AI labelling

From 2 August 2026, the transparency obligations set out in Article 50 of the AI Act will apply. These apply to both providers and users of AI systems – albeit with significantly different obligations. Companies should therefore not label all AI-supported content across the board, but should systematically review usage scenarios, roles and exceptions.

Key points at a glance:

• Article 50 of the AI Act comes into force on 2 August 2026. Providers must, amongst other things, ensure transparency in direct AI interactions and, as a general rule, mark synthetic content in a machine-readable format; users are subject to disclosure obligations, particularly in relation to deepfakes and certain texts of public interest.
• No blanket labelling requirement: Marketing, specialist or social media text generated using ChatGPT does not need to be labelled simply because AI was used. The decisive factors are the content, the purpose of publication, human verification and editorial responsibility.
• Form of disclosure: Disclosures must be clear, distinguishable, noticeable and accessible, and must be provided no later than upon first contact with the system or content. The voluntary EU icons and the Code of Practice offer practical guidance on implementation.
• Penalties and need for action: Breaches may generally be punished with fines of up to €15 million or – in the case of companies – up to 3 per cent of global annual turnover. Responsibilities, approvals and technical labelling should now be regulated in a binding manner.

What does Article 50 of the AI Act regulate?

Article 50 of Regulation (EU) 2024/1689 (AI Act) is intended to prevent natural persons from interacting with AI without realising it or mistaking artificially generated or manipulated content for authentic content. The provision covers four categories of cases. A distinction must be made between the obligations of ‘providers’ and those of ‘deployers’. Depending on the deployment model, a company may assume both roles – for example, if it uses a third-party AI service but offers it under its own name or modifies it substantially.

1. Direct interaction with AI: Providers of interactive systems must ensure that natural persons recognise when they are communicating directly with AI, for example via a customer service chatbot or voice assistant. A notice is not required if the AI nature of the interaction is obvious to a reasonably well-informed, attentive and discerning person, given the context and circumstances.
2. Machine-readable labelling of synthetic content: Providers of systems that generate synthetic audio, image, video or text content must, as a general rule, label their output in a machine-readable format and make it recognisable as artificially generated or manipulated. The technical solution must – insofar as technically feasible – be effective, interoperable, robust and reliable. Secure metadata and watermarks are particularly suitable options.
3. Emotion recognition and biometric categorisation: Users of such systems must inform the natural persons concerned about the operation of the system. In addition, data protection requirements, in particular those of the GDPR, must be observed in full.
4. Deepfakes and texts of public interest: Users must disclose when AI-generated or AI-manipulated image, audio or video content constitutes a deepfake. AI-generated or AI-manipulated texts published to inform the public about matters of public interest may also be subject to labelling requirements.

What obligations do companies have as AI users?

Article 50(4) of the AI Act is particularly relevant for business practice. The provision does not establish a general disclosure obligation for all content created using generative AI. Rather, a distinction must be made between two scenarios:

• Deepfakes: AI-generated or manipulated images, audio recordings or videos that replicate existing persons, objects, places, institutions or events and falsely appear to be authentic or true are subject to the labelling requirement. Practical examples include deceptively realistic AI videos of a real executive, voice cloning or manipulated product and event recordings.
• Texts of public interest: Texts that are generated or manipulated using AI and published with the aim of informing the public about matters of public interest are subject to disclosure requirements. This may include, for example, information of political, social, health-related or economic significance. However, this obligation does not apply if a human review or editorial check has taken place and a natural or legal person bears editorial responsibility.

For traditional corporate communications, this means that product advertising, internal texts or purely commercial content do not automatically fall into the second category. In the case of press releases, specialist articles, market information, ESG or health-related communications, however, a public interest may well be evident. A documented final human review and the clear assumption of editorial responsibility are then key compliance requirements.

What form must the labelling take?

