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

The overarching objective of this Code of Practice (“Code”) is to improve the functioning of the internal market, to promote the uptake of human-centric and trustworthy artificial intelligence (“AI”) and to support innovation pursuant to Article 1(1) AI Act, while ensuring a high level of protection of health, safety, and fundamental rights enshrined in the Charter, including democracy, the rule of law, and environmental protection, against harmful effects of AI in the Union.

To achieve this overarching objective, the specific objectives of this Section of the Code are:

a) to serve as a guiding document for demonstrating compliance with the obligations provided for in Article 50(2) and (5) AI Act, while recognising that adherence to the Code does not constitute conclusive evidence of compliance with these obligations;

b) to assist providers of AI systems generating synthetic audio, image, video or text content (“generative AI systems”) in complying with their obligations under Article 50(2) and (5) AI Act and

c) to enable the competent market surveillance authorities to assess such compliance in relation to providers who rely on the Code to demonstrate compliance with these obligations under Article 50(2) and (5) AI Act in a consistent, predictable and uniform manner across the Union.

Recitals Whereas:

a) Trust in the information ecosystem: Signatories recognise that AI systems can generate and manipulate large quantities of synthetic content and that it is becoming increasingly difficult for humans to distinguish this content from human-authored content. Such developments impact the integrity of the information ecosystem as well as the trust people have in it, and raises new risks of misinformation, manipulation at scale, fraud, impersonation, and consumer deception. Signatories recognise that transparency is fundamental to fostering trust in and integrity of the ecosystem, and to ensuring that AI systems remain reliable and trustworthy.

b) Technical solutions for marking: While at the time of publication of this Code several techniques are readily available for marking AI-generated or manipulated audio, image, video, or text content, Signatories acknowledge that in most cases, in particular for content that can be disseminated online, no single marking technique suffices to meet the four requirements in Article 50(2) AI Act, namely effectiveness, interoperability, robustness, and reliability. In such cases, Signatories recognise that, as acknowledged in scientific reports and in line with the state of the art, only an appropriate combination of marking techniques

and associated detection mechanisms can allow satisfaction of those requirements in a holistic manner, in so far as this is technically feasible, taking into account the specificities and limitations of the content, the cost of implementation, and the existence of relevant standards. The Code promotes innovation and future-proofness in a fast-evolving technological landscape by allowing Signatories to rely on alternative techniques, or even a single technique, as long as they can prove compliance with the four requirements in Article 50(2) AI Act and this Section of the Code, based on verifiable results and common benchmarks, where available.

c) Cooperation along the value chain: Signatories recognise that effective, interoperable, robust, and reliable technical solutions for marking and detection require investment of time and resources. Signatories therefore acknowledge the need for practical arrangements to make detection solutions accessible, as appropriate, and for facilitating cooperation with other actors along the value chain to enable the public to effectively identify AI-generated or manipulated content. Signatories who are providers of generative AI models play an important role in this value chain and are therefore encouraged to facilitate compliance for downstream providers of generative AI systems built on those models.

d) Advancing innovation in marking techniques and detection mechanisms: Signatories acknowledge that determining the most effective, robust, reliable, and interoperable technical methods for marking and detection remains an evolving challenge. Signatories recognise that this Section of the Code promotes innovation in various ways and encourages advancing the state of the art in AI transparency techniques and related processes and measures. For instance, the Code allows (i) innovative techniques for AI transparency, (ii) new means to facilitate interoperability among different marking techniques and associated detection mechanisms, and (iii) the development of benchmarks to evaluate the performances of the technical solutions implemented by providers of generative AI systems. Signatories are therefore encouraged to invest time and resources, as appropriate, to contribute to the advancement of the state of the art.

e) Cooperation with other stakeholders: Signatories recognise the advantages of collaborative efficiency, for example through sharing methods and/or infrastructure and relying on open standards and marking techniques implemented at the model level or provided by other third parties. Signatories further recognise the importance of enabling relevant third parties and end-users to detect marked content, and of engaging expert or lay representatives of civil society, academia, and other relevant stakeholders in understanding technical solutions for AI transparency. Signatories recognise that such cooperation may involve providing the necessary support for the smooth integration of third-party models and marking solutions into the AI system’s inference process, and entering into agreements to share information relevant to technical solutions, while ensuring proportionate protection of sensitive information and compliance with applicable Union law. Signatories also acknowledge that end users’ feedback on their implementation of marking and detection techniques is a valuable source of information in view of continuously improving its implementation. Signatories further recognise the importance of cooperating with market surveillance authorities and of fostering collaboration between providers of generative AI systems and models, deployers and other users of generative AI systems, online platforms, search engines, researchers, civil society and regulatory bodies to address emerging challenges and opportunities in transparency of AI-generated or manipulated content.

f) Promoting standardisation: Signatories recognise the value of open standards to facilitate interoperability and lower the cost of marking techniques and associated detection mechanisms for AI transparency. They recognise that further efforts will be required for such standards to emerge from international and European standard-setting organisations, considering the implementation challenges and the fast-evolving field. Signatories are encouraged to participate in relevant standardisation activities in order to accelerate the consolidation of marking techniques and associated detection mechanisms and their interoperability. In particular, Signatories recognise that metadata marking standards need to be further elaborated to provide richer information for AI transparency, as well as recognising the need for relevant watermarking standards both at technology and interface levels.

g) Proportionality for Small and medium enterprises (“SMEs”) and small mid-cap enterprises (“SMCs”): To account for differences between providers of generative AI systems regarding their size and capacity, simplified ways of compliance for SMEs and SMCs, including startups, should be possible, in a proportionate manner.

Commitments This Section of the Code applies to Signatories who are providers of generative AI systems falling within Article 2 and Article 50(2) AI Act, taking into account the scope of the legal obligations and relevant exceptions, as applicable.

Without prejudice to the primary responsibility of providers of generative AI systems under Article 50(2) AI Act, providers of generative AI models that are placed on the Union market independently from AI systems and third-party providers of marking and detection solutions may also, on a voluntary basis, adhere to the Code to demonstrate that their marking and detection solutions comply with the requirements of Article 50(2) and (5) AI Act and this Section of the Code, with a view to facilitating compliance by downstream providers of AI systems built on those models and/or using third parties’ marking and detection solutions. In such cases, the commitments and the measures of this Section of the Code shall be understood as applying to the Signatory’s generative AI model and/or marking and detection solution as appropriate.

In this Section, the terms “will”, “encouraged” and “may” should be interpreted in the following way:

1. ‘will’ is used for mandatory measures under the Code that need to be met for the Signatory to be compliant with Article 50(2) and (5) AI Act, and for which compliance will be monitored by competent market surveillance authorities;

2. ‘encouraged’ is used for optional measures under the Code that are not legally required and are purely voluntary, but are nevertheless recommended;

3. ‘may’ is used for optional measures under the Code that are not legally required and are purely voluntary or provide providers with flexibility on how to implement the respective measure or commitment.