The European Commission has made public the first draft of the Code of Practice on Transparency of AI-Generated Content, a document intended to play a central role in the implementation of Article 50 of the AI Act.
The Code aims to provide a shared operational framework to ensure that content generated or manipulated by artificial intelligence systems is recognizable, traceable, and correctly communicated to end users, thereby contributing to safeguarding trust in the information ecosystem and to the protection of fundamental rights.
The draft is the result of a broad and articulated consultation process, which involved hundreds of stakeholders, including industry operators, academia, civil society, and Member States. The work, launched in November 2025, was structured into two working groups dedicated respectively to the obligations of providers of generative AI systems and to those of deployers that use such systems for content dissemination.
The document is presented as a working basis intended to be further refined, but already capable of outlining regulatory expectations and best practices to ensure transparency, traceability, and technical compliance.
General objectives of the Code
At a general level, the Code pursues the objectives of improving the functioning of the European digital single market, promoting the adoption of trustworthy and human-centric AI, and ensuring a high level of protection of health, safety, and fundamental rights, including democracy, the rule of law, and environmental protection.
More specifically, the document identifies two main lines of action: first, it aims to serve as guidance to ensure compliance with the obligations set out in Article 50(2) and (5) of the AI Act (while clarifying that adherence to the Code does not constitute conclusive proof of compliance with those obligations); second, the Code seeks to ensure that providers and deployers of artificial intelligence systems comply with the transparency obligations laid down by the AI Act, while at the same time enabling the competent supervisory authorities to assess in concrete terms the reliability of the Code as a compliance tool.
Obligations of providers
From the providers’ perspective, the Code focuses on the marking and detectability of content generated or manipulated by AI. Pursuant to Article 50(2)(a) and Recitals 133 and 135 of the AI Act, providers are required to make such content clearly detectable and traceable, considering the specificities and limitations of different types of content, implementation costs, and the generally recognized state of the art.
To this end, the Code introduces a multi-layered approach to marking, acknowledging that no single technique is sufficient to ensure the effectiveness, robustness, reliability, and interoperability of the adopted solution (pursuant to Article 50(2) and (5) of the AI Act). The techniques envisaged include imperceptible watermarks, digital metadata, provenance certificates, and fingerprinting or logging systems. The objective is to ensure that the artificial origin of content can be verified throughout the entire provenance chain, even in the presence of subsequent modifications or reuse in different contexts.
Particular attention is devoted to multimodal content and open-weight models, for which the Code suggests structural marking solutions already at the model training stage, to facilitate downstream operators’ compliance. Alongside marking obligations, the draft provides that providers make available accessible detection tools, such as interfaces or public detectors, suitable to enable users, platforms, and competent authorities to verify whether content has been generated or manipulated by a specific AI system.
The Code also addresses the issue of interoperability and standardization, encouraging the use of open standards and shared solutions among providers, platforms, and supervisory authorities, also with a view to containing compliance costs, particularly for SMEs and startups. From this perspective, transparency of AI-generated content is conceived not only as a regulatory obligation, but also as an area of technological innovation and industrial cooperation.
Obligations of deployers
Regarding deployers, the draft Code focuses on disclosure obligations towards the public, in implementation of Article 50(4) and (5) of the AI Act. In particular, it regulates the modalities for labeling deepfakes and texts generated or manipulated by AI when intended to inform the public on matters of general interest.
To this end, the document proposes the adoption of a common taxonomy for classifying content falling within the notion of deepfakes or AI-generated or manipulated textual publications (pursuant to Article 50(4)), in particular distinguishing between the following categories of content:
- content entirely generated by AI, including content generated fully and autonomously by the artificial intelligence system, without the contribution of authentic content created by a human being;
- AI-assisted content, including content characterized by a mixed involvement of humans and artificial intelligence.
The adoption of a shared icon is also suggested, suitable to clearly and distinguishably signal the fully artificial or assisted nature of the content. The icon must be perceivable at the user’s first contact with the content and adapted to different types of media and contexts of use, including video, audio, text, and multimodal content.
Figure 1: zero-shot prompt
As specified in the Code, the icons shown are for illustrative purposes only and will be further developed during the definition of the Code of Conduct. By way of example, the figure above shows a zero-shot prompt, updated to December 2025, used to generate an icon containing the word “AI” in two different colors, indicating the distinction between fully automated content and content assisted by artificial intelligence.
Disclosure must in any case be clear, timely, and accessible, in accordance with European Union accessibility requirements, including alternative solutions for users with visual or hearing disabilities, such as audio disclosure, alternative text, and adequate color contrast.
Management of deepfakes and publications of public interest
For deepfakes, the Code establishes specific disclosure modalities depending on the type of content. For example, in real-time videos, the icon must be visible and accompanied by an initial disclaimer. In non-real-time videos, the icon must appear consistently during viewing, possibly accompanied by initial or final disclaimers. In audio content, disclosure must take place through audible notices at the beginning, at intermediate intervals, and at the end, and the icon may be shown where a screen is available.
An element of particular importance is the attention to proportionality. The Code expressly recognizes that, in the case of artistic, creative, satirical, or fictional works, transparency obligations must be applied in a way that does not compromise the enjoyment or quality of the work.
Similarly, for texts generated or manipulated by AI and published on topics of public interest, deployers must apply internal procedures suitable to ensure correct identification of content and timely disclosure.
The Code provides that the icon be placed in a fixed and clearly distinguishable manner, for example at the beginning, alongside, or at the end of the text. Any exemption from the disclosure obligation is permitted only where the content is subject to human review or editorial control.
Conclusions
Overall, the first draft of the Code of Conduct on the transparency of AI-generated content, although still a document in evolution, offers concrete guidance on regulatory expectations and on the technical and organizational tools deemed suitable to ensure a transparent and responsible use of artificial intelligence.
For companies, rights holders, and innovation operators, the Code constitutes an essential first reference point to guide compliance strategies, technological development, and the safeguarding of content reliability in the AI era.
Teresa Franza