The AI Act does not prescribe a standard wording. The information must be clear and distinguishable, provided at the latest upon the first interaction or exposure, and in compliance with accessibility requirements. It must therefore neither be hidden in the legal notice nor only become visible after the content has been consumed.

Depending on the medium, labels such as ‘AI-generated image’, ‘Voice generated by AI’ or ‘This content has been partially manipulated using AI’ may be appropriate. The EU provides optional icons for content that is fully AI-generated and for content that has been partially modified by AI. According to the implementation guidelines, the icon or text should be placed in a clearly visible position, not obscured by overlays, and, where possible, retained when the content is shared or downloaded. For audio and video content, visibility must be ensured throughout the relevant period of use.

Visible disclosure by users must be kept separate from the technical labelling provided by providers. Visible labelling does not replace the requirement for machine-readable labelling – and conversely, labelling contained solely in metadata is generally not sufficient for disclosure to humans.

What exceptions and transitional rules apply?

• Standard processing: The provider’s obligation to provide machine-readable labelling does not apply where an AI system merely performs standard processing functions or does not substantially alter the input data or its meaning. Depending on the individual case, this may include, for example, spelling, formatting or simple quality corrections.
• Human review: For texts on matters of public interest, the disclosure obligation does not apply if the content has been reviewed by a human or editorially checked, and a person or company is editorially responsible. A mere cursory nod is unlikely to suffice in this regard; the depth of the review and the assumption of responsibility should be verifiable.
• Art, satire and fiction: In the case of works that are clearly artistic, creative, satirical, fictional or similar, the disclosure must be structured in such a way that the presentation and enjoyment of the work are not impaired.
• Existing systems and legacy cases: For generative AI systems placed on the market before 2 August 2026, the deadline for implementing the technical labelling requirement under Article 50(2) of the AI Act has been extended to 2 December 2026 by Regulation (EU) 2026/1744. According to the Commission’s guidance, there is no retroactive labelling requirement for deepfakes generated before 2 August 2026; voluntary labelling is recommended.

What role do guidelines and the Code of Practice play?

In July 2026, the European Commission published guidelines on Article 50 of the AI Act, as well as a Code of Practice on the transparency of AI-generated content. The guidelines explain the scope of application, definitions, exceptions and case studies. In particular, the Code sets out in detail the technical labelling to be carried out by providers, as well as the labelling of deepfakes and relevant text publications by users.

Adherence to the Code is voluntary. However, the Commission and the AI Board have recognised it as a suitable instrument for demonstrating compliance. Signatories may rely on its measures provided they implement them correctly; other companies must demonstrate to the relevant market surveillance authorities that their alternative solutions are equally suitable. The EU icons are also voluntary and do not, in themselves, constitute legal compliance.

 What should companies do now?

1. Clarify the scope and roles: Identify interactive and generative AI systems, as well as the relevant channels. For each use case, assess whether the organisation is a provider, a user, or both.
2. Establish decision-making logic: Define robust criteria for deepfakes, public interest, material changes, standard editing and obvious AI interaction.
3. Document approvals: Establish a process for human content review, appoint editorial managers and document approval for relevant publications.
4. Ensure labelling is technically secure: Check metadata, watermarks, proof of origin, platform overlays and the persistence of visible labels during downloading and sharing.
5. Update contracts and training: Include requirements regarding labelling, the retention of metadata, evidence and responsibilities in contracts with AI providers, agencies and platforms, and raise awareness amongst marketing, communications, HR and customer service teams.

Conclusion

Article 50 of the AI Act makes transparency an integral part of AI governance, but does not require blanket labelling of all AI-assisted content. The decisive factors are role, content, purpose and human oversight. Companies that classify their use cases now, document editorial responsibility and implement labelling in a manner appropriate to the medium will not only reduce the risk of fines and reputational damage, but will also strengthen trust in their use of AI.

This article provides a non-binding overview of the subject matter and does not constitute legal advice. For further information or personal consultation, please contact: