The Artificial Intelligence Regulation (AI Act) establishes a harmonised legal framework for the development, placing on the market and use of AI systems within the European Union. Among the requirements set out in the Regulation are specific transparency obligations, which apply to certain AI systems irrespective of whether they are classified as high-risk systems.
These obligations are intended to address particular risks associated with the use of AI, notably risks of deception, impersonation, manipulation of information and erosion of public trust.
Article 50 of the AI Act constitutes the legal basis for these transparency obligations. It is supplemented by explanatory recitals, a draft code of practice and draft guidelines currently under consultation, which provide important clarifications regarding their interpretation and implementation.
I. The Legal Framework for Transparency Obligations under the AI Act
1. Information in the Case of Direct Interaction with a Natural Person – Article 50(1)
Article 50(1) requires providers to ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the persons concerned are informed that they are interacting with an AI system.
This obligation does not apply where the artificial nature of the interaction:
- is obvious from the perspective of a reasonably well‑informed, observant and circumspect person,
- taking into account the circumstances and the context of use.
Exceptions
The obligation does not apply:
- to systems whose use is authorised by law for the purposes of the prevention, detection or prosecution of criminal offences,
- subject to appropriate safeguards for the rights and freedoms of third parties,
- except where such systems are made available to the public to allow the reporting of a criminal offence.
Purpose and Protective Rationale
Recital 132 clarifies that certain AI systems, even when not classified as high-risk, may present specific risks of deception or impersonation when interacting with humans. The objective is therefore to prevent individuals from being misled into believing that they are interacting with another human being.
Practical Implication
This obligation mainly concerns direct and perceptible interactions, for example:
- customer service chatbots,
- voice assistants,
- conversational agents embedded in websites or applications,
- avatars or virtual characters interacting with users.
In practice, users must be informed from the outset of the interaction that they are dealing with an automated system, unless this is already obvious.
Examples
- A chatbot that clearly presents itself as a virtual assistant or whose interface makes its AI nature obvious complies with the obligation.
- Conversely, a system that adopts a human identity without any indication of its artificial nature may create confusion incompatible with Article 50.
2. Labelling and Detectability of AI-Generated or AI-Manipulated Content – Article 50(2)
Article 50(2) requires providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text content, shall ensure that the outputs:
- are marked in a machine-readable format;
- are detectable as artifitially generated or manipulated.
The technical solutions implemented must be as effective, interoperable, robust and reliable as technically feasible, taking into account the characteristics of the content and the costs of implementation.
Exceptions
This obligation does not apply where:
- the system performs a purely assistive function for standard editing (e.g. spell checking or formatting);
- the system does not substantially alter the input data or its semantics;
- the use is authorised by law for the prevention or prosecution of criminal offences.
Purpose and Protective Rationale
Recital 133 specifies that the marking techniques required under this provision should be as reliable, interoperable, effective and robust as the state of the art allows. This requirement may be fulfilled by a single technique or a combination of techniques, such as watermarks, metadata identifiers, cryptographic methods for proving provenance and authenticity, logging mechanisms or digital fingerprints, depending on the type of content concerned.
Practical Implications
This obligation concerns the design of the AI system itself. It does not relate to visible disclosure to the end user, but rather to the integration of technical mechanisms that allow the artificial origin of generated content to be identified.
This includes in particular:
- text generators (articles, scripts, posts),
- synthetic image or video generators,
- voice or music generation systems.
Examples
- An AI image generator integrated into a marketing tool must embed technical marking within the generated files.
- A text generation model intended for automated publication must allow the artificial origin of content to be identified even after dissemination on third-party platforms.
3. Information in the Case of Emotion Recognition or Biometric Categorisation – Article 50(3)
Article 50(3) requires deployers of emotion recognition systems or biometric categorisation systems to inform the natural persons exposed to such systems of their operation.
Personal data must be processed in accordance with applicable EU data protection rules.
Exceptions
This obligation does not apply where the use of the system is authorised by law for the purposes of prevention, detection or investigation of criminal offences, subject to appropriate safeguards.
Purpose and Protective Rationale
Recital 132 highlights the particular sensitivity of systems capable of deducing or inferring emotions, intentions or personal categories from biometric data. Such uses may have a direct impact on the fundamental rights of the persons concerned.
Practical Implications
Individuals must be informed:
- of the existence of the system;
- of the general nature of its functioning.
This obligation applies regardless of whether the system operates in real time and is cumulative with GDPR obligations.
Examples
- A facial expression analysis tool used during a video interview must be clearly disclosed to candidates.
- A system analysing voice or behaviour in a public environment must signal its operation to exposed individuals.
