Multi-layer marking of AI-generated content
Content generated or manipulated by AI systems must be marked in a machine-readable manner and must be easily identifiable as artificial content. To this end, providers will adopt a multi-layer marking approach, considering that, given the current state of the art, no single solution is sufficient on its own.
In practice, providers will be required to apply combined technical marking, using different techniques jointly and adapted to all types of content, to incorporate marking from the model training stage, to ensure that marks are not altered by subsequent uses or transformations, and to maintain content traceability whenever technically feasible.
Detection of AI-generated or AI-manipulated content
Content generated or manipulated by AI systems must be effectively detectable as such. This requires providers not only to mark content, but also to provide accessible mechanisms that allow its origin to be verified by both users and third parties.
To this end, providers must offer accessible and free detection tools that enable verification of the origin of AI-generated or AI-manipulated content, provide detection mechanisms prior to commercialisation, and implement forensic detection mechanisms for generative AI models that may be used or integrated into subsequent generative AI systems.
The results of these tools must be understandable and accessible to individuals and must remain available throughout the entire lifecycle of the system.
Requirements for marking and detection techniques
The marking and detection techniques used by providers of AI systems must be effective, reliable, robust and interoperable, taking into account the state of the art, implementation costs and the characteristics of each type of content. In addition, they must be applicable prior to the commercialisation of the system and throughout its entire lifecycle.
The Code also promotes the continuous improvement of these techniques through investment in research and collaboration with other stakeholders for the development of more advanced solutions, such as new watermarking technologies and forensic detection methods.
Testing, verification and compliance
Marking and detection obligations must not remain merely theoretical. An internal compliance framework must therefore be established, accompanied by testing, controls and supervisory mechanisms that make it possible to effectively demonstrate compliance with the AI Act.
This internal compliance framework must be based on the documentation of processes and measures applied for the marking and detection of AI-generated content, the provision of training to the personnel involved, and the information necessary to demonstrate compliance to the authorities.
Likewise, solutions must be subject to periodic testing and verification under real-world conditions, accompanied by continuous supervision and updates in order to address risks, failures or attacks.
Rules on the labelling of deepfakes and AI-generated or AI-manipulated texts applicable to AI system deployers
Disclosure of origin through a common taxonomy and icon
When publishing and disseminating content generated by AI systems, deployers must inform recipients when AI is used to generate or manipulate deepfakes or certain texts of public interest. The core objective is to harmonise disclosure across the EU through (i) a common taxonomy and (ii) a common icon, so that the public can easily recognise the origin of the content.
To this end, the Code distinguishes between content fully generated by AI and AI-assisted content, in order to calibrate the risk of deception for the public. This classification must be communicated in a clear, visible and consistent manner from the first exposure, through an identifiable icon, initially provisional and based on an acronym, which will in the future evolve into an EU-wide interactive icon allowing access to more detailed information on the involvement of AI.
Deployers must ensure that visual and audio elements comply with requirements regarding contrast, size and compatibility with screen readers, as well as the inclusion of acoustic signals for persons with visual impairments, audio descriptions for visual indicators, visual or tactile signals for audio-only content, and high-contrast icons compatible with assistive technologies.
Compliance, training and supervision
As with providers of AI systems, deployers’ internal processes must integrate documented transparency obligations and measures clearly describing how labelling requirements and the specific obligations applicable to deepfakes and AI-generated or AI-manipulated texts are applied.
In addition, the importance of training personnel involved in the creation, modification or dissemination of content subject to these transparency obligations is reinforced.
Specific measures for the disclosure of deepfakes and AI-generated or AI-manipulated texts
With regard to deepfakes, deployers must establish coherent internal identification and classification processes, based on the definition set out in the AI Act and subject to human oversight, taking into account the target audience, distribution channels and the possible application of exceptions, such as creative works.
They must also ensure clear and distinguishable disclosure from the first exposure of the content, through the use of an icon and, where appropriate, additional notices adapted to the content format (real-time or deferred video, images, audio or multimodal combinations). In the case of artistic, satirical or fictional works, disclosure must be applied in a proportionate and non-intrusive manner, preserving the user experience and establishing appropriate safeguards to protect the rights and freedoms of the individuals represented or of the public in general.
With regard to AI-generated or AI-manipulated texts, deployers must establish coherent internal identification processes, supported by human oversight, to determine when a text is subject to disclosure and when a legal exception applies. In cases subject to transparency obligations, the artificial origin of the text must be indicated in a clear, fixed and easily perceptible manner from the first exposure, through the visible placement of an icon within the content itself.
However, an exception is recognised where effective human review and the assumption of editorial responsibility can be demonstrated. This requires maintaining minimum documentation on the review process, the responsible person and the final published version, in a manner proportionate to the size and resources of the organisation.
↳ Discover more content in the section Insight.
↳ Do you want to stay up today about the sector related news? Follow us on LinkedIn.
