Abstract
This paper argues for a fluid approach to the study of agency in relation to algorithms, one that promotes crossing the boundaries of established analytical positions and breaking away from dualistic forms to frame its study. Building on various intellectual traditions, we develop three sensibilities for implementing such an approach: (a) working with tensions as an alternative to thinking about algorithmic power and human agency as an either/or binary; (b) examining mediations to reverse the tendency to treat algorithms as an ahistorical and universal force; and (c) exploring transversalities to navigate the spaces that emerge between various temporalities and levels of analysis. To make our case, we examine a crucial tension in the study of agency and algorithms, namely how scholars have either attributed power to algorithms or agency to users of algorithmic systems. The conclusion situates our argument for fluidity within larger historical debates in the study of technological power and human agency.
Introduction
Recent scholarly interest in algorithms has led to the rehashing of classic debates in social theory about the power of technological artifacts and the agency of humans. Most authors define algorithms precisely as materializations of power that seek to intervene in the behaviors and practices of users of technological systems. Hill (2016: 47), for example, conceived of algorithms as ‘finite, abstract, effective, compound control structure[s], imperatively given, accomplishing a given purpose under given provisions’ (emphasis added). For Gillespie (2014: 167), algorithms are a ‘key logic’ that shapes information flows and power distribution in contemporary societies. Other definitions center on algorithms as ‘a set of epistemic, cultural and social processes that transform and reconfigure the relationships between subjects, subjects and objects and objects themselves’ (Ricaurte, 2022: 732).
Because of the centrality that algorithms have gained in multiple domains of human activity, some authors speak of an ‘algorithmic turn’ or have theorized this historical moment as the ‘age’ of algorithms, ‘algorithmic cultures’ and ‘algorithmic societies’ (Abiteboul and Dowek, 2020; Hallinan and Striphas, 2016; Schuilenburg and Peeters, 2021). Accordingly, algorithms now occupy a privileged place in the study of digital media and technological systems.
However, most research to date has concentrated on accounting for either the ‘power’ of algorithms or the ‘agency’ that humans ‘still have’ against them (Cohn, 2019: 8). This tendency has been problematic in that it relies on a dichotomic understanding of agency in relation to algorithms: researchers have theorized either the operation of powerful technological entities shaping people’s behaviors or the practices of autonomous humans vis-à-vis algorithms. As an alternative, this paper argues for a more ‘fluid’ approach to the study of these issues, one that promotes crossing the boundaries of established analytical positions and breaking away from fixed forms to frame the study of agency in relation to algorithms.
Building on various intellectual traditions, we develop three sensibilities for implementing such an approach: (a) working with tensions as an alternative to thinking about algorithmic power and human agency as an either/or binary; (b) examining mediations to supplement the focus on algorithms themselves; and (c) exploring transversalities to navigate the spaces that emerge between various temporalities and levels of analysis. We draw on Ziewitz and Singh’s (Ziewitz and Singh, 2021) notion that ‘sensibilities acquire meaning in the empirical locations they are used [...] [they] are particularly useful for dealing with ambiguous situations and dilemmas, in which there is no obvious course of action’ (p. 10).
To make our case, we begin by examining the most central tenets of work on both algorithmic power and human agency. ‘Power’ and ‘agency’ have been used to refer to similar processes, that is, the capacity to make an effect in the world. As Ferrero (2022: 4) puts it, ‘agency is paradigmatically productive, that is, it is a matter of interventions that make a difference in the world’ and, in that sense, it is ‘productive power’. In this discussion, we follow scholars’ own analytical and conceptual preferences. In the case of algorithms, researchers have preferred to talk about power and have focused on how increasingly complex computational capacities intersect with broader structural inequalities. In the case of humans, authors have privileged the notion of agency and have concentrated instead on how people think, feel, and act in relation to algorithmic technologies.
