Abstract
This article compares political clientelism and datafied campaigning as two modes of relating politicians/parties and voters that are centred around voter surveillance. It contributes to the discussion on consequences of Big Data by showing similarities of datafied campaigns with a type of electoral politics that pre-dates the advent of mass media and is usually regarded as deficient. It thus departs from the predominant perspective on datafication and surveillance, which draws on Foucault, in order to identify the particular challenges that datafication poses in the realm of democratic electoral politics. They are related to four major aspects in which datafied campaigning resembles political clientelism, as opposed to the combination of ideology, issue-based campaigning and media appeal that characterized Western European party politics in the second half of the 20th century. It personalizes the relationship between politicians and voters with the help of intermediaries; it is based on an asymmetric and iterative monitoring of voters; it implies a strong particularism and an affinity with populist appeals; and it is ambivalent with regard to the volition of voters. The identification of these similarities renders general concerns about the consequences of datafied campaigning for democracy more concrete. It offers a mirror in which seemingly novel practices are revealed to have implications that are well known to be problematic for the quality of democracy.
Keywords
Introduction
One aspect of the datafication (Schönberger and Cukier, 2013: 73–97) of everyday life is the opportunities that it affords political parties to take recourse to instruments of voter surveillance, such as the use of microtargeting (Bennett, 2015). At first sight, this seems to be a very new phenomenon, enabled by the same combination of online data-mining and targeted advertising that is shaping ‘surveillance capitalism’ (Zuboff, 2019). However, in this paper, I look for aspects of datafied campaigning that resemble older practices of organizing electoral majorities. I go beyond a genealogy of data-based expertise that finds precursors in opinion polling and statistics and interprets the ubiquitous surveillance in the age of digitalization and datafication as a new version of biopower or a new form of governmentality (Aradau and Blanke, 2017; Rouvroy and Berns, 2013; Stark, 2018). Although this perspective offers extremely valuable insights, I propose a different lens; one especially suited to aiding our understanding of the implications of Big Data in the context of formal electoral politics. By focusing on the characteristics of the relationship between parties and voters that datafied campaigns construct and shape, I identify several features that are similar to political clientelism.
Definitions of clientelism are manifold and have different implications for the scope of the concept, which is why the consideration of defining features is already part of the comparison I suggest to undertake. In the most general sense, political clientelism refers to an informal exchange of political support, particularly in the form of the vote, in return for some direct benefit or a credible promise of such a benefit (Hicken, 2011). Political marketing, which is at the core of efforts to learn more about the interests and preferences of potential voters, mostly understands political campaigns in the same way (Scammell, 2014: 13–34). Yet while political marketing takes its concepts from commercial marketing, research on political clientelism originates from the anthropological enquiry into the power relationships in traditional societies. Where the former takes mass communication for granted, the latter looks at a link between politicians and voters that predates it. Political clientelism is based on a personalization of this link that played hardly any role in the age of media-driven campaigning (Blumler and Kavanagh, 1999) but is at the heart of datafied campaigning.
The aim of exploring an analogy of datafied political campaigning and clientelism is to tease out the re-appearance of once familiar patterns in organizing electoral majorities, thereby furthering our understanding of what datafied campaigns mean for electoral democracy. To do so, I first reiterate the predominant view on the link between datafication and surveillance, which primarily draws on Foucault. The lessons to be learnt from this literature suggest taking a closer look at the ways in which political parties have used knowledge about voters to deal with electoral uncertainty in the past. I distinguish between a mode of observing the electorate that relies on the intertwining of opinion research and mass media on the one hand and political clientelism as a mode of observation that precedes the era of mass media on the other hand. I then focus on four aspects in which datafied campaigning resembles political clientelism as opposed to the combination of ideology, issue-based campaigning and media appeal that characterized party politics in Western Europe in the second half of the 20th century. In proposing this comparison, I neither claim that datafied political campaigns are the same as political clientelism nor that such a comparison can capture all important features of datafied campaigning. Yet when it comes to understanding its consequences for the future of democracy, the comparison reveals some surprisingly old principles of organizing popular majorities at the heart of the emerging datafied politics.
Big Data and algorithmic surveillance/power
Theoretical approaches to social uses of Big Data predominantly revolve around notions of surveillance and governance. They draw on surveillance and governmentality studies, which are in turn closely related through their common roots in the work of Foucault. Their differential stress on distinctive parts of Foucault’s thinking implies slightly divergent perspectives on datafication, one more focused on characterizing a distinctive type of surveillance, the other on the role that knowledge based on numerical calculation plays in it.
From the point of view of surveillance studies, the use of Big Data and machine-learning algorithms intensifies previously existing surveillance trends by enabling searches for patterns in large, heterogeneous data sets (Kitchin, 2014; Lyon, 2014). Machine learning algorithms, especially deep learning, reinforce and evolve a mode of governance based on the prediction of behaviours and events in targeted populations. As a result, surveillance can expand beyond previously existing limits.
