Conclusion of the Artificial Intelligence Board on the assessment of the Code of Practice on Transparency of AI-generated Content (Article 50(7) AI Act)

Commitment 1 – Marking of AI Generated or Manipulated Content

Commitment 1 aims to cover the marking element of the Article 50(2) obligation. Signatories commit to implementing a machine-readable marking solution for audio, image, video or text content generated or manipulated by their AI systems placed on the market or put into service in the Union.

Measure 1.1 provides for a multi-layered approach combining at least two layers of metadata and watermarking, aligned with the state of the art and relevant standards, while allowing flexibility for a single-technique approach where equivalent compliance can be demonstrated. In the case of free-form text, where a multi-layered approach is not considered technically feasible, the Measure considers that a single layered approach is acceptable. The provision that a single layer of marking will be considered sufficient for generative AI systems embedded in physically controlled and closed environments may raise concerns regarding its susceptibility to circumvention. While this concern warrants attention, it is noted that the measure is framed, with a view to focusing on content which might pose more limited risks of deception, and in theory is subject to safeguards. The Board concludes therefore this exception does not at this stage prevent the Code from meeting its objectives.

Measure 1.2 commits Signatories to preserving marks throughout the value chain, including through inclusion of relevant provisions for downstream deployers and other stakeholders without requiring downstream policing.

Commitment 2 – Detection of Markings of AI-Generated or Manipulated Content

Commitment 2 aims to cover the detectability element of the obligation, requiring Signatories to provide the means to enable detection of the machine-readable markings present in the outputs of their AI systems.

Measure 2.1 sets out the requirements for detection solutions with regard to availability, accessibility, data protection and security. The Board acknowledges that a fragmented detection landscape may create practical difficulties for competent market surveillance authorities and other actors, thus recognising that compliance with the staged implementation of interoperability requirement will be essential for the effective implementation of the marking and detection obligations (see Measure 3.4. below). The Board considers that sub-Measure 2.1.4 does not impose any obligation on market surveillance authorities to run or maintain legacy detection systems, databases, or related infrastructure, that have been retired by Signatories.

Measure 2.3 sets out requirements to cover the disclosure element of the obligation, which specifies how detection results must be provided to natural persons in a clear, distinguishable and accessible manner in accordance with Article 50(5) AI Act.

Commitment 3 – Measures to Meet the Requirements for Marking and Detection Solutions

Commitment 3 aims to address the technical requirements applicable to marking and detection solutions under Article 50(2) AI Act.

Measures 3.1, 3.2 and 3.3 provide detailed specifications for implementing and assessing solutions that are respectively effective, reliable and robust.

Measure 3.4 introduces a staged approach for implementation of the interoperability requirements with commitments for use of available open metadata standards and minimum interoperability solutions for the watermarking detection mechanism to be implemented by 2 February 2027. In light of the fragmented state of the art at the time of application of Article 50(2) AI Act, the Board considers this approach pragmatic. It has the potential to deliver, over time, solutions that meet the interoperability and effectiveness objectives of the AI Act, provided that Signatories make the serious efforts to implement the commitments and help improve the interoperability. The Board will carefully examine the interoperability solutions that will emerge from Signatories by 2 February 2027, including to the extent to which they are implemented and contribute to improved interoperability between watermarking detection mechanisms as outlined under Measure 3.4.

Commitment 4 – Testing, Verification, and Compliance

Commitment 4 covers the processes that Signatories must put in place to ensure and be able to demonstrate continuous compliance with the marking and detection obligations across the lifecycle of their systems.

Measure 4.1 requires Signatories to document and maintain a compliance process describing how they have implemented the measures of Section 1.

Measure 4.2 sets out requirements for testing, verification and monitoring before and after market placement.

Measure 4.3 covers training of relevant personnel, and Measure 4.4 addresses cooperation with market surveillance authorities. The implementation of these measures can rely on existing processes and documentation, taking into account the size and available resources of Signatories, which ensures proportionality without necessarily affecting effectiveness.

Overall Assessment of Section 1