Citizen conceptions of democracy and support for artificial intelligence in government and politics

Theoretical assumptions

AI has the potential to help with solving cognitive tasks in many different domains. This capability can equally be brought to bear in the political realm, as AI may automate routine tasks, help to better allocate resources, manage risks and accommodate citizens’ preferences (Engin & Treleaven, 2019; Margetts & Dorobantu, 2019; Wirtz et al., 2019). Moreover, there exist initiatives for harnessing data and AI as tools in the hands of citizens to empower them by achieving greater transparency and control over political decision making (see e.g. Savaget et al., 2018). There are, however, also fundamental tensions between the operating mode of AI systems and democratic politics. Although the interplay between human decisionmakers and AI can vary, adopting AI to inform decisions means basing solutions on information, evidence and knowledge. Thus, AI helps with cognitive tasks as it serves to produce optimal outputs or decisions through processing information based on predefined objectives.

Bringing such technical problem-solving to political decision making to guide or even replace it, amounts to prioritizing effectiveness and efficiency and thus means placing output-based legitimacy over input-based legitimacy, that is, that political decisions reflect preferences of citizens (Scharpf, 1999). Greater reliance on AI could mean that citizens’ preferences as inputs to the democratic process become less relevant and that this is justified with results. Even more important consequences of AI for democracy concern the dimension of throughput legitimacy (Schmidt, 2013), because a greater reliance on AI, especially if conceded greater authority on higher levels of decision making, can change the form of the democratic political process and the way decisions are brought about.

Following Hildebrandt (2016), there is an inherent incompatibility between the decision optimization achieved with AI and the democratic process of arriving at decisions and establishing generally binding rules. AI serves to optimize predefined goals through instrumental information processing. It therefore does not accommodate the nature of democratic politics and specifically the task of figuring out what those goals should be in the first place – which entails a hermeneutic process of reinterpreting values and meanings that shape a society and polity (Hildebrandt, 2016). In this view, liberal democracy is a cybernetic system too. However, it differs from mere information processing as it demands the possibility of competing interpretations, contestability and an ongoing feedback loop that enables an iterative reflection on the rules and the values guiding the polity (Hildebrandt, 2016, pp. 23−25). Distinguishing the ‘grammar’ (Hildebrandt, 2016, p. 2) of liberal democracy from that of AI in this way highlights precisely those aspects that also lie at the core of a procedural understanding of liberal democracy (Urbinati, 2014). This view foregrounds a reflexive process that allows for contesting decisions and renegotiating the rules of the polity under conditions of pluralism. A proceduralist view on democracy that stresses pluralism and contestation as pillars of democratic politics is therefore particularly useful for highlighting tensions with AI.

On a basic level, bringing AI to social and political problems conflicts with these liberal-democratic principles. Reducing social and political problems to an optimization task, first, tends to undermine the idea of an irreducible pluralism of political views and perspectives marked by at least some degree of public deliberation. The idea that one can base decisions on the processing of data by AI systems that find optimal solutions makes any pluralistic exchange and integration of different views unnecessary. AI in political decision making, second, also conflicts with a process of ongoing contestation that occurs vertically between society and decisionmakers: If one assumes that optimal solutions can be derived from actionable insights obtained from processing information, this removes the occasion to contest decision making.

These tensions make two kinds of political attitudes particularly relevant. As the use of AI means altering to some degree the existing political system, acceptance of AI may well depend on citizens’ political support. This attitude reflects whether citizens orient themselves favourably or unfavourably toward different parts of the present political system (Easton, 1975, p. 436), such as political elites or institutions. Besides how citizens think about the status quo of politics, the kind of democratic politics that they want could be equally or even more important. Citizens’ conceptions of democracy – particularly how much they demand a reflexive democratic process based on pluralism and contestation – are therefore also relevant when it comes to their acceptance of AI in government and politics. Although citizens will not generally have a deeper knowledge of AI, they will have a basic understanding based on own experiences and particularly popular media coverage. As noted further above, people generally perceive AI to be more suitable where it deals with technical tasks. One can thus presume that they know AI as a technology that processes information to find optimal solutions – be it in translation, image recognition or playing games like chess and Go. Thus, if a person finds AI acceptable in political decision making, it seems very likely that she has a reductionist view of politics as technical problem solving and sees AI as a means to obtain better solutions. Hence, AI may conflict with basic principles of liberal-democratic politics, but if citizens do not value these principles as a part of democracy, they may accept or embrace AI in political decision making, even in high-level politics, for example, to take decisions for politicians.