Although we examine major claims and premises in this body of work, it is not our purpose to provide a systematic literature review. (For recent literature reviews on algorithms from a social science perspective, see Burrell and Fourcade, 2021; and Gandini et al., 2022). Our discussion draws on previous empirical and theoretical work on processes of algorithmization, which allowed us to identify key texts and arguments that have shaped scholarly discussions of this issue in the past years (Gómez-Cruz et al., 2023; Siles, 2023; Siles et al., 2022, 2023). Based on this assessment, we offer sensibilities for developing a more fluid understanding of agency in relation to algorithms. The conclusion situates our argument for fluidity within larger historical debates in the study of technological power and human agency.
Algorithmic Power
That algorithms carry a certain form of power is a given in most scholarship in critical algorithm studies (Beer, 2017; Cardon, 2018). But explanations of the nature of this power and how it operates vary. Scholars have emphasized different issues when accounting for the operation of so-called algorithmic power.
First, some researchers have argued that algorithms have the power to shape key aspects of society. Schwarz (2021) summarizes this notion by asserting that ‘algorithms surely have power’ and that ‘this power changes reality’ (pp. 120–121). This view builds on a particular definition of algorithms that Just and Latzer (2017: 245) capture with precision: these are considered ‘autonomous actors with power to further political and economic interests’. Power resides primarily in algorithms themselves: they determine, regulate, anticipate, and shape human behavior.
In this perspective, algorithms owe their power to sophisticated computational capacities to work with data and thus intervene in society. Through these capacities, algorithms make it possible to identify patterns, instantiate predictive models, and render populations knowable on an unprecedented scale. Authors have thus argued that algorithms have a unique form of power in human history in that they materialize decision-making processes; enact performative rules that operate as self-enclosing contracts; and transform social categorization by privileging individualization based on surveillance over aggregation based on stable classification systems (Schwarz, 2021; Zuboff, 2019).
For other researchers, power resides not in the algorithms themselves but in the assemblages that they are part of and also help shape (Fisher, 2022). In this view, power operates through algorithmic assemblages. Unlike the work discussed in the previous paragraphs, power is not the property of given artifacts but rather that which subtends the operation of such artifacts in specific ways. Savolainen and Ruckenstein (2022: 2) capture this when they claim: ‘it is not the algorithm but the overall system that has sociocultural effects’. For this reason, these authors argue, locating agency is ‘difficult’ as it is ‘distributed in mutually-reinforcing systems’ (Savolainen and Ruckenstein, 2022: p. 3). Power thus comes from how algorithms are integrated within robust sociotechnical networks that are difficult to disassemble, particularly when these networks become automated systems.
Another way to express this is to claim that algorithms are part of infrastructures, larger arrangements that are gradually naturalized as an invisible and indispensable component of daily life (Gillespie, 2014). The most common way to emphasize this idea of invisibility has been to compare algorithms to ‘black boxes’ (Pasquale, 2015): they are protected from user intervention and public regulation. This becomes problematic in that, unlike other artifacts that are not entirely understood or ‘seen’ by their users, algorithms pervade interactions with the technologies that people appropriate on a daily basis. Conceiving algorithms in this way partially complicates claims of universal effects. Some authors have suggested that algorithmic power is fractal, that is, it manifests variously in different times and places (Introna, 2016).
Another way to frame issues of algorithmic power has been to analyze how algorithms materialize broader social and cultural forces. In this view, algorithms are the latest iterations of systems of control and exploitation with a long history that work to consolidate and expand existing inequalities. Algorithms inscribe, reproduce, and amplify deep-seated forms of power and hegemony (Benjamin, 2019; Noble, 2018). The emphasis here is on how algorithms add a layer of complexity and salience to existing forms of power that have structurally and historically shaped society. A focus on algorithms allows these authors to unmask how power manifests through technologies often surrounded by the cultural veil of neutrality and progress. Authors have also criticized the adoption of Silicon Valley’s startup culture as a global standard in technology production. In this perspective, the obsession with growth and lack of diversity explains how the technologies produced through this logic will inevitably reproduce biases and problematic assumptions about who the users will be (Wachter-Boettcher, 2017). In this way, algorithms give new life to the capitalist drive for accumulation (Zuboff, 2019) and extend the history of colonialism (Couldry and Mejias, 2019).
