§3.3 Mitigation measures linked to generative AI

36. Recent technological developments in generative AI have enabled the creation and widespread use of artificial intelligence systems capable of generating text, images, videos, or other synthetic content. While such developments may bring many new opportunities, they may lead to specific systemic risks in the context of elections. Notably, generative AI can be abused to mislead voters or to manipulate electoral processes by creating and disseminating inauthentic, biased, misleading synthetic content (including text, audio, still images and video) regarding political actors, false depiction of events, election polls, contexts or narratives. Generative AI systems can also produce incorrect, incoherent, or fabricated information, so called “hallucinations”, that misrepresent reality, and which can potentially mislead voters. 37. Pursuant to Article 35(1) of Regulation (EU) 2022/2065, providers of VLOPs and VLOSEs should assess systemic risks in the context of elections, such as those mentioned above, and put in place reasonable, proportionate, and effective mitigation measures tailored to risks related to both the creation and dissemination of generative AI content, depending on the nature of their service. Best practices which may inform the relevant risk mitigation measures may be drawn already now from the AI Act and the AI Pact. Particularly relevant in this context are the obligations envisaged in the AI Act for providers of general-purpose AI models, including generative AI, requirements for labelling of ‘deep fakes’ and for providers of generative AI systems to use technical state-of-the-art solutions to ensure that content created by means of generative AI is marked in a machine-readable format and detectable as such, which will enable its detection by providers of VLOPs and VLOSEs. The Commission recommends that providers of VLOPs and VLOSEs align their policies to the AI Act in advance of its entry into application, in line with the AI Pact. 38. Mitigation measures linked to generative AI should be applied by providers of VLOPs and VLOSEs, to the extent that they are technically feasible, with particular consideration to the impacts of such measures on fundamental rights protected under the Charter. The Commission also recommends cross-industry collaboration to further develop effective mitigation measures for generative AI in the context of the Codes of Practice for general-purpose AI models and transparency of AI-generated content to be developed under the AI Act. 39. Following from the specific actual or foreseeable risks for electoral processes identified, the Commission recommends that providers of VLOPs and VLOSEs whose services can be used for the creation of deceptive, biased, false or misleading generative AI content have the following risk mitigation measures in place, to the extent this is technically feasible, and according to the current state of the art: a) Ensure that generative AI content, and other types of synthetic and manipulated media, is detectable – notably by using sufficiently reliable, interoperable, effective and robust techniques and methods, such as watermarks, metadata identifications, cryptographic methods for proving provenance and authenticity of content, logging methods, fingerprints or other techniques, as may be appropriate, taking into account existing standards. This is particularly important for any generative AI content concerning candidates, politicians, or political parties. Watermarks and metadata may also be applied to content that is based on initially authentic footage (such as videos, images or audio) that subsequently has been altered through the use of generative AI; b) Make efforts to ensure that information generated by AI systems is based to the extent possible on reliable sources in the electoral context, such as official information from relevant electoral authorities, and that any quotes or references made by the system to external sources are accurate and do not misrepresent the cited content, thus limiting the effects of ‘hallucinations’; c) Warn users about potential errors in content created by generative AI systems and suggest that they consult authoritative sources to check the veracity of such information, as well as put safeguards in place to prevent the creation of false content that may have a strong potential to influence user behaviour; d) Conduct and document red-teaming exercises and testing with a particular focus on electoral processes, with both internal teams and external experts, before releasing generative AI systems to the public and follow a staggered release approach when doing so to better control unintended consequences; e) Set appropriate performance metrics, including for safety and factual accuracy of answers given to questions on electoral content, and continually monitor the performance of generative AI systems, and take appropriate actions when needed; f) Integrate into generative AI systems safeguards that increase their safety, such as prompt classifiers, content moderation and other filters, to detect and prevent prompts that go against terms of service of the provider of a VLOP or VLOSE concerning electoral processes; take other appropriate measures that seek to prevent the misuse of the generative AI system for illegal, manipulative and disinformation purposes in the context of electoral processes; g) For text content, in particular, indicate, where possible, in the outputs generated the concrete sources of the information used as input data or provide other means to enable users to verify the reliability and further contextualise the information. 