I. Introduction

European and global democracies are under a severe threat due to extensive spread of disinformation through social and traditional media. The use of automated accounts and bots, psychographic micro-targeting, and deepfakes to proliferate fake news during elections are making the problem even more alarming. In addition, freedom of elections in the EU and the European democracy can be detrimentally affected by the use of artificial intelligence (AI) in other ways. For example, automated social bots can be (mis)used to promote political candidates and convince the voters to vote for this candidate even if they do not spread disinformation, in particular if coupled with micro-targeting.

The increasing use of artificially intelligent tools can seriously threaten public values of democracy, rule of law, freedom of elections and prevention of manipulation of voters. Nevertheless, it is also crucial to balance these values against freedom of expression, media freedom and media pluralism. EU institutions, governments, media outlets and civil society are deploying regulatory mechanisims to ensure the balance between these public values. For example, numerous policy documents have been adopted on the level of the EU to counter disinformation1 and in response to political advertising.2 This article aims to discuss the ways in which the use of AI can impact democracy and freedom of elections. More specifically, it discusses the potentially detrimental impact of the use of social bots, psychographic micro-targeting, disinformation and voting advice applications.

II. The Use of Artificial Intelligence to Impact Democracy

1. Social Bots

Social bots are automated or semi‐automated social media accounts, primarily controlled by algorithms and programmed in a way to have the ability to interact with human social media users.3 They can automatically generate and spread content, without revealing their non-human identity.4 One of the most important features of bots is that they can achieve scalability,5 which enables them to massively spread information and hence to (artificially) enhance the importance of a particular idea or popularity of a political candidate. Social bots used for political propaganda are sometimes termed also ‘political bots’,6 which strive to generate ‘likes’ and attract followers on social media.7 These bots can also identify keywords in public posts or conversations and then populate the content with their own posts (if they act as spam bots) or conversations (if they act as chat bots).8

Consequently, social bots can impact public political opinion by promoting or discrediting political candidates. Simultaneous use of numerous bots at the same time may also convey the impression that the information is coming from highly diverse sources, and is hence more reliable. Nevertheless, social bots can spread truthful information as well as disinformation alike, depending on the (potentially manipulative) goal for which they are being used. Finally, social bots can be efficient in mobilising citizens for the so-called ‘astroturf’ campaigns.9 Astroturf campaigns give the impression to be grassroot campaigns run by non-profit organisations or citizens, whereas in reality their driving force are businesses or politicians who do not reveal their identity.10

The use of social bots in election campaigns raised a considerable degree of controversy primarily during the 2016 US presidential elections.11 However, elections and referenda in Europe do not seem to be immune on the impact of bots either. For example, (dis)information was spread with the help of bots in Germany during debates on the UN migration pact12 and during the 2017 German elections,13 in Sweden during the 2018 elections14 and in France during the 2017 presidential elections.15 Moreover, the Brexit referendum campaign is a salient example of the use of social bots,16 where the latter were allegedly even used to massively sign a petition for a second Brexit referendum.17 Bots were reportedly spreading information also in the run up to the Catalan independence referendum,18 within the framework of a debate on immigration in Italy19 and in the run-up to the European elections.20

From a technical perspective, the creation of social bots on social networks is becoming increasingly easier. On Twitter, for example, the creation of social bots is greatly facilitated through the open application programming interface (API)21 and it has been found that bots represent up to a quarter of all Twitter accounts.22 Facebook is equally populated with a considerable amount of fake accounts23 and bots on its Messenger service.24 Increasingly, the creation of bots does not require specific in-depth programming skills as it is facilitated by online services such as Somiibo.

2. Psychographic Micro-Targeting

Democracy and freedom of elections can be significantly impacted also through micro-targeting of voters with political advertisements,25 particularly if coupled with the spread of disinformation. In the social media environment, profiling for the purposes of advertising is usually based on the combination of objective criteria, such as gender, age, marital status or place of residence, and subjective criteria, such as personal interests and personal history. While micro-targeting has traditionally been used for commercial advertising, nowadays it is increasingly deployed for political advertising during election campaigns.

In particular, during the 2016 US presidential elections, micro-targeting using psychographic criteria – also termed psychographic profiling – was widely used to convince the voters to cast their vote for the Republican candidate.26 In other words, political advertisements that targeted individual voters on social media, appealed on their personality type. The profiles were created by the data science company Cambridge Analytica mainly on the basis of a personality quiz launched on Facebook that enabled to classify the users as having an open, conscientious, extravert, agreeable and neurotic personality type.27 In addition, this data was coupled with a massive amount of other Facebook data about users, which were, in turn, micro-targeted with tailored political advertisements.28 This data operation allegedly helped Donald Trump win US elections in 2016.29 A similar data science manipulation apparently also contributed to the win of the ‘leave’ vote during the Brexit referendum.30

3. AI as a Tool to Create and Spread Disinformation

During the past years, democratic political systems in Europe and globally have been significantly endangered by spreading of disinformation (fake news), particularly during the run-up to elections. The use of AI can significantly exacerbate these threats in three regards.

First, as already analysed above, the spread of disinformation through automated accounts and bots, coupled with psychographic micro-targeting, does not only reach an incomparably greater number of voters, but also appeals to their sensitivities, fears and psychological characteristics.

Secondly, while automated journalism may greatly facilitate reporting, it is also highly important that the content generated by automated means is regularly verified and that accountability for such journalism is ultimately attributed to a human.31 However, unless there is a malicious intent behind such automated content, false information would be at best created incidentally and have limited capacity to harm freedom of elections. Nevertheless, the potential of automated journalism to create large-scale disinformation should not be neglected32 and should be included in the policy debates on disinformation.