Several conclusions can be drawn from these different strands of research. From a methodological standpoint, most researchers have employed mechanisms to ‘audit’ algorithms and thus reveal their hidden biases (Sandvig et al., 2014). Moreover, the tendency has been to frame the operation of algorithmic power as a universal force. Milan and Treré (2019: 325) capture this tendency in their critique of ‘data universalism’, that is, the interpretation of datafication dynamics (including algorithmic power) as a process that ‘present [s] technology [...] as something operating outside of history and of specific sociopolitical, cultural, and economic contexts’.
To make sense of this form of ‘universal’ power, most studies have made conclusions at the macro level of society (Klinger and Svensson, 2018). Thus, power has typically meant the capacity to control the resources and means to impose or naturalize hegemony (Williams and Gilbert, 2022). A specific conception of autonomy underlies this form of power, one that influences people’s thoughts and actions unconsciously or ‘beyond the self-awareness of subjects and collectives’ (Savolainen and Ruckenstein, 2022: 7). Accordingly, scholars have typically referred to algorithms as ‘control technologies’ or ‘technologies of governance’ (Introna, 2016). Researchers have also argued that algorithmic power operates through ‘affective modulations’ rather than psychological and emotional manipulation (Karppi, 2018: 115). It is this ‘affective pull’ that would make algorithmic systems prone to the spread of misinformation and conspiracy content (Grandinetti and Bruinsma, 2022: 1).
Given the emphasis on algorithms themselves, not much has been said about how people act in relation to the algorithms that control them. Going back to issues of autonomy, these studies have tended to view people as relatively powerless or, as Noble (2018) puts it, as having ‘little ability to impact the algorithm’ (p. 179). Schwarz (2021) nuanced this notion by arguing that, since algorithmic power is not mediated through human consciousness, people’s ‘space for agency’ (p. 155) is limited and often reduced to creating ‘detours’ and ‘loopholes’ (p. 139) in the operation of this power.
Human Agency
As work and concerns about the operation of algorithmic power have grown, scholars have also argued for the need to more thoroughly investigate how users relate to algorithms. These studies build on the premise that algorithmic power needs to be empirically examined rather than assumed (Cardon, 2018). The claim of this body of work is not that algorithms lack power but rather that users enact their agency in the conditions set by algorithms and wider societal structures. The underlying logic is thus one of complementarity rather than replacing the notion of algorithmic power.
First, researchers have worked to challenge a key premise of algorithmic power, namely that it is not mediated through human consciousness. Thus, some research has focused often on what Emirbayer and Mische (1998) call the projective dimension of agency, that is, how individuals imagine action trajectories. In other words, researchers contend that thinking and imagining are gateways to action and, for that reason, they have an agentic nature. By assessing how people understand algorithmic ‘black boxes’, it thus becomes possible to comprehend why people act the way they do.
The focus of these studies has been on the awareness or knowledge of users, that is, the different ways and degrees in which people understand what algorithms are, how they function, and what consequences their operation has (Cotter and Reisdorf, 2020; Dogruel, 2021; Oeldorf-Hirsch and Neubaum, 2021). To operationalize these concepts, researchers have posited such notions as folk theories (intuitive, informal theories that individuals develop to explain algorithms and their consequences) or imaginaries (ways of thinking about algorithms and how they operate) (Bucher, 2018; Ytre-Arne and Moe, 2021). Whereas some of these theories express original values and ideas, others tend to reproduce the discourses put forth by tech companies themselves to frame their operation (Siles et al., 2020). Another example of this approach is the use of speculation for anticipating the potential future of algorithms. Work on imagination and speculative fiction envisions agency as a creative power that enables the capacity of groups to inhabit the future and the possibility of acting in the world. Groups of transhack feminists have engaged in artistic and playful explorations of this kind (Zaragoza Cano and Akhmatova, 2018).