The panopticon, as Foucault (1977) reconstructed it from Bentham, serves as a powerful metaphor for this trend, in conjunction with references to Deleuze and Guattari (Deleuze, 1992; Deleuze and Guattari, 1988). Foucault understood the panopticon not only as a specific site of surveillance, aimed at a specific group of people: he linked it to the notion of discipline, a form of power that seeps through society and produces bodies routinely conforming to norms of behaviour (Elmer, 2012). Since surveillance studies tend to take the panopticon literally as an instrument of surveillance, they add notions like rhizomatic surveillance and surveillant assemblage (Haggerty and Ericson, 2000) to stress both the exponential growth and the levelling effect of contemporary surveillance. Compared to earlier surveillance practices, nobody is exempt and the value of the surplus of information is distinctly economic; two points reiterated by Zuboff’s notion of surveillance capitalism (Zuboff, 2019).
While surveillance studies (Haggerty and Ericson, 2000; Lyon, 2014) suggest a shift from discipline to control in the sense of Deleuze, Foucault (2007, 2008) proposes the notions of biopower and governmentality as successors to disciplinary power. These are linked to statistical knowledge and the capacity to monitor populations, observe normality in terms of averages and direct the attention to the diverging and therefore risky characteristics and behaviour of subpopulations.
Much of the research that takes its lead from the link between governmentality and prediction focuses on aspects of security, such as insurance, predictive policing or border control. The taming of contingency (Aradau and Blanke, 2017: 375–378) is the common denominator of all these practices. Big Data analytics serves the anticipation and prediction of behaviour in a superior way (Hansen, 2015; Lyon, 2014). It results in a radicalization of biopower, based on statistical knowledge, by creating individual profiles that no longer prescribe anything, instead informing the intervention into individuals’ environment in ways that are invisible to them (Rouvroy and Berns, 2013). This algorithmic governmentality does not only promise to predict the future but it also enables near real-time influence on behaviour in environments suffused with digital communications, the data traces of which are added to the existing data profiles to identify new points of intervention (Aradau and Blanke, 2017).
Haggerty and Ericson (2000) propose the notion of data doubles, circulating in centres of calculation, to capture how the results of data-mining (fail to) represent the individuals who are the sources of the data. It clarifies that surveillance transforms the human body into infinitely mobile and comparable data, the goal not being accuracy, but discriminations between groups of people according to variable criteria. Stark (2018) suggests the ‘scalable subject’ as a refinement of the notion of data double, stressing the role of psychometrics, i.e. psychological measurements and modelling, in the new, soft biopower. This role has gained prominence due to the fact that Cambridge Analytica, the company whose activities drew public attention to practices of political microtargeting (Cadwalladr, 2019), claimed to use it. Since it is unclear whether they actually made much use of it, and since microtargeting can employ other types of data (Chester and Montgomery, 2017), I adopt the term ‘data double’ to distinguish between what the combination of algorithms and Big Data creates as knowledge about a person, and the person beyond its data traces. Moreover, I leave aside the question of whether the data double is best understood as a standardizing, prescriptive power, forcing subjects to adjust to the scales used for calculating its characteristics (Stark, 2018), or whether it sidesteps subjectivation completely by treating social life like organic life, purely in terms of stimulus and reaction (Rouvroy and Berns, 2013). The second possibility may well be the ultimate goal of personalized offers in digitized environments, while the first may be more relevant in areas other than political campaigning.
As theoretical frameworks, both governmentality and surveillance are indifferent to the possibility that the surveillant assemblage or algorithmic governmentality superimposes on other ordering principles, which may demarcate different areas of application. Studies can certainly pick out specific areas in which algorithmic governmentality or surveillance unfolds. Yet, neither of the two perspectives lends itself to asking what specific challenges datafication poses in specific fields (as for example formalized politics). Nevertheless, three general points can be distilled from the works directly or indirectly drawing on Foucault, which need to be followed up on, and substantiated in, the context of electoral competition and campaigning.
Firstly, the elimination of spatial and temporal limits to data-mining notwithstanding, there is a common core problem that drives old and new forms of surveillance or governmentality. They are supposed to deal with the contingency of conduct. In the world of political parties, this problem concerns political competitors, their own members or activists, but first and foremost the electorate, and a theoretical analysis of political microtargeting can use this as its point of departure.
Secondly, the data doubles or scalable subjects that feature in Big Data analytics need not be accurate transformations of human bodies and psyches for them to have an impact. This is true for all ensembles of instruments that political campaigns have used to predict voter behaviour and to adjust campaign communication accordingly. The ways in which political campaigns perceive voters (Hersh, 2015) may change over time, but this change is not simply in terms of greater accuracy.
Thirdly, the kind of surveillance/power that the mining of Big Data affords is not about the normalization of averages and the observation of risks in relation to those averages: it aims at the individual. This implies a significant departure from the logic of political campaigning that characterized electoral politics in most Western countries in past decades, namely a focus on mass media and their broad, diverse audiences on the one hand and a temporary mobilization of party members before elections on the other (Blumler and Kavanagh, 1999). The shift to a continuous engagement with members of the electorate, undergirded by detailed knowledge about their wants and needs, brings contemporary efforts to organize electoral majorities closer to a mode of linking politicians and voters that preceded the era of mass media. The next section will explore the three points in more detail.
