Authors

Aurora Agostini

Partner

Giovanni Lombardi

Associate

On 20 July 2026, the European Commission adopted the “Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689 (the ‘AI Act’)” (the “Guidelines”), which provide guidance on the transparency obligations applicable to providers and deployers of AI systems (“AI”). The Guidelines clarify the scope of the obligations laid down in Article 50 of Regulation (EU) 2024/1689 (the “AI Act”), applicable from 2 August 2026 and, for generative systems already placed on the market as of that date, from 2 December 2026, limited to the marking and detection obligations under Article 50(2) of the AI Act. The Guidelines provide practical guidance not only to providers and deployers, but also to competent authorities, to promote a coherent, effective, proportionate and uniform application of the AI Act.

Context and objectives of the Guidelines

The Guidelines were prepared by the Commission pursuant to Article 96(1)(d) of the AI Act and are non-binding in nature: they serve as practical guidance for competent authorities, providers and deployers in applying the transparency obligations under Article 50 of the AI Act. The Guidelines were drafted on the basis of contributions collected during a broad public consultation organised by the Commission and input from the Member States in the AI Board.

The purpose of the transparency obligations, as indicated in Recitals 132-136 of the AI Act, is to reduce the risks of impersonation, deception, disinformation, large-scale manipulation and fraud, and to mitigate the potential negative impacts on democratic processes and collective trust caused by AI-generated or AI-manipulated content or AI-mediated interactions. Informing people about the artificial origin of interactions and content helps them make informed decisions, thereby contributing to safeguarding trust in and the integrity of the information ecosystem.

A central message of the Guidelines is that transparency is not limited to “high-risk” systems: Article 50 of the AI Act also applies to interactive systems (chatbots, avatars, virtual assistants), generative systems, deepfakes, emotion recognition systems and biometric categorisation systems. The obligations may apply simultaneously to the same system or the same output, involving providers and deployers cumulatively.

The four transparency obligations under Article 50 of the AI Act

Article 50 of the AI Act provides for four transparency obligations, applicable to different types of AI systems or their outputs. As clarified by the Guidelines, these obligations do not apply to use that is exclusively personal and non-professional. Two obligations are imposed on providers (those who develop an AI system and place it on the market under their own name or trademark) and two on deployers (those who use the system under their authority for professional purposes):

1. AI systems intended to interact directly with natural persons (Article 50(1)): providers must ensure that such systems are designed and developed in such a way that natural persons are informed that they are interacting with an AI system, unless this is obvious from the circumstances and the context of use; the information must be provided at the latest at the time of the first interaction and maintained for the duration of the interaction. The obligation applies, among others, to chatbots, avatars, virtual assistants and AI agents.
2. AI systems that generate or manipulate synthetic content (Article 50(2)): providers must ensure that outputs (audio, images, video and text) are marked in a machine-readable format and detectable as artificially generated or manipulated by means of effective, interoperable and robust technical solutions (e.g., metadata, watermarking and C2PA standards), and make corresponding detection tools available.
3. Emotion recognition and biometric categorisation systems (Article 50(3)): deployers must inform natural persons exposed to the operation of such systems, both in real time and ex post; the obligation is relevant, among others, in HR, retail, security, education and customer experience contexts.
4. Deepfakes and AI-generated text on matters of public interest (Article 50(4)): deployers must clearly and distinguishably disclose that the content has been artificially generated or manipulated, by means of a visible or audible disclosure (the provider’s technical marking is not sufficient). A less stringent regime applies to deepfakes forming part of artistic, creative, satirical or fictional works. For these purposes, “deepfake” means audio, image or video content that resembles existing persons, objects, places or other entities or events and would falsely appear to a person to be authentic or truthful (Article 3(60)); accordingly, the obligation does not apply to fantastical or non-realistic content.

Article 50(5) of the AI Act provides that the information must be provided to natural persons in a clear and distinguishable manner, at the latest at the time of the first interaction or exposure (e.g., when a conversation with a chatbot starts, when a video containing a deepfake begins, or when a text generated by AI on matters of public interest is opened), and must comply with the applicable accessibility requirements.