Although comparatively less frequently, other researchers have focused instead on affective issues. In contrast to those who have theorized affect as a prime vehicle of algorithmic control, these authors have examined it as a means to enact agency. That is the case of researchers interested in people’s attitudes, appreciation, aversion, or irritation in relation to algorithms (Oeldorf-Hirsch and Neubaum, 2021; Siles, Segura-Castillo, et al., 2019; Ytre-Arne and Moe, 2021). These studies tend to blend the projective and the practical-evaluative dimensions of agency (ie agency is expressed as people evaluate the demands of present situations) (Emirbayer and Mische, 1998) in the sense that affect connects intensities of feeling and attachment and thus involves ways of imagining and feeling.
Finally, work on human agency has also focused on user actions and practices to relate to algorithms (Siles et al., 2019). Ziewitz (2017: 11) aptly summarized the premise that underlies this intellectual endeavor: ‘If we look at algorithms not as objects to be known in theory but as figures to be used in practice, then calls for transparency, accountability, or ethical design appear in rather different light’. Research on people’s practices has focused primarily on the practical-evaluative and iterative dimensions of agency (ie agency is shaped by the past through habits and routines) (Airoldi, 2022). Scholars have also drawn on practice theory to account for how people relate to algorithms (Stephansen and Treré, 2019). Viewed in this way, the power of algorithms is not given but rather constantly made through specific actions (Seaver, 2017).
As work on human agency has proliferated, scholars have developed more encompassing frameworks aimed at combining issues of knowledge, affect, and practice. Key examples of this tendency are the notions of algorithmic literacies (Dogruel, 2021; Oeldorf-Hirsch and Neubaum, 2021), expertise (Bassett et al., 2015), lore (MacDonald, 2021), and the study of skill (Gruber and Hargittai, 2023; Hargittai et al., 2020).
To summarize, research on human agency vis-à-vis algorithms has sought to show that, even if people don’t necessarily understand algorithmic ‘black boxes’, they still enact their agency as they live with these technologies (Siles, 2023). Employing primarily qualitative techniques (most notably interviews and focus groups), surveys, and more sporadically computational methods, they have accounted for these agency enactments by examining one specific temporal dimension at a time (either past, present, or future). Most scholars have examined individuals at a micro level, accounting for how they act individually or collectively.
Challenging some of the premises of the control paradigm, these studies have also worked to demonstrate that people retain a certain sense of autonomy when relating to algorithms (Dogruel et al., 2021). Compared to work on algorithmic power, this analysis has relied on a distinct conception of autonomy, one that Savolainen and Ruckenstein (2022: 10) capture with precision:
Autonomy [is] best understood as activated by a related but analytically separable development in human–algorithm relations, where the aim is not to control at a distance but to cultivate an increasingly close and emotion-based relationship with the consumer through the use of personalization-enabling algorithmic techniques.
Autonomy is thus enacted both in the conditions and activities of people’s daily lives and, more explicitly, in efforts to resist algorithms. Accordingly, a prime object of concern in work on human agency has been the significance of resistance as a form of counter-hegemony (Cohn, 2019; Fotopoulou, 2019). Research on resistance also reveals the importance given to collective forms of agency in studies on human agency, illustrated by the practices of so-called ‘gig workers’ and organizational employees (Hidalgo Cordero and Salazar Daza, 2021; Kellogg et al., 2020).