Electoral uncertainty and modes of observing the electorate
Focusing on elections and political campaigning as a key process in formal democratic politics suggests the use of a notion of power that emphasizes its ordering effects (Haugaard, 2003). Without such ordering effects, the outcome of an election under universal suffrage would be completely contingent and indeterminable. In democracies, the power of the electorate derives from the rules and procedures that translate a multitude of narrowly circumscribed, formalized communications via the ballot into an allocation of political offices or parliamentary seats (Luhmann, 1983). Yet, this power in itself has little ordering effect in view of a mass electorate faced with various parties and candidates. Parties and candidates need to provide at least minimal information about past achievements, future plans or general capacities and beliefs to render the exercise of the power conveyed by the ballot meaningful (Aldrich, 1995; Cox, 1997). In structuring what potential voters can learn about them, parties and candidates exercise power over the electorate (Luhmann, 1990). Concurrently, parties and candidates face the contingency of conduct in the electorate, which is why the observation and (selective) information of potential voters is a way to reduce uncertainty about what voters will do at the ballot box.
The observation of voters with the help of opinion polls and statistical techniques became a standard element in election campaigns in the 1960s (Delli Carpini, 2011). The notion of target groups was applied to broadly defined groups, such as ‘the college-educated’ or ‘women’ (Levy, 1984) and campaigns aimed to address the issues which target groups supposedly cared about most. At the same time, attempts to identify target groups based on opinion research and find ways to appeal specifically to them were inevitably entangled with observations via mass media (Blumler and Kavanagh, 1999). On the one hand, mass media mediated the observation of the electorate by commissioning their own opinion surveys and reporting on their results (Ismach, 1984). On the other hand, politicians and parties dealt with mass media – in the form of both advertising and news management – based on a construction of the electorate that used opinion research. Mass media effects are not only direct but also mediated by opinion leaders in a two-step flow of information (Katz and Lazarsfeld, 1955). Yet this was of little practical consequence to political campaigns, which had no means to observe this flow and had to rely on general opinion polls instead.
Both the entanglement with public observation in mass media and the focus on (due to limited available data) broad target groups distinguish voter observation from the aspirations of datafied campaigning as it has emerged over the last decade. Digital marketing techniques such as lookalike modelling and geolocation targeting (Chester and Montgomery, 2017) target individuals based on similarities of their data doubles with characteristics of likely supporters and can identify the online devices they use wherever they go. Classification systems of commercial marketers link postal codes to lifestyle groups that dominate in the respective neighbourhood, and political parties combine them with their own data on potential voters to draw inferences about policies that may appeal to these groups (Bennett, 2015: 376). For example, an off-the-shelf Canadian consumer segmentation system distinguishes 68 lifestyle types, based on geographic, demographic and psychographic data (Environics Analytics, 2019).
The granularity of possible inferences varies with the expansiveness of parties’ voter files, and the accuracy of inferences is an open question (Hersh, 2015). Nevertheless, the datafied approach to political campaigning can result in targeted messages addressing extremely particularistic issues. An early example from a United States governor’s race in 2006 is a message that Republicans microtargeted to working-class snowmobilers, warning them that the Democratic candidate’s environmental policy would undermine snowmobiling opportunities (Hillygus and Shields, 2008: 14). Yet although datafied campaigning is still mostly associated with the United States and a few other cases such as Canada and the United Kingdom, examples of parties building extensive voter files also come from countries such as Chile or Kenya (Tactical Tech, 2019: 21). This suggests that politicians in very different democratic settings are attracted by the notion of datafied campaigning.
Kreiss (2016: 217–220) characterizes datafied campaigning as ‘networked ward politics’, comparing it to an era of American politics before the emergence of mass media, when machine politics prevailed in many American cities. Part of machine politics was an extensive network of ward captains who had a deep knowledge of large parts of the electorate. By interacting with constituents in their daily lives, ward captains gained intimate insights into their needs, problems and preferences (Banfield and Wilson, 1963).
There is a benevolent reading of the return of this ‘deeply social and personalized form of political representation’ (Kreiss, 2016: 214) in the age of datafication: ‘Technology and data-intensive campaigning need not be dehumanizing. Instead, it can enable political parties to better understand the needs of the public and communicate with them in increasingly targeted and participatory ways’ (Schrock, 2017: 472–473). This positive appraisal seems to be based on the absence of a transactional dimension (Kreiss, 2016: 220), which was essential for the ward politics of old, either in the form of patronage jobs or other material benefits.
