In the legal debate surrounding artificial intelligence, recent attention has largely focused on the obligation to label deep fakes – realistic images, audio recordings and videos created or manipulated using AI. From 2 August 2026, the law requires such content to be appropriately labelled so that recipients are aware that it has been artificially generated or manipulated. However, the transparency obligations imposed by the AI Act are not limited to deep fakes. Less obvious – and less frequently discussed – are the rules governing AI-generated text. In this area, the mere use of AI to prepare a text does not, as a rule, trigger an obligation to inform recipients that AI has been used. In certain circumstances, however, such an obligation may arise.

The obligation to label text arises only where an AI system generates or manipulates text which is published for the purpose of informing the public on matters of public interest (Article 50(4), second subparagraph). The provision lays down three conditions, all of which must be satisfied: the text must be published, must inform the public, and must concern a matter of public interest. If any one of these conditions is not met, the obligation does not apply. The scope of these concepts is further clarified in the European Commission’s Guidelines on Article 50[1].

A natural area of application is news portals and press agencies – the Guidelines expressly refer, by way of example, to a newsroom whose journalists use AI to assist them in reporting on current events[2]. However, the provision is not limited to the media and may also apply to public institutions, civil society organisations and businesses publishing reports, analyses or communications addressed to the general public.

When is a text ‘published’?

The Guidelines adopt a broad interpretation: a text is considered published if it is made available to an undefined, relatively large number of unrelated potential readers – including on a paid basis, such as under a subscription model[3]. Individual correspondence, communications within a closed private group and internal organisational materials, for example content published on an intranet, fall outside the concept of publication.

Neither payment nor a restricted group of recipients therefore automatically excludes the obligation. A newsletter distributed to a broad customer base may satisfy the publication criterion even if it is not publicly accessible.

When does a text ‘inform the public’?

A text informs the public where it is intended to communicate knowledge, opinions or facts[4]. The concept therefore encompasses not only factual or informational materials, but also commentary and opinion pieces – what matters is the intention to communicate content, rather than whether that content is neutral. By contrast, short texts which do not in substance communicate knowledge, opinions or facts fall outside the scope of the obligation.

When does a text concern a matter of public interest?

The AI Act does not define this concept or provide an exhaustive list of matters falling within it. The Guidelines indicate that it covers matters of significance to society – at local, national, EU or international level – which warrant public debate or scrutiny, including in particular[5]:

• politics and democratic processes;
• public administration and public services;
• the administration of justice and law enforcement;
• the protection of fundamental rights and public safety;
• public health, environmental protection and consumer safety;
• economic, financial, political, scientific or cultural developments capable of becoming a significant subject of public debate.

The list is non-exhaustive, and the Commission notes that the scope of matters of public interest may evolve over time and depend on the relevant context. According to the examples given in the Guidelines, advertising and product descriptions fall outside the obligation, provided they do not contain claims relating to health, consumer safety or sustainability, as does compliance advice provided to a client. By contrast, corporate reports containing information for investors and published on the website of a listed company are subject to the obligation.

Human review may remove the obligation – but not every review will suffice

A text concerning a matter of public interest does not need to be labelled if it has undergone human review or editorial control and editorial responsibility for its publication is held by a natural or legal person. Both conditions must be satisfied.

The exemption does not apply automatically. Review means a deliberate examination of the substance of the content by a person with appropriate expertise and professional judgement, with verification of the accuracy of the information constituting the minimum requirement. Merely correcting spelling or grammar, having an editorial policy in place, conducting an automated review or giving the text cursory approval without substantive involvement is not sufficient[6].

A key caveat is that, in the Commission’s view, any material AI intervention after editorial approval invalidates the exemption. If, following approval, the text is materially amended, supplemented or redrafted by AI, it must be labelled[7].

Summary

For textual content, the obligation to label is not the rule – it applies to informational and opinion-based publications concerning matters of public interest. Even then, labelling is not required where the content has undergone proper human review or editorial control and editorial responsibility for its publication has been assigned. In practice, however, the key challenge may lie not in classifying the content itself, but in documenting the review process and being able to demonstrate that review in relation to the final intervention made in the text.

The lawyers at Jabłoński Koźmiński i Wspólnicy advise businesses and publishers on assessing published content, determining their role as deployers under the AI Act, and implementing appropriate review and approval procedures.

[1] European Commission Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689, C(2026) 5054 final, Brussels, 20 July 2026.

[2] Paragraph 14 of the Guidelines.

[3] Paragraph 131(i) of the Guidelines.

[4] Paragraph 131(ii) of the Guidelines.

[5] Paragraph 131(iii) of the Guidelines.

[6] Paragraphs 133–138 and 140 of the Guidelines.

[7] Paragraph 136 of the Guidelines.