Sensibilities for Studying Fluid Agencies
To supplement the dominant analytical preferences discussed above, we argue for a more fluid approach to the study of agency in relation to algorithms than what has been the case to date in the literature. The concept of fluidity emphasizes notions of flow, convergence, instability, coexistence, friction, and change, rather than solid, opposite poles or definitive states. Fluidity, as Linstead and Brewis (2004: 355) put it, ‘bespeaks uninhibited or potential movement within the space between dualistic categories that substantiate opposing poles’. In critical studies of data and algorithms, similar ways to frame issues of fluidity have centered on the notion of ambivalence. For Hepp and collaborators (2022:10), power must be theorized as ambivalent in that it ‘open [s] up many potentials that can be emancipative’ while also having ‘many negative elements that should not go unnoticed’. Yet, these authors argue, it is precisely this ambivalence that invites (and demands) new perspectives of analysis.
Critical frameworks, such as queer theory as well as decolonial and postcolonial scholarship, have often alluded to fluidity in their accounts of the links between agency, identity, and subjectivity. In this body of work, identity is understood as indefinite or what Aymara scholar Silvia Rivera Cusicanqui, 2018 refers to as the ch’ixi condition. In Aymara, ch’ixi designates a mottled gray tonality produced by the juxtaposition of black and white but that does not entirely result in mixture or hybridization. Ch’ixi thus represents ‘the potency to cross boundaries and embody polar opposites in a reverberating manner’ (Rivera Cusicanqui (2018): 79). The tension between opposite worlds is the source of transformative power and agency in the ch’ixi condition. These ideas also resonate with the ‘borderlands’ concept that Gloria Anzaldúa used in both geographical and metaphorical meaning to name ‘the painful yet also potentially transformational spaces where opposites converge, conflict, and transform’ (Keating, 2009: 319).
Theorized as fluid and liminal, agency is a zone of both creative contact and friction between worlds (Anzaldúa, 2004). Agency, then, does not imply having to always choose between fixed options but rather having the possibility to transit between them or being both, even when these contradict each other. Put differently, fluidity opens up the possibility to occupy different sites of agenciation (Kapoor, 2003).
In what follows, we articulate three sensibilities to move forward the scholarly conversation through an approach that, in its fluidity, can allow grasping the complex ways in which power and agency operate in relation to algorithms: (a) tensions; (b) mediations; and (c) transversalities. We build on various theoretical and empirical traditions, including Latin American communication scholarship, which has not typically been featured in this conversation. This body of work offers productive insights to further understanding of agency as it allows the incorporation of a series of concerns and issues that are not necessarily prevalent in agency theorizations developed in the Global North (for example, how fickle infrastructures, peripheral technologies, and conditions of structural precarity lead to different kinds of relationships with algorithmic systems) (Chan, 2014), opening possibilities for more comparative studies (Tseng, 2022).
Tensions
A first sensibility centers on the need to recognize agency as fraught with tensions. As noted above, there has been a tendency in the literature on algorithmic power to adopt an underlying paradigm of control. Freedman (2014) identified a similar pattern in discussions of the media and argued for rethinking this dominant assumption through a logic of contradiction instead. For Freedman, this alternative ‘emphasizes structure and agency, contradiction and action, consensus and conflict; [and] recognizes the existence of unequal power frameworks but acknowledges that they are not forever frozen’ (Freedman, 2014: 29; emphasis in original). We follow Freedman’s call for theorizing power as fraught with tensions but stress instead how agency can be manifested as both compliance and resistance (singly and simultaneously), rather than exclusively through the logic of opposition (as the notion of contradiction implies).
Interdisciplinary work on the notion of friction is useful to further theorize agency as fraught with tensions. In physics, Cresswell (2014: 108) reminds us, friction ‘is a force which resists the relative motion of two materials sliding and rubbing against each other’ (p. 108). Scholars from different fields have developed the friction metaphor to argue against the view of social phenomena as seamless flows that unfold without tension (Bates, 2017; Lehuedé, 2022). Against this view, as Tsing (2005: 6) puts it, ‘speaking of friction is a reminder of the importance of interaction in defining movement, cultural form, and agency’.
