Section 50(4) of the AI Act: What organisations must label as AI content from August 2026

For further details:

With just under three months to go until the effective date of 2 August 2026, on 8 May 2026 the European Commission published a first draft of the guidelines on the implementation of transparency obligations under Article 50 of the AI Act. The Commission thereby answers two key questions left open by the text of the AI Act.

• What content must operators actually label? and
• What form must the labelling take?

These answers provide a robust assessment framework for the first time, but they are not final until the end of the consultation phase on 3 June 2026.

Why does Article 50(4) of the AI Act require guidelines at all?

This article has been in place since the entry into force of Regulation (EU) 2024/1689 in August 2024. However, deployer of AI systems have so far been unaware of how to implement it in practice. The text of the AI Act itself uses vague legal terms such as ‘deepfake’, ‘in the public interest’ and ‘clear and distinguishable’, which require further clarification.

This clarification is to be provided through guidelines and a code of conduct.

• Whilst the guidelines focus primarily on the interpretation of abstract legal terms.
• The voluntary Code of Practice on Transparency of AI-Generated Content (as of the second draft, dated 5 March 2026) primarily addresses practical and technical implementation issues.

Neither instrument is legally binding, but both serve as recognised compliance benchmarks. Those who adhere to them provide strong evidence of compliance with their obligations. Those who ignore them bear the full risk of a fine of up to €15 million or 3% of their global annual turnover (Art. 99(4) AI Act).

What content falls under the labelling obligations under Article 50 of the AI Act?

Article 50(4) of the AI Act distinguishes between two categories: deepfakes (UA 1) and texts in the public interest (UA 2), which are clarified in the draft guidelines.

When does a deepfake exist within the meaning of the AI Act?

According to the draft, a deepfake exists if:

• the image, audio or video content is AI-generated or manipulated;
• it resembles real persons, objects, places, institutions or events, including realistic-looking fictional subjects; and
• it could be mistakenly perceived by a person as genuine or truthful; and
• the effect on the audience is assessed objectively – intent to deceive is not a prerequisite.

The definition is thus significantly broader than in general usage. It covers not only fake politicians or cloned voices, but also realistic-looking AI product images, fabricated press photos or synthetic stock images that appear authentic. However, it does not cover obviously stylised content (cartoon versions of historical events) or purely technical audio processing such as normalisation or noise reduction.

Which texts are considered to be ‘in the public interest’?

Texts are subject to the labelling requirement only if they cumulatively:

• are distributed to a large number of people (e.g. articles, push notifications, news tickers, public posts); and
• are intended to inform the public (conveying knowledge, opinions or facts, rather than being purely for entertainment); and
• relating to matters of public interest, such as politics, administration, health, the environment, the economy, science, education, security and culture of general relevance.

This covers traditional editorial content, such as election coverage, legislative analysis, business and health news, and official warnings. Marketing texts, internal memos and entertainment-only formats are generally excluded, but may be included if they address topics of general relevance.

Which exceptions to the labelling requirement apply to texts?

For texts (note: not for deepfakes), the labelling requirement under Article 50(4) UA 2, Sentence 2 of the AI Act does not apply if the following conditions are met cumulatively:

• a human review or editorial check takes place prior to publication, involving a deliberate examination of the content for accuracy, plausibility and sources, with the genuine possibility of amending or rejecting the text; and
• editorial responsibility is assumed by a clearly identifiable natural or legal person.

The draft guidelines clarify: A mere check for spelling or formal correctness is not sufficient.

In what way does Article 50(4) differ from Article 50(2)?

Article 50(2) applies to providers of AI systems. Providers of generative AI systems must mark all synthetic outputs in a way that makes them machine-recognisable as AI-generated, for example through watermarks or metadata. Please note that under the draft Digital Omnibus Act, this deadline is expected to be postponed to 2 December 2026.

Paragraph 2 applies to providers, i.e. primarily developers of AI systems, while paragraph 4 applies to deployers of AI systems, i.e. natural persons or companies that use AI systems in the course of their professional activities at their own responsibility. While paragraph 2 covers all synthetic content, paragraph 4 only covers ‘deepfakes’ and ‘texts in the public interest’.

The key difference from paragraph 2 is that, while the machine-readable provider marking runs in the background, the deployer marking under paragraph 4 must be visible and recognisable to humans. Hidden HTML tags or metadata are not sufficient.

What are the practical steps for labelling?

The draft guidelines flesh out the basic principles of Article 50(4) of the AI Act in more detail. The key requirements are clear and distinguishable labelling, which should be applied at the latest upon first encounter with the content. It should be understandable to laypeople and vulnerable groups, accessible, and not hidden in terms and conditions or sub-menus.

The draft code of conduct provides for a standardised, high-contrast ‘AI’ icon, which can be supplemented with a short text (‘Generated with AI’) where necessary. As the icon is not yet finalised, it is recommended that the plain-text label ‘AI-generated’ continues to be used until the icon is adopted.

Placement by medium:

• Image: Add a caption (‘AI-generated’) or a label/icon directly to the image. For social media thumbnails: The label should be overlaid on the image so that it remains visible when shared.
• Audio: A spoken notice at the start stating that the audio contains AI-generated content. For embedded players, include a visual label as well.
• Video: For live broadcasts, include a label at the start and throughout, plus regular notifications. For short videos, label the content from the start; for longer videos, label the content at the start, at intervals, and in the credits where applicable.
• Text: Consistently position the label in a clearly recognisable place, such as above the text, in the headline or in the legal notice block at the start of the post. For texts that are only partially AI-generated, it is sufficient to label the relevant section.

