Parte 3
(18) Article 50(2) AI Act sets out a transparency obligation for providers of generative AI systems comprising three distinct, but inherently, interlinked elements. According to that provision, providers of generative AI systems must ensure:
• that the outputs of the AI system are marked in a machine-readable format.
• that the outputs are detectable as artificially generated or manipulated.
• that the combination of the technical solutions for marking and detection meet the requirements of effectiveness, interoperability, robustness and reliability, insofar as this is technically feasible, taking into account the specificities and limitations of various types of content, the costs of implementation and the generally acknowledged state of the art, as may be reflected in relevant technical standards.
(19) Article 50(5) AI Act further specifies that the information referred to under Article 50(2) AI Act – notably the result of the detection – should be provided to the natural persons exposed to the content in a clear and distinguishable manner at the latest at the time of the first interaction or exposure.
(20) The Commitments and measures set out in Section 1 of the Code cover each of the aforementioned elements of the obligation laid down in Articles 50(2) and (5) AI Act and facilitate their effective implementation by providers of generative AI systems as follows.
(21) Commitment 1 facilitates the effective implementation of the marking element of the transparency obligation with commitments from Signatories to implement a marking solution with regard to the audio, image, video or text content, or any combination thereof, generated or manipulated by the AI system which is placed on the market or put into service in the Union. Measure 1.1 specifies the machine-readable marking techniques that providers of generative AI systems may rely on to comply with that obligation, including a multi-layered approach of at least two layers (metadata and watermarks) for increased robustness aligned with the state of the art and relevant standards, while allowing flexibility for a single layer approach if an equivalent level of compliance can be demonstrated by the provider of the generative AI system. Measure 1.2 outlines the commitments for markings not to be removed. This measure
EN 4 EN
aims to advance the objective of effective marking, including through cooperation with other stakeholders in the value chain to preserve the marks.
(22) Commitment 2 facilitates the effective implementation of the detectability element of the transparency obligation with commitments from Signatories to ensure that there are means to enable the detection of the machine-readable markings present in audio, image, video or text content, or a combination thereof, as generated or manipulated by their AI system. Measure 2.1 clarifies the requirements for the detection solution with regard to their availability, accessibility, personal data and security protection, and withdrawal. This is supported by recital 135 AI Act, which clarifies that the Code’s purpose is to support practical arrangements for making, as appropriate, the detection mechanisms accessible and to enable the public to effectively distinguish AI-generated content from other content.
(23) Commitment 2 also facilitates the disclosure element of the transparency obligation with commitments from Signatories to ensure that the detection results are provided to natural persons that may be exposed to the content in a clear, distinguishable and accessible manner. Measure 2.3 further specifies this commitment with regards to the specific measures to implement clear and accessible disclosure of detection results to all natural persons in accordance with Article 50(5) AI Act.
(24) Commitment 3 facilitates the effective implementation of the requirements for marking and detection solutions specified in Article 50(2) AI Act. Measure 3.1 details how to implement and assess in practice marking and detection solutions that are effective. Measure 3.2. details how to implement and assess in practice marking and detection solutions that are reliable. Measure 3.3 details how to implement and assess in practice marking and detection solutions that are robust. Measure 3.4 sets out a staged approach to the implementation of the watermarking and detection solutions that are interoperable. The latter measure includes clear commitments for the use of available open metadata standards, starting from the entry into application of the rules, and a minimum level of interoperability solutions for the watermarking detection mechanism to be implemented by 2 February 2027 to reduce the burden on providers of AI systems and enable a more seamless access for users. Such an approach constitutes a pragmatic step forward in a context where the state of the art does not yet support more mature and standardised solutions for watermarking and where the current ecosystem is fragmented and requires a certain level of protection to preserve the robustness of watermarking solutions. Furthermore, Signatories commit to cooperate towards advancing interoperability of watermarking and detection solutions and to implement improved and state of the art interoperability solutions and standards in a timely manner once they emerge. The Commission will monitor the implementation of this measure under the Code and reserves its right, in case of a lack of progress, to consider taking appropriate measures.