However, there is more than only one notion of democracy that diverges from liberal-democratic principles and that citizens may sympathize with. As research on conceptions of democracy and process preferences (e.g., Bengtsson & Mattila, 2009; Bertsou & Caramani, 2020; Font et al., 2015; Hibbing & Theiss-Morse, 2002; König, 2022; Kriesi et al., 2016; Webb, 2013) has shown, citizens have different views regarding what form democracy should take. They place variable importance, for example, on expert decisionmakers, on more direct responsiveness to the will of the people, or on representative-democratic arrangements that leave politics to professional party politicians. Such views of democracy can vary with regard to whether they are compatible with AI in political decision making.

Based on the preceding considerations, the following section will formulate hypotheses about the role of political support and conceptions of democracy for the acceptance of AI in government and politics. Taking liberal democracy as the starting point, it will discuss four further conceptions of democracy (see also König, 2022) which, following Urbinati (2014), can be understood as disfigurations of liberal democracy: while they are compatible with the general idea of the rule of people, they subvert liberal democracy through undermining pluralism and ongoing contestation needed to sustain a reflexive political process. The set of four disfigurations described in the next section – a technocratic, a populist, a post-democratic and a majoritarian-relativist conception – has been chosen for two reasons. First, they are, with the exception of majoritarian-relativism, prominently treated in the literature. Second, they can be analytically anchored in contrasts regarding how they deviate from realizing pluralism and ongoing contestation. In this way, the analysis can make sure that they are conceptualized and measured on the same level, rooted in a proceduralist understanding of democracy.

Hypotheses

In sum, the four described disfigurations of liberal democracy all undermine pluralism and ongoing contestation, but whereas the relativism that marks post-democracy and majoritarian relativism – ‘anything goes’ – and the knowledge-based monism of technocracy are compatible with AI in political decision making, this cannot be said of the monism that marks populism, which clings to a true popular will. The differing directionality of the expected associations with AI support can be tied to an underlying contrast between the democracy conceptions: technocracy, post-democracy and even majoritarian relativism are largely about performance and politics realizing satisfying outputs, without inputs to the political process having to take a specific shape. Liberal democracy and populism, in turn, are more about input and throughput – about how citizens’ interests are represented and about a particular political will being realized respectively. Further, the preceding hypotheses are most likely to hold true for more far-reaching uses of AI which imply greater decision making authority and substantially change the way political decisions are produced. Where AI intervenes less heavily in democratic politics, citizens’ conceptions of democracy may be less relevant.

Besides their preferred realization of democracy, it could be citizens’ political support (Easton, 1975) that shapes AI acceptance. Implementing AI can amount to a substantial change to political decision making and institutions. Citizens might therefore show greater acceptance for making such changes if they are dissatisfied with the status quo of politics, whereas they will hardly see a need for tinkering with the political process if they are highly satisfied with politics. Testing the role of political support together with conceptions of democracy is also important, as already hinted at above, because these conceptions can go along with higher or lower political support – which can lead to spurious associations when not including both as predictors. For instance, populism has been shown to be associated with lower, and technocracy with higher, institutional trust (Bertsou & Caramani, 2020).

Sample

The data used in the analysis stem from a survey among 1115 participants recruited by respondi AG from an online panel representative of the German population aged 18 to 74 (see Supporting Information Appendix A1 for a description of the sample composition).³ Although the data come from a single country, their relevance extends beyond the German case. First, evaluations of AI in general (European Commission, 2017) are mostly as or more positive than in Germany and support for AI in a political context (IE University, 2019, 2021) seems to be largely comparable between Germany and at least several other European countries, like France, Poland and the Netherlands. Besides the level of acceptance of AI in political decision making, the individual-level correlates of AI acceptance are likely to be similar across contexts as well, especially when considering that these correlates concern rather basic political attitudes that are likely to play a similar role in different settings.