Building on this body of work, we mobilize friction as an analytical category to further the conversation on agency vis-à-vis algorithms. In short, we argue, friction invites recognition of the tensions involved in enacting agency. As we showed above, scholarly discussions of agency have rested on a marked duality: either power is assigned to algorithms or agency is attributed to humans. Even in work that draws on theoretical approaches that have sought to overcome this form of binary thinking (such as actor-network theory), people are usually relegated to specific roles: they are made to act by algorithms more than capable of acting on them. Similarly, underlying the empirical study of users is a tendency to treat agency as an all-or-nothing condition: people either resist algorithms or yield to their power.
Friction requires thinking beyond the choice between either algorithmic power or human agency. As Cresswell (2014: 109) puts it by drawing on the metaphor of mobility, this is because friction ‘suggests an ambiguous, two-sided form of relative stillness that is both impeding mobility and enabling it’. Friction thus posits agency as diverse expressions that contradictorily live together. This approach can inform analyses of algorithms by stressing how individuals and groups can simultaneously resist, comply, challenge and obey algorithms in their daily lives.
Consistent with such an emphasis on tensions, some researchers have begun to empirically show how algorithmic power and human agency depend on one another in complex ways. Considering the case of collective action, Treré (2018) theorized the relationship between algorithms and social movements in terms of ‘mutual shaping’. In a similar manner, Siles (2023) examined the ‘mutual domestication’ between users and platforms’ recommendation algorithms. The notion of domestication emphasizes how people assign meaning to algorithms rather than try to transform or subvert them. Savolainen and Ruckenstein (2022: 13) also developed the notion of ‘co-evolution’ as ‘a collaboration where action and intentionality are produced in the interactions between human and machine’.
Michel Callon provided a vivid illustration of this sensibility toward tension in his discussion of Amazon’s algorithmic recommendations. For Callon (2017: 197–198), algorithms are ‘equipments’ combined in a multiplicity of ‘sociocognitive prostheses’. In his words, algorithms ‘constitute a common space of shared and structured infrastructures that act, make act and are acted in return, transforming and redistributing themselves with the events they help to frame’ (Callon, 2017: 212, our translation). Callon thus envisions the user-algorithm relationship as a process in constant tension from which both algorithms and people acquire and enact shifting capacities to act (Siles et al., 2022).
Tension helps to further investigate–using Cresswell’s (2014) formulation–the ‘two-sided form of relative stillness’ that is at the heart of stability and transformation: it is both a force that slows things down and that enables movement. In other words, tension demands further consideration of the hypothesis that agency is enacted both in the transformation of algorithms as well as in their permanence. In keeping with this observation, researchers could continue examining the tension experienced and created by efforts to resist algorithmic power. But, as authors such as Cresswell (2014) and Tsing (2005) remind us, tension also keeps power in motion. Scholars should thus also explore tensions as enacted through the domain of the infrapolitical, that is, practices that are not aimed at disrupting algorithms but that still represent a claim for agency and autonomy in their own right (Scott, 1990; Siles, 2023).
Mediations
Building on the analytical significance of tensions introduced above, our second sensibility reframes the algorithmic power/human agency binary as a situated process. We take a cue from the work of Jesús Martín-Barbero to theorize this relationship as a ‘mediation’. Martín-Barbero argued for ‘tak [ing] the study of reception out of communication defined as circulation of messages, effects and reactions, and put it into the field of culture: the conflicts which articulate culture, the mestizajes which weave it together and the anachronisms which sustain it’ (Martín-Barbero, 1993: 221–222, emphasis in original).
