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

Signatories will implement marking and detection solutions which, in conjunction, achieve a high level of reliability in different expected contexts and across use cases, to the extent technically feasible and in alignment with the state of the art. Reliability refers to the capability of the marking and detection solutions to accurately identify and distinguish the origin of AI-generated or manipulated content from other content. The reliability assessment consists of two components: (i) how accurate the detection solution is with regard to detecting the marking in nominal conditions where there is no alteration to the content, and (ii) how the accuracy of the marking and detection solutions varies with respect to the length, size, diversity, and semantics of the content. Signatories will measure the accuracy of the detection of AI-generated or manipulated content by using relevant metrics for the type of markings, such as the error rate of the detection. Low error rates will be demonstrated on a variety of samples, including both content generated or manipulated by their AI system with their own marking and other content (possibly including content generated or manipulated by other AI systems). The data for the assessment of reliability shall, to the extent possible, not have been used during the training and development of the AI system. Signatories will ensure these accuracy measurements are performed and reported with content intended to be generated or manipulated by the AI system of various length, size, diversity, and semantics, in order to demonstrate that the performance of the implemented marking and detection solutions generalises across diverse contexts and use cases.