Third, an even more significant threat to democracy could be posed by the creation of political deepfakes.33 Deepfake technology enables creation of video or audio material that appears to be real, but is actually fake. It can take a form of interposing parts of the video with other content (such as swapping the face) or manipulating the video so that it seems that the person on the video is saying something she is actually not.

There is currently a disagreement as to whether deepfakes effectively represent a threat to democracy: some see it as a potentially severe source for political manipulation,34 whereas others consider it to be a ‘false alarm’ whose threat ‘hasn’t materialised’.35 Nonetheless, the mere potential for political manipulation is a sufficient source for concern and for an early regulatory response. Numerous examples have been put forward to depict these threats that can be, in the framework of elections, broadly categorised in two categories: videos aimed to harm political opponents and those seeking to enhance the candidates’ political popularity. The first category could include videos depicting politicians involved in corruption or another controversial or criminal activity and uttering statements with inappropriate or offensive content.36 The second category could include fake videos of politicians attending high-level international meetings they never attended, shaking hands with prominent world leaders or offering support to vulnerable societal groups, such as homeless, sick or otherwise affected. The recent doctoring of videos of Nancy Pelosi demonstrates that deepfakes could represent a serious threat for democracy and freedom of elections.37

However deepfakes are used, they have the capacity to lead to manipulation of elections where timing is of essential importance; if such a video is released shortly before elections, it can severely damage a candidate’s political reputation or even sway election results.38 This potentially harmful effects can be exacerbated by difficulties to effectively detect and debunk these quasi-realistic videos that give the audience an appearance of truth.

4. Algorithmic Voting Advice Applications

The aim of the voting advice applications is to help users take a decision which political party corresponds best to their political opinions. This is particularly important in a multi-party system with numerous smaller or mid-size political parties whose political agendas do not different considerably, yet aspire for their voice to be heard. Voting advice applications have become widespread in the run-up to numerous elections, both on the national as well as on the European level. For example, in the run-up to the 2019 European elections, different voting advice applications were offered to the European citizens, not only on the national level, but also through a Europe-wide recommender systems such as euandi201939 or EUvox40.

In the Netherlands, for example, voting recommender systems are widely used to help voters choose their preferred party; different websites are available that offer a specific EU version of the recommender system for European elections, such as StemWijzer41, mijnstem42, Kieskompas43, or the MVO Kieswijzer44. According to statistical data, around 10% of the Dutch population used the StemWijzer recommender system before the 2019 European elections; for the national elections, this number was much higher.45 In Slovenia, similar voter recommender systems were used before the 2019 European elections, such as the one from the newspapers Večer46 or Delo47. In Poland, a similar recommender system was used for previous elections, named ‘Latarnik wyborczy’ (meaning Election lighthouse).48

The most pressing question with regard to the voting advice applications is whether the recommender algorithms do not suffer from an engrained bias that would lead to favouring of a particular political party. Such potential presence of bias would depend also on whether the organisation that sets up such an application is an independent body or a body potentially affiliated with a certain political party. Unfortunately, on many of the abovementioned applications, this information is not available. Moreover, as with many other algorithmic decisions, transparency regarding how exactly the matching between the users’ answers and those provided for by political parties is not available. Further work therefore needs to be done to enhance the transparency of these recommender systems, to avoid potential bias.49

Moreover, the voting advice applications collect sensitive data about peoples’ political preferences and political opinions. Even though many of the applications seemingly function on an anonymous basis, which would render the General Data Protection Regulation (GDPR) inapplicable,50 they sometimes nevertheless collect personal data that allows for identification of users. For example, the Dutch voting advice application Kieskompas requires data about the year of birth, province, postal code and even (optionally) e-mail address of users, which is clearly information that enables the identification of the user.51 The possibility to identify users would lead to the application of the GDPR and the requirement of processing such data on one of the legitimate grounds for processing from Article 9(2) GDPR, with the most common ground being explicit consent (Article 9(2)(a)). However, the websites offering voting advice applications that collect personal data in principle do not require the users to give their explicit consent.

However, the impact that the voting advice applications have on voters’ electoral choices seems to be limited. Research suggests that the voters were mostly impacted in their electoral choice when the suggested party already coincides with the party they were already considering to vote, but little impact has been noticed when the system suggested the user to vote for a party she did not previously consider.52 This demonstrates that voting advice applications might not have a detrimental effect on democracy and freedom of elections.

III. Conclusion

The use of artificial intelligence can threaten and protect democracy at the same time. As demonstrated above, numerous application of technologies deploying AI can be detrimental to democracy. Due to this broad array of impacts, it seems appropriate to concur with Polonski who somewhat controversially, albeit truthfully, opined that the AI ‘silently took over democracy’.53 In response to this ‘takeover’, the EU has already taken numerous legal and policy measures to protect democratic processes, especially through protection of voters’ data. The European Parliament for example called for more transparency with regard to political advertising54 and the rules regarding funding of European political parties have been amended with the aim to prevent misuse of data to impact the outcome of elections on the EU level.55 It seems that the rules on data protection are therefore no longer serving only the protection of private individuals, but also safeguarding public values, including democracy. Finally, it is perhaps important to note that AI itself can also play a significant role in protecting democracy and democratic values, such as freedom of expression and fairness of elections. The detection of controversial uses of AI technologies is sometimes possible only with the AI itself. Typical examples are automated detection of deepfakes which cannot be recognisable with a naked eye or the establishment of existence of bots on social media. It is important to recognise that any technology is in and of itself neutral and that democracy and other public values are impacted through the human use of this technology and its purpose, as determined by humans.