4. Transparency Regarding Deepfakes and Texts on Matters of Public Interest – Article 50(4)
A. Deepfakes
Deployers of AI systems that generate or manipulate image, audio or video content constituting deepfakes must disclose that such content has been artificially generated or manipulated.
Recital 134 specifies that deepfakes are contents that may falsely appear authentic or truthful to the public. The assessment is based on the risk of misleading perception, not on the creator’s intent.
Where the content forms part of an evidently artistic, creative, satirical, fictional or similar work, the transparency obligation is limited to disclosing the existence of AI-generated or AI-manipulated content in an appropriate manner that does not impair the enjoyment of the work.
Exceptions
This obligation does not apply where the use is authorised by law for criminal offence prevention, detection, investigation or prosecution.
B. Texts Published on Matters of Public Interest
Deployers of AI systems that generate or manipulate texts published for the purpose of informing the public on matters of public interest must disclose that the text has been artificially generated or manipulated.
This includes in particular texts relating to politics, economics, health, the environment or science. The objective is to enable the public to assess the origin of the content, the level of human involvement and the existence of editorial responsibility.
Exceptions
This obligation does not apply where:
- the use is authorised by law for criminal offence prevention or prosecution;
- the content has undergone human review or editorial control;
- a natural or legal person assumes editorial responsibility for the publication.
5. Article 50, paragraphe 5 – Modalités communes de fourniture de l’information
Information required under paragraphs 1 to 4 must be provided:
- in a clear and recognisable manner;
- at the latest at the first interaction or first exposure;
- in accordance with applicable accessibility requirements.
The information must not be hidden in terms and conditions or require multiple user actions to access.
II. European Commission Guidelines
The European Commission has recently published draft guidelines on the implementation of the transparency obligations under Article 50 of the AI Act, currently subject to public consultation until 3 June 2026. These guidelines are not legally binding and may evolve before final adoption.
They notably clarify:
- the concept of direct interaction, defined as a bidirectional exchange in real time or near real time, and strictly limit the “obviousness” exception;
- that marking and detectability of synthetic content must operate jointly, and that only purely technical or minor interventions qualify as standard assistance;
- that the obligation to inform in relation to emotion recognition or biometric categorisation applies regardless of when processing occurs, independently of data protection law requirements;
- the qualification of deepfakes based on the risk of false authenticity, and a restrictive interpretation of exceptions related to artistic works and human editorial control;
- the overarching requirement that information be clear, distinct and perceptible to each exposed person from the first contact, including in cases of deferred or repeated exposure.
The guidelines also emphasise the role of codes of practice as a means of demonstrating compliance: adherence to an approved code is voluntary but constitutes a privileged reference for supervisory authorities, while non-signatories must provide more detailed justification of the effectiveness of their measures.
III. Code of Practice on Marking and Labelling AI-Generated Content
The European Commission is facilitating the development of a voluntary code of practice, currently in draft form, to assist providers and deployers in implementing the transparency obligations under Article 50 of the AI Act. The code is not legally binding and may evolve before finalisation.
Two drafts have been published so far: a first in December 2024 and a second in March 2025. The final version is expected in May/June 2026.
The draft code is structured into two sections:
- one addressed to providers under Article 50(2), detailing technical modalities for marking and detectability of AI-generated content (notably a combined approach using secure metadata and watermarks);
- one addressed to deployers under Article 50(4), providing practical guidance on the labelling of deepfakes and certain texts of public interest, and defining common principles for presenting transparency labels or notices.
The objective of the code is to facilitate the demonstration of compliance and promote harmonised practices, without creating new legal obligations.
IV. Adjustment of the Timeline via the AI Omnibus
As part of the “AI Omnibus” initiative aimed at simplifying certain AI Act obligations, a political agreement was reached in early May 2026 to postpone the application of the machine-readable marking obligations under Article 50(2) until 2 December 2026.
At the time of publication of this article, this measure has not yet been formally adopted and remains subject to legislative finalisation.
All other transparency obligations under Article 50 will nevertheless apply from 2 August 2026, in line with the general AI Act timeline.
Implementing Transparency Over Time
Compliance with the AI Act’s transparency obligations goes beyond the one-off addition of labels or notices. It requires an ongoing ability to identify relevant systems, determine applicable obligations based on use cases, and demonstrate that information provided to individuals remains clear, accessible and effective throughout the AI system lifecycle.
This notably involves structuring internal practices to:
- link transparency requirements to the relevant systems and use cases;
- translate legal obligations into concrete information, marking and documentation mechanisms;
- document design and deployment choices;
- ensure consistent and auditable monitoring of obligations over time.
👉 In this context, the Naaia platform has been designed to support organisations in the operational implementation of AI Act compliance, notably by facilitating governance, traceability and the management of transparency obligations applicable to artificial intelligence systems.