A less benevolent perspective would point out that ward politics and the political machine in which it was embedded were an instance of political clientelism, a way of organizing electoral support that, at different times, was common in various democratic settings all over the world and has persisted in some until today (Clark, 1994; Kitschelt and Wilkinson, 2007; Scott, 1969). It induces members of the electorate to vote for a particular party or candidate by offering tangible benefits exclusively to supportive voters, generating sufficient knowledge about individuals or groups of voters to assess whether they are in fact supportive or not (Hicken, 2011; Kitschelt and Wilkinson, 2007; Piattoni, 2001). Although datafied campaigns do not have a transactional dimension, the knowledge dimension of (networked) ward politics deserves further exploration, especially in view of the potential geographical scope of both datafied campaigning and political clientelism. Predictive knowledge about individuals’ voting intention and about actual individual vote choice is key to stable clientelistic relationships and to the claims of Big Data applications in politics. Both seem to rely on a capacity for monitoring individuals; the question is whether this parallel implies further similarities between clientelism and datafied political campaigning.
Personalization and its intermediation
Until the advent of digitalization and datafication, political campaigns were personalized only in the sense that media coverage and advertisements drew (too) much attention to the personality of candidates (Holtz-Bacha et al., 2014). This pattern is now superseded by a non-public personalization of political issues in the sense of customized messages targeted to individual voters. Alexander Nix, as founder of Cambridge Analytica at the epicentre of the first public outrage at the datafied organizing of political majorities, described this personalization in the following way: We are trying to make sure that voters receive messages on the issues and policies that they care most about, and we are trying to make sure that they are not bombarded with irrelevant materials. That can only be good. That can only be good for politics, it can only be good for democracy […] (quoted in: House of Commons, 2018: 27)
Of course, gathering of all these data does not imply that campaigns actually make use of them to the degree of sophistication that data analytics firms claim in an effort to generate further business (Baldwin-Philippi, 2017). Nevertheless, it is the promise of connecting to voters in a personalized way and almost in real time that distinguishes datafied campaigning from earlier campaign strategies and that attracts parties and other political actors to campaigning on social media. However, it is important to note that datafied campaigns predate the ubiquity of social media in a few countries, in particular the United States. For this case, Hillygus and Shields (2008) show how both parties already in the 2004 presidential campaign used direct mail to render issues that were largely absent from the public agenda salient to those voters who generally leaned towards supporting the other party’s candidate but had a specific policy preference that did not fit his platform.
This resembles the way in which the clientelistic relationship centres on the problems of individual voters and promises to ameliorate their particular situation (Auyero, 2001). Research on the patron–client relationship and clientelism began in the 1960s, and it has witnessed various changes of focus and preferred theoretical foundation over time. It has often overlapped with research on informal political exchanges more generally. Those authors who have insisted on conceptual clarity and delimitation have however highlighted the personalized character of clientelism (Hilgers, 2011). In the absence of personal contact between patron and client, chains of brokers personalize the clientelistic relationship (Hicken, 2011: 290–291; Lemarchand and Legg, 1972). Brokers are in regular personal contact with voters to gauge their political preferences and loyalty and to distribute benefits or mediate access to them on politicians’ behalf (Stokes et al., 2013). It is this personal, long-term and diffuse character of clientelistic relationships which distinguishes them from related variants of particularistic distribution, such as vote-buying or pork-barrelling (Hilgers, 2011).
Clientelism is not an alternative but a complement to other strategies that aim to attract voters, such as broad ideological commitments and mass media appeals. Clientelistic relationships can enhance the organizational capacities of parties and candidates by turning clients into active supporters who try to mobilize and persuade additional voters (Trantidis, 2016: 10). Yet, different from typical interest group organizations, clientelistic networks do not pursue collective goods; the relationship is constituted at the individual level of favours and support exchanged between client and patron (Kitschelt, 2000). This includes the possibility of collectives benefiting from clientelism. However, whether these are poor neighbourhoods being connected to the water supply (Chandra, 2007: 107–108) or professional groups and unions being granted special pension schemes (Mavrogordatos, 1997; Petmesidou, 1996: 329–337), their interests are considered due to the intermediation of brokers who maintain the individualistic, as opposed to collective, character of the relationship.
One effect of personalization based on intermediaries is that political clientelism is very open to voters (Piattoni, 2001). It is able to forge links to members of the electorate who are otherwise marginalized (Holzner, 2010: 45–50), although it deals in promises that are only sometimes fulfilled and cannot be claimed as rights (Auyero, 2001; Chubb, 1981). Concurrently, it discourages the formation of independent civil society groups and genuine political participation (Gay, 1998; Sotiropoulos, 2004). The orientation towards individuals and their problems means that parties operating on such grounds do not contribute to interest aggregation and generalization (Kitschelt, 2000; Piattoni, 2001).
Although neither clientelism nor datafied campaigning encourages voters to directly connect with each other, they are heavily based on networks that weave intermediated dyads together. In the case of clientelism, the network of relationships is often based on personal contact in close geographical proximity (Clark, 1994; Levitsky, 2003), although clientelistic networks can also mediate special interests of professional and other groups (Mavrogordatos, 1997). In the case of datafied campaigning, geolocation targeting and lookalike modelling make it possible for digital marketing to use either inferred geographical location or specific interests (Chester and Montgomery, 2017) to not only forge links to individual voters but also encourage them to mobilize others (Penney, 2016). The sharing of ad content and posts on social media turns users into intermediaries who give a campaign indirect access to their personal networks (Kreiss, 2016: 217). Their role as opinion leaders suggests the relevance of Katz and Lazarsfeld’s (1955) two-step flow model of communication, originally proposed in the context of political communication via mass media, for social media as well.