Second, Germany's societal values and political culture are similar to other central and Northern European countries (Inglehart & Welzel, 2005; Schwartz, 2006). Furthermore, as in many other European states, one can expect to find substantial variation regarding citizens’ conceptions of democracy in German society that is suitable for testing the hypotheses. Since World War II, Germany has developed a democratic political culture marked by pronounced self-expression values (Inglehart & Welzel, 2005). Yet, there is also historical experience with authoritarian regimes characterized by technocratic traits, and German post-war political culture has been described as a subject orientation marked by strong passivism and pragmatism (Almond & Verba, 1963). Still today, at least some parts of society show weak democratic values and an acceptance of unaccountable government structures (Decker et al., 2016). Further, the success of a right-wing populist party in more recent history and survey evidence on a populist attitude (Vehrkamp & Merkel, 2018) attest to the presence of a populist notion of democracy. In sum, there are several reasons for why it is unlikely that the findings are specific only to the German case.

Measures

Dependent variable: Support for AI in government and politics. The dependent variable is based on six five-point rating items. In the survey, these items followed an introductory text that briefly described AI and listed several real applications indicating that AI can assist, if not replace, decision making by professionals in certain non-political domains (see Supporting Information Appendices A2 and A3). To capture whether citizens show a general acceptance of AI in certain areas, the question asked how respondents generally evaluate the use of AI while considering that this technology will likely become more advanced in the future. The six items have been formulated with specific scaling properties in mind: they refer to the use of AI on different levels of decision-making authority such that agreement with some of the items can be expected to be lower than for others. While people may easily agree with using AI for (1) routine or (2) complex administrative tasks, this does not mean that they will also support that (3) AI informs political decisions for example, in ministries, (4) assists parliamentarians in their decision making and even less that AI (5) takes decisions that politicians would otherwise take or (6) may compete at elections, as it has occurred in Tokyo (Cole, 2018).

Conceptions of democracy. While the literature offers various measures expressing demand for different conceptions of democracy, such as for populism and technocracy, these have been developed independently. As the analysis covers five conceptions of democracy overall, directly drawing on existing measures would lead to an eclectic mix. Also, some existing measures incorporate aspects of political support as one of their elements. Specifically, a populist attitude (for an overview, see Wuttke et al., 2020) and stealth democracy as a form of expert-led government in the interest of the people (Bengtsson & Mattila, 2009; Hibbing & Theiss-Morse, 2002; VanderMolen, 2017; Webb, 2013), but also certain measures of technocracy (Bertsou, 2021), include discontent with political elites. Yet other measures, such as for liberal democracy, do not (see e.g., Kriesi et al., 2016). To deal with these issues, the measures used in the analysis are still closely linked to the literature on democratic preferences but are designed to capture the different notions of democracy on the same conceptual level while also keeping them separate from political support: Based on the proceduralist perspective formulated above, the measures focus narrowly on preferences regarding the shape of the democratic political process. They thus reflect support for different ways of organizing the political process, particularly how it translates citizens’ views into political decisions.

The items have been formulated to reflect the conceptualized similarities and differences described in the preceding section. The populist, technocratic, post-democratic and majoritarian-relativist conceptions of democracy formulated above, are all measured with three rating items each. The items for populism capture the idea that politics should realize some true and unitary will of people as directly as possible; majoritarian relativism combines immediacy with a relativism that does not care about a true popular will and sees all political positions and citizen preferences as equally valid; technocracy reflects the idea that experts should find the best political decisions for the people based on knowledge; and the post-democracy items capture that political elites do not need to have a firm commitment to any political positions and should simply be left to go about their business to govern for the people. Finally, the liberal-democratic conception is based on four items on the importance of pluralism and four on the importance of ongoing contestation (see Appendix A4).

The items for each conception of democracy have been combined into variables. A principal component analysis with oblique rotation recovers the six components (see Appendices A5 to A8 for details on scaling analyses), with dynamic contestation and pluralism as elements of liberal democracy emerging as two separable dimensions. Since these are strongly correlated (r = 0.6) and are posited as elements of a liberal-democratic conception, they have been added into one variable. Through including the five conceptions of democracy together, the analysis can account for mixed attitudes since people may favour the realization of different conceptions of democracy at the same time and to varying degrees (see e.g., Font et al., 2015; Webb, 2013).

