Pursuant to Article 50 of the AI Act, as of 2 August 2026, providers and deployers of AI systems are subject to a transparency obligation vis-à-vis users (I). This obligation is subject to limited exceptions (II) and entails practical implications in the manner in which AI systems are designed and placed on the market (III).
I. The transparency obligation
The Regulation provides that providers shall inform users (i) that they are interacting directly with an AI system and (ii) that content generated or manipulated by AI must be clearly marked and detectable as generated or manipulated by AI.
Deployers, for their part, shall (i) inform the user of the operation of the AI system where it is capable of emotion recognition or biometric categorisation, and (ii) disclose that the content has been generated or manipulated by AI in the case of the generation or manipulation of deepfakes or of text published for the purpose of informing the public on matters of public interest.
II. Exceptions to the transparency obligation
As a general rule, the transparency obligation does not apply where systems are authorised by law to detect, prevent, investigate or prosecute criminal offences.
In addition to this general derogation, other specific exceptions listed below apply. These exceptions stem from a balancing of the transparency obligation with the fundamental rights and freedoms of conducting a business, artistic expression, and freedom of expression and information.
- The obligation for providers to mark content generated or manipulated by AI does not apply where AI systems perform an assistance function for standard formatting or do not substantially alter the input data provided by the deployer or their semantics.
- The obligation for deployers to disclose that “deepfake” content has been generated or manipulated by AI is limited to the disclosure of the existence of such content so as not to hamper the display or enjoyment of the work where the content forms part of a manifestly artistic, creative, satirical, fictional or analogous work or program.
- The obligation for deployers to disclose that content intended to inform the public on matters of public interest has been generated by AI does not apply where the content has been subject to a human review process or editorial control and where a natural or legal person assumes editorial responsibility for the publication of the content.
III. Practical implications of the transparency obligation
In order to understand the practical implications of the transparency obligation for providers and deployers, reference should be made to the draft Code of Practice on transparency of AI-generated content, developed under the auspices of the European Commission, which is intended to structure and operationalise the obligations laid down in Article 50 of the AI Act.
The draft Code of Practice recommends that providers embed in the metadata of generated content the information required by the Regulation, accompanied by digital signature mechanisms ensuring its integrity and, where possible, “watermarks” (imperceptible watermarks) directly embedded in the output, capable of surviving transformation operations as well as potential alterations of the content. Where it is technically impossible to implement such measures, the draft Code of Practice prescribes the use of complementary “logging” mechanisms or “digital footprint” mechanisms.
According to the draft, these requirements should be integrated by providers from the design stage of AI systems and positioned upstream in order to facilitate compliance by downstream operators and to ensure traceability throughout the value chain.
To ensure the effectiveness of the transparency obligation, the draft Code of Practice provides that providers shall make available free of charge an interface (for example, an API or a user interface) or a publicly accessible detection tool enabling users and other interested parties to verify, by means of confidence scores, whether content has been generated or manipulated by their AI system or model.
Furthermore, the draft provides for the creation of a common icon within the European Union to signal content generated or manipulated by AI.
The draft is particularly detailed: by way of illustration, the recommendations relating to “deepfakes” (A) and to content intended to inform on matters of public interest (B) are addressed below.
A) Recommendations concerning “deepfakes”
The draft Code of Practice on the transparency of AI-generated content recommends that:
- “Deepfakes” be signaled in a clear and distinct manner from the first exposure.
- In the case of real-time video, there be a continuous display of a non-intrusive icon accompanied by a warning at the beginning of the video.
- In the case of non-real-time video, there be a notification by means of a visible icon accompanied by a warning at the beginning and/or a permanent icon.
- In the case of multimodal content (image-text-audio, etc.), there be an icon displayed permanently, clearly visible without user action.
- In the case of a “deepfake” image, there be a fixed, visible icon distinct from the image and not concealed.
- In the case of audio < 30 seconds, there be a brief audio warning at the beginning indicating the artificial origin.
- In the case of long-form audio (podcast, etc.), there be repeated audio warnings.
B) Recommendations concerning content intended to inform on matters of public interest
The draft Code of Practice on the transparency of AI-generated content recommends that, for content intended to inform the user, the deployer adopts procedures and retains documentation proportionate to its size, but including at least the following elements:
- identification of the natural or legal person responsible at editorial level (name, function and contact details);
- an overview of the concrete organisational and technical measures as well as the human resources put in place to ensure adequate human or editorial control prior to the publication of textual content generated or manipulated by AI, taking into account, where applicable, national specificities (linguistic, cultural or contextual factors capable of influencing interpretation or impact);
- the date of review and approval;
- a reference to the final approved version of the content (for example, file name, URL or any other internal identifier).
In conclusion, for undertakings, it is not a matter of awaiting the formal application of Article 50, but of anticipating, from now on, the implementation of transparency mechanisms that are legally sound, technically robust and compliant with what is set to become the European standard of transparency in the field of content generated by artificial intelligence.