Yet, microtargeting uses data doubles to profile individuals, which adds a technological, expert-based mediation to the network of intermediaries. Soffer (2019) suggests that this mediation by algorithms can also be understood as a two-step flow. Algorithms personalize exposure to contents by making inferences from the digital behaviour not only of the individual but also of other users who are deemed similar in relevant respects. In contrast to the two-step flow in the context of mass media, the groups to which an individual is linked are calculated from data doubles and therefore constantly changing. Different from the social circle of an opinion leader, the groups also lack any awareness of connection or commonality since they are dynamically determined by algorithms making use of the most recent data.
Politicians and parties have to rely on their own and/or hired data analysts as well as on the services of major platforms such as Google and Facebook (Kreiss and McGregor, 2018) to make use of this dynamic, automated intermediation. These experts can tap into a continuous stream of individual-level data on various aspects of individuals’ life, much of which is not related to vote choice in any obvious way. Since they often specifically target members of the electorate who are habitual non-voters, they can contribute to higher involvement in politics by previously marginalized groups (Dennis, 2019). Yet the terms of involvement are narrowly prescribed by the characteristics attributed to their data doubles.
To what extent microtargeted campaigning might undermine party-guided interest aggregation as a result is an open question. The answer depends on the relative weight of microtargeting compared to more conventional types of campaigning and, significantly, on the actual content of targeted advertising. At least parties or candidates heavily relying on messages to data doubles, segmented according to a variety of inferred interests, would need to take recourse to some other means when it comes to the question of which interests to prioritize once in government.
Monitoring: Asymmetry and iteration
The benefits of political clientelism are not allocated unconditionally. Their distribution is based on the understanding that beneficiaries will vote for a specific party or candidate in return. Voters who renege this implicit deal are likely to be excluded from future benefits. A perennial question in the research on clientelism has been how parties can ensure this conditionality if the ballot is secret. This puts the problem of monitoring or surveillance of voters at the centre of clientelistic arrangements in democracies. Provided the secrecy of the ballot is intact, the act of voting is impossible to monitor. Instead, vote choice is inferred by the brokers of the clientelistic relationship, based on their deep knowledge of the voters (Stokes, 2005). This invites comparisons to the emerging voter surveillance based on data-mining and machine-learning algorithms for pattern recognition, even though its goal is not to enforce an implicit deal but to suggest that a party or candidate already cares exactly about what is most important to a particular voter.
Since datafied campaigning does not have a distributive dimension as such, conditionality here applies only to the communication with potential voters. Depending on the inferences made from data doubles, whether voters receive messages from campaigns and what kind of messages these are will differ (Hillygus and Shields, 2008). Likely supporters will be encouraged to vote and become active for the campaign; likely supporters of competitors may be discouraged from voting with the help of targeted negative messages about their preferred party or candidate (Ansolabehere et al., 1994). Large parts of the electorate may be excluded altogether from targeted efforts (Gimpel et al., 2007; Lipsitz, 2009).
Clientelism is considered to be an asymmetric, hierarchical relationship, in which one side is more powerful than the other due to differences in socioeconomic status or access to resources (Hicken, 2011: 292; Weingrod, 1968). In the case of datafied campaigns, the notion of hierarchy would be misleading, and yet there is a pronounced and irremediable asymmetry with regard to knowledge. This asymmetry is of general concern when it comes to Big Data and the algorithms analysing them (Danaher, 2016). In political campaigning, it manifests as a lack of knowledge on the part of voters about when and why they are targeted with messages which others do not see in the same way.
As Chadwick (2017) points out, political communication on social media is part of an ecosystem of information flows that combines old and new media logics and also includes individuals’ offline conversations. Consequently, parties’ and candidates’ employment of Big Data for social media campaigns and microtargeting does not tell us anything about the effectiveness of strategies to isolate parts of the electorate from each other when addressing them online. Empirical studies on selective exposure and the existence of echo chambers, using data from the United States (Barberá et al., 2015; Eady et al., 2019) as well as Germany and Italy (Vaccari et al., 2016), suggest that individuals differ in the extent to which they engage in networks that support instead of challenge their political views and that only a relatively small subset of the respective electorate is exclusively exposed to political communication that confirms their views. Moreover, Hersh (2015) demonstrates for the United States that most of the predictive power of Big Data analytics in electoral campaigns depends on publicly available data, such as voter registration records, whereas data points not directly related to political preferences hardly help campaigns with getting a more accurate picture of the electorate.
Although these caveats warn against overstating the actual monitoring capacities of datafied campaigns, they do not invalidate concerns about voter surveillance and its consequences. Once again, the comparison to political clientelism is helpful. As a result of its dependence on intermediation by brokers, it also struggles with a gap between how parties and candidates perceive the electorate and how it actually responds to campaign messages and promises that are crafted based on this perception. A conflict of interest between brokers and party leaders or candidates, respectively, in combination with an information asymmetry between the two sides, leads to the inefficiency of clientelistic strategies (Stokes et al., 2013). Although it would be in the interest of party leaders to target undecided and swing voters, the identification of potential clients in this group as well as their subsequent monitoring is particularly difficult for brokers, who consequently prefer to target loyal supporters. Since party leaders and candidates lack the intimate knowledge that brokers have, they are unable to distinguish whether the latter actually attempt to win those voters whose recruitment as clients would most significantly improve the election result. Yet, no matter whether the clientelistic strategy is actually beneficial to them or not, it perceives and addresses voters as individuals who are amenable to promises of tangible benefits in exchange for their vote and who should be monitored as to whether they comply with the implicit deal.
The parallel to datafied campaigning and the monitoring it employs does not lie in the assumption of such a deal but in a particular perception of the electorate that is consequential even when it is inaccurate. It perceives the electorate as a jumble of particular interests that can be inferred from data doubles and turned into the content of messages that only those voters see who are assumed to be amenable to them. Datafied campaigns do not necessarily make use of this additional power to segment the electorate and address the segments in ways that cannot be observed publicly. They may well restrict themselves to highlighting different aspects of a publicly available programme to different parts of the electorate. Yet they have at least the potential to tip the balance between the formal power of voters at the ballot box and the informal power of politicians and parties to present themselves selectively to the electorate in favour of the latter; especially in combination with their iterative character.
Iteration is key to clientelism, the ongoing nature of which distinguishes it from related phenomena like bribery (Hilgers, 2011). It means that the actual capabilities of monitoring the political behaviour of voters need not be perfect, because repeated interaction leads to the development of expectations that orient the relationship. The individual expectation of being monitored replaces actual monitoring at the individual level. Concurrently, iteration does not only increase the reliability of clients, but also provides information on the reliability of patrons regarding the actual delivery of benefits (Hicken, 2011: 292–293). This implies that incumbents often enjoy a huge advantage over challengers, since they can access state benefits amenable to targeted distribution and thereby ensure their reliable delivery and the continuation of established clientelistic relationships (Medina and Stokes, 2007).
Iteration is also key to the kind of information on which microtargeted campaigns rely; however, it takes on a mathematical meaning. Data doubles and the inferences drawn from them cannot be created based on one-off communications. Targeted communication is embedded in a constant feedback loop of behavioural data that determines the content and presentation of messages to individuals, based on their data doubles, and the observation of changes to the data doubles as a result of intervention (Rouvroy and Berns, 2013: 176–184; Stark, 2018). Online environments facilitate iteration by enabling campaigns to compare two or more variants of an advertisement or message as to the responses they elicit from users online. Examples of this so-called A/B testing have been reported from the Vote Leave campaign for the United Kingdom’s 2016 referendum on EU membership as well as United States presidential campaigns since 2008, with the 2016 Trump campaign making particularly extensive use of it (Tactical Tech, 2019: 38–39). Since the iterated relationship is at the level of data doubles, it altogether excludes the question of actual outcome for the voter and does not add to the knowledge of voters about the qualities of parties and candidates in terms of actual delivery. The reliance on as many data points about as many people as possible implies that incumbents, who already have abundantly filled databases, may have a bigger advantage over new challengers than was the case in the era of mass media and opinion polling (Kreiss, 2016: 206–209), at least where data protection regulation permits such an accumulation of data.
This points to the crucial role of institutional rules in limiting the informal power campaigns can derive from datafication. The comparison with political clientelism also suggests the importance of such rules in shaping linkages between politicians and voters, thus making clientelistic strategies more likely in some countries than in others (Kitschelt, 2000: 859–862). For example, small electoral districts, or the publication of results at the level of precincts, facilitate clientelism by making it possible to observe the voting behaviour of relatively small groups of voters, delimited by territory (Medina and Stokes, 2007: 76–79). Similarly, the asymmetry related to datafication is more or less pronounced depending on which kinds of data organizations are allowed to collect and which kinds of data are considered as highly sensitive due to the information on political opinions that may be inferred from them. Even in non-clientelistic contexts, the perception of ballot secrecy seems to be considerably lower than actual ballot secrecy (Gerber et al., 2013). A growing awareness of data-mining and microtargeting in the population may consequently strengthen the assumption that parties can actually acquire knowledge about an individual’s vote choice. Data protection regulation and its application to new kinds of data analysis in organizations (ICO, 2018) are therefore crucial for the shape of the future relationship between parties and voters.
Particularism and populism
Political campaigning has always included an element of particularism, with different political parties catering to different social groups. Yet the possibility of having to justify specific pledges publicly limits the attraction of particularism. Parties that aim for political power have to factor in a possible alienation of groups with other interests. When parties run on a platform of policy projects and ideological commitments and their communications can be observed by a general public, excessively particularistic pledges are therefore deterred.
