Do You Have to Label AI Text in the EU? The 2026 Rules

Most people who write with AI in Europe do not have to put a label on anything. The rule that started on 2 August 2026 is much narrower than the headlines made it sound. It is aimed at one kind of writing: text published to inform the public about matters of public interest, with nobody taking editorial responsibility for it.

What Article 50 actually says

Article 50 is the transparency part of the EU AI Act. It is short, and it is split between two groups of people. Providers are the companies that build an AI system and put it on the market. Deployers are the people and organisations that use one in their work. The two groups have different duties, and most of the confusion online comes from mixing them up.

Four of its paragraphs matter for anyone who writes with AI. The table below shows who each one binds and whether it touches written text.

Paragraph 3 covers emotion recognition and biometric sorting, which has nothing to do with writing, so it is left out here.

Who counts as a deployer

The Act defines a deployer as any person, company, public body or other organisation using an AI system under its authority, except where the AI system is used in the course of a personal non professional activity. Article 2(10) repeats the point: the deployer duties do not apply to people using AI for purely personal reasons.

So the first question is simple. Are you using the AI tool for work, for a business, for an organisation or for a public body? If yes, you are a deployer. If you are writing a birthday message, planning a trip or drafting a personal diary, you are not.

Does your text need a label? 12 common cases

The Act does not define a matter of public interest, and the Commission's guidelines are still the main reference. The table below is our reading of the text for common situations. It is not legal advice, and where the answer really depends on the facts it says so.

Two things decide almost every case. The first is purpose: is the text published to inform the public about something that matters to the public, such as politics, health, safety, the economy or public services? The second is responsibility: did a real person read it, and does a person or organisation stand behind it as its publisher?

What counts as human review or editorial control

The exemption in Article 50(4) has two parts, and both are needed. The text has to go through human review or editorial control, and a natural or legal person has to hold editorial responsibility for publishing it. A newsroom where an editor approves every piece meets both. A script that turns press releases into articles and posts them overnight meets neither.

In practice, a review that would stand up if anyone asked looks like this:

• ✦
  A named person reads the whole textNot a skim of the headline. They check the facts and the claims before it goes out.
• ✦
  They can change or stop itReview only counts if the reviewer has the power to edit the piece or refuse to publish it.
• ✦
  Someone is accountableThe publisher, a newsroom or a company, is named and answers for the content.
• ✦
  There is a recordWho approved it and when. A short note in your publishing tool is enough.

What AI tool makers have to do

Article 50(2) puts a separate duty on providers. Any system that generates synthetic audio, images, video or text has to mark its output in a machine readable format so it can be detected as AI made. The Act asks for marking that works, holds up and can be read by other systems, as far as that is technically possible, which in practice means watermarks, metadata or both.

There is an exemption for tools that only assist with standard editing, or that do not substantially alter the input. A spelling checker or a grammar fixer is the obvious example. The Commission published a voluntary Code of Practice on marking and labelling in June 2026 to show providers how to comply, and according to law firm Stibbe, providers whose systems were already on the market before 2 August 2026 have until 2 December 2026 to bring the marking into line.

None of this moves the deployer's duty. If a tool marked your text and you publish it on a public interest topic without review, the label is still yours to add.

Fines and who enforces them

Article 99 sets the ceiling. Breaking the Article 50 obligations can cost up to €15 million or 3% of total worldwide annual turnover for the previous year, whichever is higher. For small and medium sized companies the rule flips to whichever is lower. Giving incorrect or misleading information to an authority is a separate offence, capped at €7.5 million or 1% of turnover. Enforcement sits with the market surveillance authority in each EU country.

The AI Act timeline so far

Content that was generated and published before 2 August 2026 does not need a label after the fact. Content generated before that date but published on or after it does, according to Stibbe's reading of the rules.

A simple checklist if you publish for EU readers

1. 01
   Decide whether the text is public interestPolitics, health, safety, the economy, public services and current affairs are the clear cases. Product copy and marketing usually are not.
2. 02
   Put a real reviewer in the loopSomeone who reads the whole piece, checks it and can stop it. This is the single step that removes the need for a label.
3. 03
   Keep a short approval recordName, date and the piece. It takes seconds and it is your evidence if a regulator ever asks.
4. 04
   Label anything that skips reviewA plain line such as "This article was generated with AI" placed where readers see it first meets the clear and distinguishable test.
5. 05
   Label deep fakes separatelyAny realistic AI image, audio or video of real people, places or events needs its own disclosure, whatever the text says.
6. 06
   Check your own sector's rulesPress codes, health advertising rules and platform policies can ask for more than the AI Act does.

If you are outside the EU

The Act reaches past Europe's borders. Article 2 applies it to providers and deployers based outside the EU where the output of the AI system is used in the Union. A US news site with a large European audience, or a UK agency writing for a German client, can fall inside it. If none of your published output is used in the EU, Article 50 does not apply to you, though your own country may have rules of its own.

The bottom line

The EU did not ban AI writing and it did not ask everyone to label it. It asked one thing of publishers: if you put AI written text in front of the public on something that matters to the public, either have a person take responsibility for it or tell readers where it came from. For students, job seekers, marketers writing product copy and people using AI in their private lives, nothing changed on 2 August. For newsrooms, public bodies and anyone automating public information, a named reviewer is now the cheapest compliance tool there is.

HumanGPT (2026). Do You Have to Label AI Text in the EU? The 2026 Rules. Published 26 September 2026. https://humangpt.io/blog/label-ai-text-eu-ai-act-2026