High-risk § 3.8.2

Point (8)(b) of Annex III lists as high-risk AI systems intended to be used for (i) influencing the outcome of an election or referendum or (ii) the voting behaviour of natural persons in the exercise of their vote in elections or referenda. That provision clarifies that this use case does not include AI systems the output of which natural persons are not directly exposed to, such as tools used to organise, optimise or structure political campaigns from an administrative or logistical point of view. Recital 62 AI Act explains that the classification of AI systems falling within this use case as high-risk aims to address the risks of undue external interference with the right to vote enshrined in Article 39 of the Charter, and of adverse effects on democracy and the rule of law.

For an AI system to be classified as high-risk under the use case listed in point 8(b) of Annex III AI Act, two conditions must be fulfilled : first, the intended purpose of the AI system must be to influence the outcome of an election or referendum, or to influence the voting behaviour of natural persons in the exercise of their vote in elections or referenda ; second, natural persons must be directly exposed to the outputs of the AI system. These conditions and the concepts necessary to interpret them are analysed in the subsections below.

Only AI systems intended to be used to influence the outcome of an election or referendum or the voting behaviour of natural persons, i.e. the AI system would have to be specifically intended to have an effect on the electorate’s choice or turnout, fall within the scope of the use case listed in point 8(b) of Annex III. The focus of the use case is thus on underlying objectives: the system must be directed toward s electoral influence, not merely be incidentally capable of it. Therefore, AI systems with non-intentional influence on the outcome of an election or referendum or the voting behaviour of natural persons , such as general content recommender systems that , amongst other content, also recommend political content and thus, de facto, have the potential to influence political opinions, should not be classified as high-risk. General-purpose AI systems, which offer sufficient safeguards against a use influencing electoral processes, e.g. a chatbot clearly replying that it cannot give any voting advice when asked by a user, or providing only neutral, factual, and informational content about elections (voting procedures (how, when, where), registration procedures; the institutional and legal framework; objective information on parties and candidates (e.g. references to their official websites); and general civic education on democratic processes) would also not fall under high-risk category as not being intended to influence the outcome of an election or referendum or the voting behaviour of natural persons.

That the AI system should be intended to be used to influence the outcome of an election or a referendum or to influence the voting behaviour of natural persons in the exercise of their vote in elections or referenda means that the scope of the use caseis both collective influence (the outcome of the elections or referenda as a whole) and individual influence (the behaviour of individual voters). Despite their alternative formulation, those two use cases are partially overlapping. That is because the influence of voting behaviour has as its necessary consequence an influence on the outcome of an election or referendum, whether by shaping the electorate’s choices or by affecting voter turnout, and thus altering the distribution and the outcomes of votes. The wording of point 8(b) of Annex III explicitly ties the influencing of voting behaviour to the ‘exercise of [a] vote in elections or referenda’.

The inclusion of ‘voting behaviour’ in point 8(b) of Annex III clarifies that this high-risk use case covers AI systems which are intended to shape the decision-making of individual electors, even if the final outcome (the overall result) is not directly targeted. Without the reference to voting behaviour, a narrow reading could suggest that the high-risk rules only apply where there is a demonstrable risk of altering the collective outcomeof an election or a referendum. However, such a reading would be contrary to the primary objective of including this use case for protecting democratic processes from undue external interference . Conversely, influencing the outcome of elections or referenda has a broader meaning, as it can go beyond voting and relate to other stages of the electoral process, such as redrawing the boundaries of constituencies.

Point 8(b) of Annex III refers to ‘voting behaviour of natural persons’. In this contextvoting , behaviour should be understood to cover both electoral choice and electoral participation. Therefore, ‘voting behaviour of natural persons’ should be understood as covering not only how individuals cast their votes (e.g., which party or candidate they support), but also whether they choose to participate in elections or referendum in the first place. Those terms also cover both active voting and abstentions (not turning out to vote or through casting a blank/invalid ballot).

Point 8(b) of Annex III does not require that the voting behaviour of natural persons is effectively influenced. To fall within the use case, the AI system must be intended to be used for that purpose (incl. when being advertised to be used for a particular purpose). The actual result of the AI system after its deployment is immaterial.

The two use cases listed in point 8(b) of Annex III both explicitly refer to the outcome or exercise of a vote in elections or referenda. This should be understood as a requirement that the AI system is developed to be used in the context of an election or referendum. The notion of elections or referenda should be understood as encompassing those that may take place at the European, national, regional, or local level.

Point 8(b) of Annex III refers to AI systems intended to be used for a particular purpose, therefore requiring all providers of such AI systems to classify them as high-risk. The provision does not specify who provides, puts into service, or uses the AI system. The decisive factor is the intended purpose of the AI system as specified by the provider, not who the deployer will be.

In practice, AI systems intended to be used to influence the outcome of elections or referendum, or the voting behaviour of natural persons, may be expected to be provided or put into service for use by the following types of deployers (non-exhaustive list):

• Political parties, political foundations, campaign organisations , and candidates in elections: these can be expected as the primary actors who would deploy AI systems to influence voting behaviour or the outcome of an election or referendum. The exclusion clause in the second sentence of point 8(b) supports this. Third parties (e.g. advocacy groups): these persons or bodies may also have an interest in influencing the outcome of an election or referendum or the voting behaviour of natural persons, or may do so in the interest of political actors. Media outlets in many cases fall outside the scope; however, this assumption may not hold in all situations, for example, where media outlets are affiliated to a campaign organisation or party-run, or where they offer voting advice applications.

Point 8(b) of Annex III clarifies that it does not include AI systems to the output of which natural persons are not directly exposed, such as tools used to organise, optimise or structure political campaigns from an administrative or logistical point of view . This generally excludes AI systems used by political parties or candidates to increase back-office efficiency.

Political influencing is a legitimate activity which is part of the electoral context, insofar as it is in compliance with existing EU and national law. The high-risk classification of point (8)(b) of Annex III does not aim at restricting the development oruse of AI systems for these purposes, but ensures that they are developed and used in a safe and trustworthy manner with safeguards to address risks to fundamental rights and democratic values.

Manipulation or interference with elections, in particular foreign information manipulation and interference (FIMI), is an illegitimate activity and addressed by specific EU acquis, such as the Regulation (EU) 2024/900 (Political Advertising Regulation)¹⁰⁷, Regulation (EU) 2022/2065 (‘the Digital Services Act’)¹⁰⁸ and the European Democracy Shield¹⁰⁹¹¹⁰. The AI Act adopts a product safety approach and addresses a different layer, namely the safety and trustworthiness of AI systems which are legitimately used. Under the AI Act, only a certain number of AI practices that pertain to placing on the market, putting into service, or use of AI systems with un acceptable risk is prohibited; the classification of AI systems as high-risk does not serve that purpose.