency obligations broadly apply from August 2, 2026, after the Commission finalized its implementing guidelines on July 20, 2026.

Does content I generated before August 2026 need to be labeled retroactively?

No, the guidelines specifically exclude content generated before that date from a retroactive labeling requirement.

Who actually has to label AI-generated text?

Deployers publishing AI-generated or manipulated text to inform the public on matters of public interest, such as news-style content or public communications, plus separate duties for chatbot operators and deepfake deployers.

Where can I read the rule itself?

The official Article 50 text and the Commission’s FAQ on transparency obligations are the primary sources.

Download Wrivio for Windows to draft the routine work correspondence this rule was never aimed at, with a clear diff showing exactly what changed.