Exceptions and less stringent regimes

The Guidelines devote considerable attention to exceptions and less stringent regimes, which are essential to calibrating compliance obligations correctly:

1. Obviousness of the interaction (Article 50(1)): the obligation to provide information does not apply if the artificial nature of the interaction is “obvious” to a reasonably well-informed, observant and circumspect person, taking into account the circumstances and the context of use (e.g., a virtual assistant clearly presented as such).
2. Law enforcement: exceptions are provided for systems authorised by law to detect, prevent, investigate or prosecute criminal offences; however, these exceptions are to be construed narrowly and do not cover uses made available to the public for reporting crimes.
3. Standard editing and assistive functions (Article 50(2)): systems intended to perform an assistive function for standard editing, or that do not substantially alter the input data provided by the deployer or the semantics thereof, fall outside the scope.
4. Creative and satirical deepfakes (Article 50(4)): for deepfakes forming part of artistic, creative, satirical, fictional or analogous works, a less stringent regime applies: a disclosure signal is nevertheless required, but it must be presented in a manner compatible with the viewing or enjoyment of the work, without hindering its display or enjoyment, i.e., without cumbersome visual elements.
5. Texts subject to human review/editorial responsibility (Article 50(4)): the obligation to label as AI-generated does not apply to texts on matters of public interest that have undergone human review or editorial control and for which a natural or legal person has editorial responsibility.

Supervision, penalties and support tools

The Guidelines devote a section to supervisory authorities, identified primarily as the national market surveillance authorities competent under the AI Act. For certain categories of systems, the AI Office is also involved, specifically in relation to systems built on general-purpose AI models or integrated into very large online platforms or very large online search engines designated under the Digital Services Act, as is the European Data Protection Supervisor (“EDPS”) where providers or deployers are institutions, bodies or agencies of the European Union.

Breaches of the transparency obligations may result in fines of up to EUR 15 million or up to 3% of an undertaking’s total worldwide annual turnover for the preceding financial year, with proportionality criteria applicable to small and medium-sized enterprises (SMEs) and small mid-cap companies; for institutions, bodies and agencies of the Union, fines may amount to up to EUR 750,000.

The Guidelines also address the “Code of Practice on Transparency of AI-generated Content” (the “Code of Practice”), published in its final version on 10 June 2026, as a voluntary tool for demonstrating compliance with the marking and labelling obligations under Articles 50(2), 50(4) and 50(5) of the AI Act: adherence to the Code constitutes “a straightforward, predictable, and legally certain way of demonstrating compliance”.

Separately from the Guidelines, the Commission has made available additional information and support materials to facilitate compliance:

1. Dedicated FAQs (Transparency obligations under Article 50 of the AI Act): answers to frequently asked questions on the concepts and obligations clarified in the Guidelines;
2. Quick Facts: a concise information sheet on the transparency rules for AI systems;
3. EU labelling icons: a set of icons that deployers may use to label certain AI-generated content (deepfakes and AI-generated or manipulated text).

Recommendations and conclusions

The key issue for businesses is no longer simply to determine whether an AI system is “high-risk”, but to assess whether — and to what extent — it interacts with people or produces content that people see, read or hear; in such cases, the transparency obligations under the AI Act, as interpreted by the Commission’s Guidelines, apply in practice.

In light of these Guidelines, to ensure compliance with these obligations, businesses may implement a structured process along the following lines:

1. Map the AI systems in use or under development and classify them for the purposes of Article 50 (interactive, generative, emotion recognition, biometric categorisation, deepfakes, and text on matters of public interest).
2. Clarify the role (provider, deployer or both) and allocate internal responsibilities, including in contracts with cloud and AI technology providers.
3. Adapt the User Experience (UX), information notices and policies to ensure that information about interactions with AI and exposure to AI-generated or manipulated content is clear, distinguishable, accessible and provided no later than the first interaction or exposure.
4. Implement or update technical solutions for the marking and detection of generative outputs in line with Article 50(2), planning in good time for the adaptation of systems already placed on the market by 2 December 2026.
5. Define procedures for the visible labelling of deepfakes and AI-generated text on matters of public interest, in line with the Guidelines and, where appropriate, the Code of Practice if the organisation adheres to it.
6. Integrate the AI Act perspective into existing governance frameworks (GDPR, DSA, NIS 2, Data Act), adopting documentation that makes it possible to demonstrate accountability for decisions concerning design, marking and disclosure.

The Guidelines constitute a significant interpretative instrument, providing operational criteria for identifying systems subject to transparency obligations, distinguishing between the roles of provider and deployer, and determining the practical means of compliance.