The new form of personalized campaigning undercuts (potential) observation by a general political public, represented by the political coverage in mass media (Hillygus and Shields, 2008). This makes it more likely that highly particularistic pledges remain undetected, which may render them more attractive as a means to sway voters. Such pledges sideline the problem of interest aggregation in a way that is similar to political clientelism, for which particularism and non-publicness are typical (Kaufman, 1974: 285), although based on face-to-face contact.
Political clientelism conveys a particularistic understanding of politics by putting individual interests and direct, short-term rewards at the centre. Simultaneously, in its iterative practice it limits clients’ expectations by only sometimes giving people something they have asked for (Auyero, 2001; Chubb, 1981). By contrast, datafied campaigns iterate only observations of data doubles and remain disconnected from questions of delivery. The picture of democracy that they paint consequently does nothing to moderate expectations, at least for those who rarely search for political information elsewhere. Provided only those targeted perceive microtargeted messages, there is no hint that other members of the electorate may have different priorities. Personalized messages suggest that those targeted can get exactly what they want, provided that the right party or candidate wins.
Considering this, particularistic promises in the form of clientelism and perhaps even more so in the form of microtargeted ads have two implications. Firstly, they raise the imagined stakes of the wrong candidate or party winning. Secondly, they impose constraints on the content of public statements with which campaigns may wish to complement personalized appeals. Both implications suggest an affinity of particularism and populism. Such an affinity has been suggested in connection with the sidestepping of mass media and traditional forms of party organization that datafied campaigns render possible (Persily, 2017). Yet it also manifests in the complementary relationship of clientelism and populism that has been noted in a number of examples, such as Greece (Pappas, 2013) and Argentina (Auyero, 2001).
Although populism is often ill-defined, its potential role in organizing electoral majorities can be pinned down to two aspects. On the one hand, it posits a distinction between ‘the people’ and an adversary of this people, which may variably be a domestic elite, foreign forces or some internal ‘other’, such as the immigrant (Aslanidis and Kaltwasser, 2016; Mudde and Kaltwasser, 2013). In the context of electoral campaigning, all other parties in the race are depicted as representing this adversary. On the other hand, the populist rhetoric does not address particular interests or concerns but suggests that those will quasi-automatically be fulfilled once the true representatives of ‘the people’ are in power. Populism is thus capable of reconciling strictly particularistic appeals in personalized communications with a public message that includes large parts of the electorate without having to get into specifics of policies, which would only undermine its plausibility. In this respect, it is ideational, but not an ideology, which (despite its high level of generality) would require a minimum specification of values to be pursued and a ranking in case of conflicts between values (Luhmann, 1962).
Volition and misrecognition
The most contested key element of clientelism is volition (Hicken, 2011: 293–294). Scholars differ in their views regarding the extent to which clients are coerced to enter into a clientelistic relationship as opposed to voluntarily choosing it. Disagreement on this point is partly due to the choice of research cases and partly due to different theoretical underpinnings. Associating with a patron may be the only way to access resources in authoritarian settings that suppress the formation of grassroots groups demanding such access as a right (Escobar, 1994). This intuitively suggests a coercive dimension that is absent when people can freely associate to raise collective demands. Yet a rational choice approach would model both situations in terms of costs and benefits and posit that people enter a clientelistic relationship when its perceived costs are lower than its perceived benefits. An approach stressing asymmetries of power would tend to attribute different degrees of coercion depending on how pronounced the asymmetry is (Hicken, 2011: 293).
In relation to datafied campaigning, the matter of volition is similarly ambiguous. Users of Facebook, Google or other advert-based online services can hardly opt out of microtargeting. The first line of defence these services have against their users opting out is ignorance. As long as users have no or only a vague idea about how their data is used, they will simply see no reason to opt out. Yet even users who have a clear notion of the principles of microtargeting can hardly escape it. With the implementation of the General Data Protection Regulation 2018 (GDPR) in the European Union (European Parliament and Council, 2016), they can in principle reject the personalization of adverts and the collection of data to this end. In practice, it is very burdensome to do so and provision of the service is often coupled with giving consent to extensive data collection. Moreover, the GDPR does not address the possibility that algorithms may transform data not considered sensitive at the point of collection into data doubles from which highly sensitive inferences are made (Veale et al., 2018; Wachter, 2019).
Yet, at least a rational choice approach would point out that users apparently weigh the benefits of using the services higher than the costs that they have to bear in terms of personal data collection and exposure to targeted advertising. In the United States, weaker data protection regulation has encouraged parties to collect their own data on voters, to a significant extent based on voter registration records (Hersh, 2015; Kreiss, 2016). Opting out of being the targets of data collection and custom-made political messages in such a setting would be the same as opting out of participating in elections.
Both political clientelism and datafied campaigning thus constitute relationships that can be described in terms of an exchange logic, and yet such a description does not really capture their specificities. Indeed the absence of an explicit notion or even awareness of exchange turns out to be an essential element of both. In the case of clientelism, the exchange remains implicit and is enacted (and misrecognized in Bourdieu’s sense) as a personalized relationship anchored in an understanding of reciprocity and mutual support (Auyero, 2001; Garrigou, 1998). The accounts that those involved in political clientelism give of the relationship consequently often differ from reasoning in terms of a long-term quid pro quo. Explicit monitoring is thereby rendered unnecessary; a clientelistic understanding of doing politics is instead a taken-for-granted practice. At the same time, clientelism encourages participation in politics, precisely because it provides powerful incentives and mobilizes via powerful networks (Holzner, 2010: 45–50), even though it is a version of politics that does not meet certain normative expectations about democracy.