40. Following from the specific actual or foreseeable risks for electoral processes identified, the Commission recommends that providers of VLOPs and VLOSEs whose services can be used to disseminate deceptive, false or misleading generative AI content consider the following risk mitigation measures, to the extent technically feasible according to the current state of the art: a) Adapt their terms and conditions and ensure their enforcement, to significantly decrease the reach and impact of generative AI content that depicts disinformation or misinformation on the electoral process, such as election irregularities. i. The Commission recommends that providers of VLOPs and VLOSEs provide clear public information on which internal processes and mitigation measures, such as labelling, marking, demoting or removing, are in place to enforce these policies; ii. The Commission recommends that providers of VLOPs and VLOSEs cooperate and share information about such deceptive content with fact checkers to ensure that the risk of amplification in other platforms is minimised. b) Clearly label, or otherwise make distinguishable through prominent markings, synthetic or manipulated images, audio or videos that appreciably resemble existing persons, objects, places, entities, events, or depict events as real that did not happen or misrepresent them, and falsely appear to a person to be authentic or truthful (i.e., deepfakes). i. The Commission recommends that providers of VLOPs and VLOSEs provide users with standard and easy to use interfaces and tools to add labels to AI generated content; ii. When labelling generative AI content, the Commission recommends that providers of VLOPs and VLOSEs apply efficient labels, easily recognised by users, taking into account aspects such as graphics, position and timing, drawing on scientific research on the effectiveness of labels (31). Providers of VLOPs and VLOSEs should also test the effectiveness of such labels before release and adapt and improve them based on feedback and experience; iii. The Commission recommends that providers of VLOPs and VLOSEs make sure the labelled generative AI content retains its label once it is shared by other users on the platform. c) The Commission recommends that providers of VLOPs and VLOSEs adapt their advertising systems, for example by providing advertisers with options to clearly label content created with generative AI in advertisements or promoted posts and require in their advertising policy that this label is used when the advertisement includes generative AI content. d) To enforce these policies, providers of VLOPs and VLOSEs should adapt their content moderation processes and algorithmic systems in such a way as to detect AI generated or manipulated content via watermarks, metadata identifications, cryptographic methods for proving provenance and authenticity of content, logging methods, fingerprints or other techniques. i. In this context, providers of VLOPs and VLOSEs should cooperate with providers of generative AI systems and follow leading state of the art measures to ensure that such watermarks, metadata identifications, cryptographic methods for proving provenance and authenticity of content, logging methods, fingerprints or other techniques are detected in a reliable and effective manner; they are also recommended to support new technology innovations to improve the effectiveness and interoperability of such tools. e) Media literacy measures mentioned in section 2 should also focus on generative AI systems, for instance to explain how the technology works and the possibilities for its misuse. 41. Pursuant to Article 35(1) of the Regulation (EU) 2022/2065, when providers of VLOPs and VLOSEs address legal but harmful forms of generative AI content that can influence voters’ behaviour, they should give particular consideration to the impact their policies and measures may have on fundamental rights, notably freedom of expression, including political expression, parody and satire. Such a fundamental rights impact assessment is in particular required when developing policies on what type of deceptive generative AI content a provider of a VLOP or VLOSE does not allow on their service and will remove from it. 42. As AI generated content bears specific risks, it should be subject to heightened scrutiny, also through the development of ad hoc tools and technologies, e.g., to perform research aimed at identifying and understanding specific risks related to electoral processes. Providers of VLOPs and VLOSEs are encouraged to consider setting up dedicated tools for researchers to get access to and specifically identify and analyse AI generated content.

← 3.2.3 · All articles · 3.4 →

Source: EUR-Lex (Cellar) · retrieved 2026-09-07