Political support. The analysis includes measures that have commonly been used as indicators of political support. Specific political support, which refers to the performance of the political authorities, is captured with a standard item on satisfaction with the working of democracy in the country and additionally, a variable for political disaffection based on five items taken from a question used in the German Longitudinal Election study 2017 (Roßteutscher et al., 2018). These express discontent with how politicians represent citizens and a certain alienation from politics (see Appendix A9). To measure a central form of diffuse support, the survey contains standard items on institutional trust asking about trust in five political institutions.⁴

Control variables. A central control variable in the analysis is citizens’ general attitude toward AI, that is, AI used beyond political domains. Two measures will be used in the analysis: a simple evaluative question about people's positive versus negative view of AI and the seven highest-loading items of a validated AI optimism scale (Schepman & Rodway, 2020). Only seven items have been chosen in order to avoid respondent fatigue in a survey that features several longer item batteries (see Appendices A10 and A11). Besides AI evaluations, the analysis will also include a variable for self-assessed knowledge about AI (see Appendices A12 and A13).

As a further control, the analysis employs a seven-point item asking about left-right ideology. An authoritarian attitude that is more commonly found on the political right is likely to translate into greater acceptance of political arrangements that are characterized by weaker accountability (Bengtsson & Mattila, 2009; Webb, 2013). For a more fine-grained measurement of ideology, seven-point items about economic policy preferences (more welfare and higher taxes versus lower taxes and less welfare), and immigration policy preferences are also included. Ambiguity intolerance has been included as a further control variable because it may be underlying demand for a kind of politics that reduces complexity and uncertainty (Gründl & Aichholzer, 2020). It might lead citizens to be more favourable towards AI delivering seemingly clear-cut solutions for given problems, thus avoiding the conflict and ambiguity that usually marks politics. Yet the common opacity of AI, however, could also counter any such relationship. To measure ambiguity intolerance, the survey features an item battery used in the Austrian National Election Study 2017 (Wagner et al., 2018). The analysis also includes political self-efficacy, which is based on two items taken from the German Longitudinal Election study 2017.

Finally, further control variables are political interest, education, age and gender (female, male, diverse; female coded as 1).⁵ The survey also featured an attention check and a control question on the last page of the survey asking whether responses were honest and can be used for the analysis. These questions are applied as a filter together with a speeding check (at least five minutes for a survey designed to take around 10–15 minutes).

Description of the dependent variable

Figure 1 shows the average support for the six items of the dependent variable. Citizens are overall reluctant to support the use of AI in government and politics. In line with the idea that citizens are more accepting of less intrusive AI applications, some AI uses are found more acceptable than others. Similar to what has been shown for human expert decisionmakers, the level or stage of decision making can make an important difference (Bertsou, 2021). Figure 1 suggests that the higher the decision-making authority the less acceptable the use of AI. Only the score for AI taking on routine administrative tasks is above the mid-point of the scale. For AI taking decisions that politicians would otherwise make and that AI can compete at elections, support is very low. Given how drastically these AI uses would interfere with democratic politics, it is to be expected that many see them as unacceptable. Still, some citizens are willing to entertain the idea of partially delegating high-level politics to machines. What the low mean value hides: a respectable 10 per cent of the respondents have scores of three and higher.

A principal component analysis (see Appendix A15) yields three dimensions behind the six items: three item pairs in the order shown in Figure 1. The analysis in the next section will therefore examine three dependent variables measuring AI (1) in an administrative context, (2) for assisting political decision making and (3) for replacing politicians. While these three dimensions lie behind the six items, these items also yield a strong Mokken scale when taken together (see Appendices A16 and A17). The regression analysis below will therefore additionally be performed with a dependent variable combining all six items.