In a similar manner, the multiple accounts of the relationship between parties and voters that datafied campaigns establish should be considered as part of the phenomenon. From the point of view of a party organization that attempts to emulate commercial marketing ideas, data-mining and microtargeting are elements of relationship marketing, which is the best available substitute for weakened ties to specific milieus and social groups. Voters who receive messages that correspond to their interests and views may feel encouraged to engage with politics in the way that the messages suggest and, for example, share them with friends. Yet, critical observers may identify power differentials and the knowledge asymmetry that datafication implies as equally central to understanding how it shapes political campaigning and the organization of electoral majorities.
Conclusion
In contrast to the predominant theoretical approaches that link datafication and surveillance, a comparison between datafied campaigning and political clientelism highlights effects that are both specific to and far from new for democratic politics. Firstly, the comparison clarifies both the advantages and the drawbacks of personalizing the linkage between politicians and voters. On the one hand, personalization has the potential to mobilize marginalized voters who would otherwise remain excluded from electoral politics. On the other hand, such personalization discourages voters from conceiving themselves as part of potential collectives with common interests; in the case of clientelism this is a consequence of its transactional dimension, in the case of datafied campaigns it is a result of the knowledge asymmetry that leads to a fluid, constantly changing grouping of voters by algorithms.
Secondly, the monitoring of voters is central, although not necessarily efficient, for both clientelism and datafied campaigning. The awareness of being monitored in terms of particular interests and corresponding behaviour is likely to affect voters’ behaviour in both cases, even if there remains a gap between how voters are perceived by politicians and parties and how they actually decide at the ballot box.
Thirdly, both clientelism and datafied campaigning encourage a particularistic understanding of democratic politics, at least in a part of the electorate. This may render a populist rhetoric an attractive complement for politicians and parties heavily relying on one of these strategies.
Fourthly, both clientelism and datafied campaigning only seemingly follow an exchange logic. It is confounded by power differentials and knowledge asymmetries, which undercut the element of volition that the notion of exchange implies.
However, the comparison also suggests that the effects of personalization, the monitoring of behaviour, heightened particularism and misrecognition are neither inevitable nor likely to be uniform across all democratic settings. Neither clientelism nor datafied campaigning is an exclusive or inevitable strategy. In the case of clientelism, there are examples of party competition where some parties chose to rely on clientelistic links and others stuck to programme and ideology (Chubb, 1981) as well as examples where parties combine clientelistic appeals to poor voters with programmatic ones to members of the middle class (Magaloni et al., 2007). In an analogous manner, differentiated datafied campaign strategies may attempt to distinguish between different patterns of media use and selective exposure in the electorate, although this distinction itself is likely to be made based on Big Data analytics.
The gap between how campaigns perceive voters and what voters actually do is also unlikely to be seen in the same way across all countries and parties. On the one hand, institutional rules partly determine what information politicians and parties can gain from clientelistic networks or Big Data analytics, respectively. In the case of political clientelism, rules pertaining to the electoral procedure and the granularity of published election results are particularly relevant (Medina and Stokes, 2007); in the case of datafied campaigning, rules regarding data protection (Wachter, 2019) as well as the content and access to electoral registers and other public records (Hersh, 2015) are of similar significance. Such rules will inevitably matter when parties and candidates attempt to assess the possible gains of clientelism and datafication, respectively.
Moreover, politicians are at a disadvantage to detect discrepancies between their perception of voters and actual behaviours, interests and motives in the electorate, lacking the knowledge of data experts and of local brokers, respectively. Consequently, neither clientelistic practices nor datafication is amenable to scrutiny regarding their actual electoral payoff. Belief in their effectivity ensures their continuation; with past successes, anecdotes as well as the persuasiveness of numbers, and brokers’ personality being factors that can nurture the belief but are hardly uniform across all settings. Nevertheless, both clientelism and datafied campaigning appear to have (had) a particular persuasiveness for politicians in a wide variety of democracies. It may lie in a similar promise of control vis-à-vis members of the electorate, the extent of which is foreign to campaigning based on issues, mass media coverage and opinion polling.
Footnotes
Acknowledgements
Earlier versions of this paper were presented at the Work-in-Progress Seminar at the Institute for Advanced Studies in the Humanities (IASH), University of Edinburgh, and as part of the Sub-Theme ‘Organizing in the Age of Digitalization and Datafication’ at the 35th Colloquium of the European Group for Organizational Studies in Edinburgh. I would like to thank the participants for helpful comments and suggestions. I am also grateful to Amy McMonagle for her careful reading of the manuscript.
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the European Institutes for Advanced Study (EURIAS) Fellowship Programme, funded by the Marie Curie Actions within the Seventh Framework Programme of the European Union.