Special case: artistic, creative and satirical deepfakes:

According to Article 50(4), first subparagraph, sentence 3 of the AI Act, simplified disclosure that does not impair the enjoyment of the work is permitted. A note in the opening and closing credits of a feature film is usually enough, and permanent on-screen displays are not necessary. Note: The simplified labelling requirement does not ‘cure’ any infringement of rights. Rights of personality, copyright, trademark rights and GDPR regulations must continue to be fully respected.

Which questions does the draft leave unanswered?

Definition of ‘significant manipulation’: At what level of editing does an image created with the help of AI become ‘manipulated’ content within the meaning of the deepfake offence? While the draft cites technical audio editing as a negative example, it leaves grey areas such as AI retouching of stock photos, background replacement, and facial cropping open. Supervisory practice from August 2026 onwards will refine this threshold further.

Scope of the editorial exception for UGC and agency material: When content is supplied by freelancers, agencies or users (UGC), who bears editorial responsibility? According to the draft, a ‘clearly identifiable’ person is required, and the contractual chain of responsibility between the media company, agency and freelancer must be regulated by contract.

Interaction with the GDPR, rights of personality and copyright: The labelling requirement exists independently of other regulatory frameworks. A labelled deepfake may constitute an infringement of personal rights, a copyright infringement, or unlawful data processing. The practical relationship, particularly with regard to Sections 22 and 23 of the German Artistic Copyright Act (KUG) and Article 6 of the GDPR, must be examined on a case-by-case basis.

Platform responsibility: While the guidelines encourage platforms to preserve existing labels when content is shared, they do not impose a separate obligation on platforms. The interaction with DSA obligations remains open.

What does this mean in practice terms for the media, businesses and public authorities?

The labelling obligation under Article 50 of the AI Act applies to more than just traditional media organisations. Whenever AI-generated or AI-manipulated content — whether images, video, audio or text — is shared with an indefinite number of people, the question of Article 50(4) of the AI Act arises. The draft guidelines provide a clear assessment framework for this situation for the first time.

For media companies, publishers and editorial teams:

The editorial exception is the key lever. However, it only applies where there is a genuine review of content with a documented assignment of responsibility. Anyone using AI workflows should now check the following: Which content is created with AI support? Which content undergoes genuine editorial review? Where is the chain of responsibility documented? Most organisations will need to update their internal editorial guidelines and freelance contracts.

For companies with AI-supported marketing, communications and customer service:

The labelling requirement may also apply outside of traditional editorial departments, particularly for AI-generated product images intended to convey authenticity, and for external communication formats that serve an informative purpose. It is essential to have an AI inventory and clear workflows for labelling and documentation.

For public authorities and bodies:

As deployer, they are covered by Article 50(4) of the AI Act. Official warnings, press releases and information services must be examined particularly carefully, as they often fulfil the criterion of being ‘in the public interest’.

For AI providers:

Although the primary responsibility under paragraph 4 lies with the operators, the providers must fulfil their obligations under paragraphs 1 and 2 (interaction notice and machine-readable labelling, respectively). The key compliance indicator here is compliance with the Code of Conduct.

Recommended measures by 2 August 2026:

In order to implement the labelling requirement under Article 50 of the AI Act by the deadline, we recommend addressing the following points:

• AI inventory: Which systems generate or manipulate which content?
• Workflow adjustment: Which content is labelled and which is exempt?
• Documentation: Tool, version, person responsible for checking and date – all traceable in the CMS.
• Contract clauses: Check freelancer and agency contracts for AI clauses and the allocation of responsibility.
• Training: Prepare the editorial, marketing and communications teams for their new obligations.

Frequently Asked Questions (FAQ)

Does the labelling requirement apply retroactively to archived content?

No, Article 50(4) of the AI Act only applies to content published or modified on or after 2 August 2026. Neither Article 50(4) nor the draft guidelines provide for a retrospective labelling requirement for archives. However, anyone republishing or substantially modifying archived content from August 2026 onwards will trigger the obligation.

When is the editorial exception truly met?

Only if a human being checks the content for accuracy, plausibility and sources, and the editorial team takes clear responsibility. This covers AI-assisted article drafts with genuine editorial revision, as well as routine AI reports (e.g. weather, brief financial news, sports) with plausibility checks. Not covered are comments written almost entirely by AI with only linguistic refinement, automated translations without substantive checks and aggregation tickers without prior verification.

What applies to marketing content and social media posts?

In principle, pure product advertising and corporate communications do not fall under the category of ‘in the public interest’. However, marketing content designed to look like editorial content, such as native advertising on health or financial topics, may fall within the scope of the regulations. Furthermore, the deepfake provisions apply to AI-generated product images that appear authentic, regardless of the marketing context.

What is the specific risk of a fine?

The fine can be up to €15 million or 3% of the company’s global annual turnover, whichever is higher (Art. 99(4) AI Act). There is no provision for issuing a warning prior to imposing a fine. Under the Digital Omnibus draft, SMEs and start-ups are expected to receive automatic reductions in fines (50% for SMEs and 75% for micro-enterprises).

Do the obligations also apply to deployers outside the EU?

Yes, the AI Act adheres to the market location principle. According to Article 50, anyone who publishes or makes AI-generated content accessible in the EU falls under its remit, regardless of the deployer‘s place of establishment.

Consultancy services

We advise media companies, publishers, public bodies, eductaoinal companies, edtec providers and AI providers on all aspects of compliance with the AI Act, from conducting baseline assessments and designing workflows and contracts to liaising with supervisory authorities.

We offer structured initial analyses to assess your current situation and handle the adaptation of editorial guidelines, contracts and documentation for implementation. We also conduct workshops and training sessions for internal teams, which can be tailored to your industry if required.

To arrange an appointment or a no-obligation preliminary discussion: email Dr Kerstin Bäcker