Regression analyses

Figure 2 visualizes the ordinary least square regression results for the three subscales of support for AI in government in politics as dependent variables (for regression tables and bivariate correlations, see Appendix A18 and Appendix A19). Turning first to the role of conceptions of democracy, there are striking differences between the models. For uses of AI concerning the first two levels of decision-making authority – in public administration and for assisting political decision making – citizens’ ideas of what democracy should look like are largely irrelevant.⁶ Citizens appear to see different notions of democracy to be equally compatible even with AI uses that imply a notable change from the status quo.

Only where AI uses would mean the strongest change to democratic politics, that is, on the highest level of decision-making authority (third model in Figure 2), are conceptions of democracy linked to AI support. The results largely support the formulated hypotheses. Notably, the most pronounced association is that for a liberal-democratic conception. In line with H1, the results imply that the more people endorse democracy as liberal democracy based on pluralism and ongoing contestation, the lower is their support for AI taking decisions for political actors. A negative association also emerges for populism, which supports H3. The majoritarian-relativist and post-democratic conception, in contrast, are positively associated with the third subscale in Figure 2, thus conforming to H4 and H5. Remarkably, a technocratic conception (H2) of democracy is not associated with the dependent variables, challenging common wisdom about the affinity between technocracy and governance by AI.

It is instructive to look at the results in combination with one of the control variables: citizens’ general AI evaluations and especially AI optimism, which is by far the most important predictor in the analysis in the first two models shown in Figure 2. In other words, whether citizens accept AI used for administration tasks and assisting political decision making is not a question of their conceptions of democracy but is heavily shaped simply by how they perceive AI in general, that is, in non-political contexts. For the more far-reaching uses of AI (subscale 3), the coefficient of AI optimism is still significant, but more than halved in comparison. Citizens seem to understand that there is a tension between the idea of finding optimal solutions based on information processing and the nature of democratic politics: general AI optimism does not as directly translate into support for AI in high-level politics whereas citizens’ conceptions of democracy are more important predictors.

Turning to the question whether political support is linked to people's acceptance of AI in government and politics, the evidence is weak. Only satisfaction with democracy reaches conventional levels of statistical significance, with a negative coefficient in the third model. Overall, the findings contradict the idea that AI support is attributable to political discontent.⁷ Only the kind of democracy that people want emerges as relevant together with AI optimism.⁸

The explained variance (adjusted R²) for the first two subscales in Figure 2 is clearly higher than for the third subscale: 0.41, and 0.35 as opposed to 0.16. AI optimism contributes heavily to the explanatory power in the first two models, but several other control variables are also relevant particularly in the first model (AI in an administration context). Women appear to have less favourable perceptions of AI used in this setting. One reason for this could be that women are more likely to be employed in the public sector and therefore perceive a higher risk of being replaced by AI. A similar reason might lie behind the significant positive coefficient of higher formal education and age. Since higher education means people are less likely to perform non-routine tasks, they are less exposed to automation risks (Arntz et al., 2016), and higher age means a lower likelihood of being replaced by AI in one's future career.

The reported results stem from regression models using the three subscales of AI support. When taking all six underlying items together into a single scale, the results largely mirror those for the third subscale in Figure 2, but AI optimism emerges as more important, while the liberal-democratic conception loses statistical significance (see Appendix A23). Also, when performing the regression analyses with the individual variables capturing citizens’ conceptions of democracy, the findings are overall corroborated (see Appendices A24 to A26). When using different variants (individual items and combinations of items) for the technocratic conception of democracy, only the first item and the first and the third item taken together show a significant but weak positive coefficient in line with H2 (see Appendices A27 to A30).

The fact that the analysis covers conceptions of democracy and political support rather comprehensively makes it unlikely that there are missing conceptually related variables that would markedly change the findings. A sensitivity analysis also shows that among the highly significant conceptions of democracy variables, the smallest existing bias due to omitted confounding variables would have to be about 45 per cent (for ‘post-democratic’) to invalidate the significant coefficient (Appendix A31). Finally, one should note that conclusions about any possible causal interpretation of the registered relationships rest on theoretical assumptions. A causal relation, opposite to what has been modelled above, would mean that support for AI in government and politics comes first and influences how citizens think about the shape that democratic politics should take and their political support. This seems less plausible, however, given that these political attitudes are more general and arguably the result of a longer socialization, whereas ideas about AI in government and politics are more specific and have become more relevant only through more recent developments.