Does AI-written text have to be labelled? Article 50 explained

Usually not. Article 50 of the EU AI Act requires a label on AI-generated text only when it is published to inform the public on matters of public interest. Even that text can go out unlabelled if a person has reviewed it properly and someone holds editorial responsibility for publishing it. The review has to meet a higher bar than many assume, and the exemption itself has been in the regulation since 2024.

What Article 50 requires

Article 50 of the EU AI Act sets out four duties, split between two roles.

Providers of chatbots must make sure users are told they are talking to an AI. Providers of generative systems must mark what the system produces in a machine-readable format, so the content can be detected as artificially generated. Deployers of emotion-recognition and biometric-categorisation systems must inform the people exposed to them. And deployers who create deepfakes, or who publish AI-generated text to inform the public, must disclose that the content was artificially generated or manipulated.

A deployer is anyone using an AI system in a professional capacity. If you put text from ChatGPT or Copilot on your website, you are a deployer. The company behind the tool is the provider. Text published by deployers is the part most organisations will run into, and it is the only part where human review can lift the duty.

The human-review exemption

Article 50(4) says the duty does not apply where the AI-generated content has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for publishing it. Both conditions must be met.

It covers text only. Deepfake images, audio and video have no equivalent exemption. The only relief there is a law-enforcement exception and a lighter duty for work that is evidently artistic, satirical or fictional.

Nor does it help providers. The machine-readable marking in Article 50(2) is the toolmaker's responsibility, whatever an editor later does with the text.

When the exemption came in

It was in Article 50 as adopted in 2024. The Norwegian Union of Journalists discussed it in its consultation response on the draft Norwegian AI law in September 2025, two months before the Commission proposed the Digital Omnibus simplification package on 19 November 2025. In the version of the article shown in the AI Act Explorer, only paragraph 7 is marked as amended.

What changed in 2026 was the material around it:

• The final Code of Practice on marking and labelling AI-generated content, published on 10 June, dropped the "fully AI-generated" versus "AI-assisted" distinction from its first draft.
• The Commission's guidelines on Article 50 were finalised on 20 July and describe what counts as review.
• The Digital Omnibus, in force since 27 July, gave providers of generative systems already on the market before 2 August until 2 December 2026 for machine-readable marking. Deployer duties under Article 50(4) were not postponed.

What counts as human review

The Commission's guidelines (C(2026) 5054, section 6.2.3, as reproduced by the MQM Council) set four conditions for a review to count:

1. Someone with relevant knowledge and professional judgement has to assess what the text says. Fact-checking is the minimum.
2. Spell-checking, grammar correction, automated review, a written editorial policy or a quick sign-off are not enough on their own.
3. A named person or function holds final legal responsibility for the publication, and their contact details are public and easy to find, for example in a colophon or legal notice.
4. The review comes last. If AI changes the text substantially after sign-off, the exemption is lost. A machine translation of an approved text needs its own review.

The Code of Practice asks deployers to name who holds editorial responsibility and to document the routines that make sure review happens. Logging each individual review is not required.

Text the duty does not cover

Most text never gets as far as the exemption, because the duty does not apply to it. It covers text that is published, meant to inform, and about a matter of public interest. All three conditions have to be met.

According to the guidelines, "published" means available to an indeterminate and fairly large number of unrelated readers. Internal texts, intranet pages and private professional correspondence are outside. "Public interest" covers politics, public administration, the justice system, fundamental rights, public health, the environment and consumer safety, plus economic, scientific and cultural developments that may be the subject of public debate.

The Commission's examples of text outside the duty include AI-written novels, advertising and product descriptions, unless they make claims about health, consumer safety or sustainability. Investor information on a listed company's website is given as an example of text that is covered.

For most Norwegian small businesses, ordinary marketing and product information is outside. Articles about social issues, public administration or health may be inside.

Status in Norway

The AI Act is assessed as EEA-relevant but has not been incorporated into the EEA Agreement, so it is not Norwegian law. Minister for Digitalisation and Public Governance Karianne Tung has said the government aims for a bill in spring 2027, but that progress depends on negotiations with the EU on EEA adaptations. NTB reported in late summer that the requirements do not apply in Norway as long as the AI law is missing.

Norwegian organisations are still not entirely free of it. Companies outside the EU must follow the rules if their AI systems or outputs are used in the EU. Article 50 also says expressly that it does not displace other transparency duties in EU or national law, and the Norwegian Union of Journalists notes in its consultation response that press ethics already require openness and editorial responsibility for AI-produced content.

We have looked at what the AI Act means for small businesses in Norway

My objection to the exemption

A label is meant to tell readers how much of a text a machine wrote. The exemption answers a different question: whether someone has taken responsibility for it. A reader who sees an editor's name learns nothing about how the text was produced.

It is also hard to check. Because individual reviews do not have to be logged, a supervisory authority has little to go on, and only the person who approved a text knows whether they read it closely.

The case for the exemption is a good one. Editorial responsibility is how publishing already works: a newspaper does not mark each story "written by a journalist", but someone can be held to account when one is wrong. If every text AI had touched carried a label, the label would soon stop meaning anything. And the guidelines have narrowed the exemption by demanding fact-checking, subject knowledge and a public name.

I would still give the reader more weight. A visible label tells the reader something, and an internal routine does not. The EU struck the balance the other way, and that choice can be defended.

What to do now

None of these steps depends on when the rules reach Norwegian law:

1. Go through what you publish and decide which texts are covered. Marketing and product information usually are not. Claims about health, safety or sustainability are.
2. Name the person or function with editorial responsibility, and put the name and contact details somewhere easy to find on your website.
3. Make review the last step, after any AI editing. The reviewer needs subject knowledge and the authority to stop publication.
4. Write down the routine: who reviews what, and how. You do not need to log each review.
5. Handle images, audio and video separately. The exemption does not cover them.

Frequently asked questions