AI Act: The Commission clarifies the transparency requirements of Section 50
Section 50 of the IA Act
As a reminder, Section 50 imposes transparency obligations on providers and operators of certain AI systems, requiring that users be informed when they interact with an AI system or when content is generated by AI. These obligations extend to providers and operators of open-source AI systems, who are not eligible for any exemptions.
This article establishes transparency obligations in four cases:
- when AI interacts directly with people
- when AI generates synthetic content
- when AI is used for emotion recognition or biometric classification
- when AI generates deepfakes or publishes texts on topics of public interest.
These obligations apply:
- to all AI systems used in the four situations mentioned above, not just high-risk systems
- effective August 2, 2026, subject to the adoption of the political agreement reached by the Council and the Parliament on May 7, which would formalize a postponement of this deadline
The Commission has published draft guidelines proposing a practical interpretation of Section 50. This important document specifies in concrete terms:
- The precise categories of content subject to the labeling requirement
- The definition of a “deepfake”
- What constitutes AI-generated or AI-manipulated content
- The obligations of providers and distributors
- Applicable exceptions
- Expected technical criteria: effective, robust, reliable, and interoperable solutions
The document also confirms several key policy directions:
- Multi-layered protection: mandatory combination of digital watermarking, metadata, source traceability, and detection tools
- Regulation of deepfakes—visual, audio, video, and textual montages (excluding short-form content)
- Accountability of AI agents whenever their outputs or interventions enter the realm of human perception
- Requirement to adopt solutions capable of withstanding file modifications, successive reposts, and malicious attempts to circumvent (attacks)
These guidelines should be distinguished from the upcoming Code of Conduct (to be published in June):
- The Code of Conduct proposes a voluntary operational framework
- The Commission’s Guidelines detail the Commission’s doctrinal interpretation
Expert Opinion
Although not legally binding, the Commission’s guidelines effectively dictate the enforcement and sanctioning policy that regulatory authorities would apply. Companies are advised to prepare now, rather than waiting until August 2, 2026.
Firsh helps companies conduct an audit of their AI solutions and assists them in ensuring compliance with these guidelines and the future code of conduct in order to benefit from a protective presumption of compliance.