(25) Commitment 4 facilitates the effective implementation of the transparency obligation under Articles 50(2) and (5) AI Act as a whole with commitments from Signatories to set up, keep up to date, and implement compliance, testing, verification, and monitoring processes to ensure the effective implementation of and to demonstrate compliance with the obligations set out in that provision and the commitments and the measures envisaged in Section 1 of the Code. Measure 4.1 specifies that Signatories will document, implement, and keep up to date, in line with the state of the art, a compliance process that describes, at a high-level, how they have implemented the different
EN 5 EN
measures of Section 1 of the Code. Measure 4.2 clarifies ways for testing, verification, and monitoring to ensure continuous compliance of the marking and detection solutions before and after the AI systems are placed on the market or put into service in the Union. Measure 4.3 covers appropriate training for personnel with roles relevant to ensure compliance with Articles 50(2) and (5) and Article 4 AI Act. Measure 4.4 addresses cooperation with market surveillance authorities. The implementation of the measures may rely on existing processes and documentation and take into account the size and available resources of Signatories, which ensures proportionality and simplification, without necessarily affecting effectiveness. Commitment 4 is motivated by recital 135 AI Act, which clarifies that the Code’s purpose is to facilitate cooperation with other actors along the value chain.
(26) Section 1 of the Code further contributes to the effective implementation of Article 50(2) AI Act by suggesting optional measures that are not required to ensure compliance with that provision and are therefore voluntary, but are nevertheless recommended. Those measures are conducive to achieving the goals of Article 50(2) AI Act, since the purpose of the Code, as clarified by recital 135 AI Act, is to enable natural persons to effectively distinguish AI-generated content. Measure 1.3 encourages providers of AI systems to provide provenance information and richer metadata that will help natural persons further obtain relevant information on the nature of the content and will increase trust in the information ecosystem. Measure 1.4 encourages providers to provide an optional functionality for a perceptible label that will also facilitate compliance for downstream deployers of AI systems subject to labelling obligations. Measure 2.2 opens the possibility for providers to include a forensic detection mechanism as part of their detection solution to address cases where metadata has been stripped and watermarks removed. Measure 2.4 supports literacy on AI marking and detection solutions that are essential for the broader effectiveness of the transparency obligations and the ability of natural persons to effectively distinguish AI-generated content from other content. Measure 3.5 encourages providers to invest in advancing the state of the art of marking and detection solutions to improve robustness and the effectiveness of those solutions. All these measures are also motivated by recital 135 AI Act, which clarifies that the Code’s purpose is to facilitate cooperation with other actors along the value chain, since these measures encourage cooperation with model providers, downstream deployers, end-users, civil society organisations, academia, independent researchers, providers of online platforms and search engines, market surveillance authorities, and other regulators.
(27) Furthermore, Section 1 of the Code allows providers not covered by the transparency obligation laid down in Article 50(2) AI Act, such as providers of AI models and third- party providers of marking techniques and detection mechanisms that are provided independently from the AI system, to adhere to the Code as specified in the introduction to the Commitments. Those providers can also adhere on a voluntary basis and rely on the Code to demonstrate that their AI model or marking and detection solutions comply with Articles 50(2) and (5) AI Act, thus facilitating effective implementation and compliance for downstream providers of generative AI systems. The Code clarifies that references to Signatories throughout this Section of the Code shall be understood accordingly and that the commitments and the measures of Section 1 of the Code are applicable respectively to the Signatory’s generative AI model or marking and detection solutions. In doing so, the Code effectively achieves the objective set out in recital 135 AI Act, which is to facilitate cooperation with other actors along the value chain, and
EN 6 EN
Link a questa unità: https://fwks.it/news/normativa/testi/parere-commissione-codice?u=parte-3