Code of Practice on Transparency of AI-Generated Content (marcatura ed etichettatura dei contenuti generati o manipolati dall'IA)

Prior to placing their generative AI system on the market or putting it into service, and regularly thereafter, Signatories will test the compliance of their marking and detection solutions with the requirements and the measures specified in this Section of the Code. This will be carried out under conditions that are representative of real-world use, with the objective of ensuring alignment with the evolving state of the art. Signatories who are downstream providers of generative AI systems may rely on the results of testing performed by an upstream model provider or a third-party provider of marking and detection solutions, so long as those solutions comply with the requirements specified in this measure and without prejudice to the ultimate responsibility of the Signatory to ensure compliance with Article 50(2) and (5) AI Act. Until recognised performance evaluation methods and benchmarks for marking techniques and detection mechanisms emerge, in particular methods and benchmarks endorsed by the AI Office after consultation of the AI Board and the taskforce described in Measure 3.5, Signatories will test and report on the performances of their solution according to internal benchmarks and industry best practices. Signatories may involve independent experts in the testing of their solutions, especially to evaluate robustness to adversarial attacks through red teaming and/or conduct such testing and evaluation under regulatory supervision in the context of AI regulatory sandboxes, as provided for in Article 57 AI Act. Signatories will monitor any substantiated compliance shortcomings identified internally or reported by relevant third parties and implement appropriate follow-up corrective actions.