AI-generated texts: when is a label required?

Part 2 of a three-part series on the new European guidelines on AI transparency

From 2 August 2026, the transparency obligations under the AI Act apply. But does using ChatGPT or another generative AI tool mean that an article, report or social media post must now be labelled as AI-generated content?

No. Not every text created with the assistance of AI needs an AI label. What matters, among other things, is the purpose for which the text is published and whether human review and editorial responsibility are in place.

In Part 1 of this series, we discussed when the use of AI must be disclosed. In this second part, we focus specifically on AI-generated and AI-edited texts.

Not every AI-generated text needs to be labelled

Organisations increasingly use AI to draft, summarise, rewrite or translate texts — including articles, news reports, reports and social media posts. This does not mean that every text created with the assistance of AI must visibly be labelled as AI-generated content.

Article 50(4), second subparagraph, of the AI Act specifically concerns texts that are published for the purpose of informing the public on matters of public interest.

Three cumulative conditions

A visible AI disclosure is required only where all three of the following conditions are met:

1. the text has been artificially generated or manipulated;

2. the text is published for the purpose of informing the public;

3. the text concerns a matter of public interest.

   
   

The mere fact that AI was used during the writing process is therefore not sufficient to trigger a labelling obligation.

When is a text considered ‘published’?

Publication means that the text is made accessible to an indefinite and relatively large number of people — including where access requires payment or a subscription.

Examples include:

• articles on a publicly accessible website;

• news reports;

• publicly accessible social media posts;

• public corporate reports.

Individual correspondence, advice provided to a specific client, internal communications, messages within a small closed group, or a chatbot response visible only to an individual user are not considered publication in this sense.

What is a ‘matter of public interest’?

This concept is interpreted broadly. It covers matters that are relevant to society and may give rise to public debate or public scrutiny.

The guidelines refer, among other things, to:

• politics and democratic processes;

• public administration and public services;

• the administration of justice and fundamental rights;

• public safety;

• public health and consumer safety;

• the environment and climate;

• economic, financial, scientific and cultural developments.

Marketing texts and product descriptions will generally not fall within the labelling obligation. This may be different where they contain claims relating, for example, to health, consumer safety, the environment, sustainability or financial developments. In such cases, it must be assessed on a case-by-case basis whether the text is also intended to inform the public about a matter of public interest.

What should an AI label look like?

Where an AI disclosure is required, it must be perceptible to the reader.

A technical record contained only in metadata is not sufficient. Nor is a disclosure that is buried solely in general terms and conditions or other general policies.

The reader must therefore be able to identify, in connection with the relevant content, that it has been artificially generated or manipulated.

The important exception: human review and editorial responsibility

Even where AI-generated or AI-edited text is published for the purpose of informing the public on a matter of public interest, a separate AI label is not always required.

A separate AI disclosure is not required where both of the following conditions are met:

1. the text has undergone a process of human review or editorial control;

2. a natural or legal person holds editorial responsibility for its publication.

This distinction is particularly important for organisations using generative AI for articles, reports and other external communications.

What is not sufficient as human review?

Not every human intervention in relation to a text qualifies as sufficient review.

The following, for example, are not sufficient:

• checking spelling and grammar only;

• briefly reading through the text;

• approving the text without substantive review;

• using only a second AI system to review the text;

• having an editorial policy without actually reviewing the specific text.

What is required is a deliberate substantive review by someone with relevant knowledge and professional judgement, with fact-checking forming at least part of that review.

The final version is decisive

An important practical point is that human review must relate to the final version that is published.

If a text is reviewed by a person and is subsequently substantially rewritten by AI, the new version must be reviewed again.

A robust process could therefore look like this:

AI-assisted draft → substantive human review → corrections → final approval → publication

Organisations that systematically use AI for external communications should therefore establish not only rules governing the use of AI, but also when the final human review must take place.

Who holds editorial responsibility?

Editorial responsibility may rest with the publishing organisation, an editorial team, a responsible subject-matter expert or an individual author.

The relevant person or organisation must, however, genuinely have the authority to approve, amend or reject the text.

The Commission recommends making the party with editorial responsibility easy to identify, for example through an imprint or the legal information on a website, in order to support public accountability.

This is a recommended practice that may help substantiate reliance on the exception. Article 50 itself does not require an individual editor to be identified for every publication.

What does this mean for internal governance?

Organisations using generative AI for external publications should consider documenting:

• where AI is used for public-facing content;

• who reviews the content and verifies the facts;

• who approves the final version;

• who holds editorial responsibility;

• whether any further AI edits are made after the final review.

  
  

A burdensome new control process is not necessarily required. However, it should be demonstrable that the final published version has undergone substantive review by an authorised and appropriately qualified person.

Alignment with and adherence to the European Code of Practice on the transparency of AI-generated content, published on 10 June 2026, may help organisations substantiate their processes and compliance.

The Code of Practice does not, however, provide an exemption from Article 50, nor does it offer an absolute guarantee of compliance with the legal requirements.

Conclusion: When Is an AI Label Required?

The central question is therefore not simply: was AI used to write the text?

The first question is whether the text is published for the purpose of informing the public on a matter of public interest. If so, it then becomes relevant whether the final text has undergone substantive human review and whether it is clear who holds editorial responsibility for its publication.

Using AI to draft, summarise, rewrite or translate a text therefore does not automatically result in an obligation to apply an AI label.

In Part 3 of this series, we will examine chatbots, AI assistants and deepfakes.

Earlier in this series: Part 1 — When must the use of AI be disclosed?

Want to know more?

DUFINCO helps organisations inventory their AI applications and determine which transparency obligations under the AI Act apply. Does your organisation use generative AI for articles, reports, websites, social media or other external communications? We can assess when an AI disclosure is required and how human review and editorial responsibility can be implemented in practice.

Would you like to know what the AI transparency requirements mean in practice for your organisation?

Contact us at info@dufinco.nl or call +31 (0)6 512 47 217 for a no-obligation introductory discussion.