Martín-Barbero’s (1993) dictum of moving away from the study of the media to think about mediations challenges the premise that algorithms are a ‘source’ of power to privilege instead how people relate to algorithms within the conditions that shape their daily lives and experiences. In this sense, our second sensibility fosters exploring the situated interconnections between algorithmic power and human agency. Otherwise put, a focus on mediations underscores the need to better situate people’s relationships with algorithms within the particular cultural conditions in which these relationships take place. This is because agency is neither experienced nor enacted in the same way everywhere and should thus be situated in specific contexts and acknowledged as a result of historical processes (Mahmood, 2001).
Interdisciplinary work on friction is once again useful to make our case for algorithmic mediations. In her exploration of friction in the forests of Indonesia, Tsing (2005) argued that universal principles need to become specific in order to exist. Friction thus takes place when the universal lands on the real terrain of particular places. In Tsing’s (2005: (4) words, ‘Cultures are continually co-produced in the interactions I call ‘friction’: the awkward, unequal, unstable, and creative qualities of interconnection across difference’. In this sense, friction refuses the notion that universal forms of power operate in seamless ways. Applying this insight to our case at hand, we argue that algorithmic power can only experience becoming when it encounters the specificity of friction in particular sites or bodies, that is, particular mediations.
As noted above, there has been a tendency to theorize algorithmic power as a universal force (Milan and Treré, 2019). But, examined through the lens of mediation, algorithmic power needs to encounter the real terrain of situated places in order to exist. In Tsing’s (2005: 10) terms, this mediation ‘gives purchase to universals, allowing them to spread as frameworks for the practice of power. But engaged universals are never fully successful in being everywhere the same because of this same friction’.
A sensibility toward mediation thus seeks to further acknowledge the conditions that make the prevalence of specific forms of algorithmic power and the enactment of knowledge, affect, and practice possible in the first place. To this end, it is key to go beyond the focus on privileged, young users – the most frequently studied group in critical algorithm studies, as Hargittai et al. (2020) note – and incorporate into the analytical enterprise an empirical consideration of marginalized and less privileged groups. People relate to algorithmic mediations differently, depending on intersectionalities of race, gender, class, and ability, among other markers of identity, and do so with varying degrees of access to capital, knowledges, and technologies. For example, Noble (2018) showed the magnitude of harms entailed by biases in Google’s algorithms specifically for Black, young, and marginalized women. Buolamwini and Gebru (2018) demonstrated the problematic relationship between gender and race in facial recognition systems and the relevance of intersectional analysis for auditing algorithms. Conducting these kinds of analyses beyond the Global North is crucial to understand agency not only as a political and economic process but also as an embodied experience.
Accounting for algorithmic mediations requires methods that properly situate agency in its cultural and social contexts. Longitudinal surveys or panels could help identify specific sociodemographic variables that shape literacies, experiences, and expertise with algorithms (Gran et al., 2021). Moreover, ethnography also has the potential to help turn algorithmic power from an analytical premise to the product of a situated encounter in local places (Tseng, 2022). Recent studies that have employed such methods as diaries have begun to show how the formation of algorithmic awareness is a thoroughly cultural and temporal process (Schellewald, 2021; Siles et al., 2022). Mixed-method research designs, which have been rare in the study of agency and algorithms, could help combine data-centric approaches with a sensibility for context and the lived experiences of people.
In their analysis of the role of algorithms in shaping consumer cultures, Airoldi and Rokka (2022) build on Du Gay et al.’s (1997) notion of articulation to provide an insightful example of potential work on mediations. These authors seek to overcome the tendency to consider that either markets can completely control people through algorithms or that empowered consumers act independently from algorithms. Instead, they argue that ‘algorithmic systems “articulate” consumption and production processes within digital environments” (Airoldi and Rokka, 2022: 412). In this view, algorithms acquire meaning only by articulating production and consumption dynamics in specific cultures. By examining how this work of articulation occurs, future work on mediations could help bridge the interest in the production of algorithms and the research on users of algorithmic systems, which have remained relatively separate to date (Christin, 2020; Seaver, 2022).
Transversalities
Implementing our previous sensibilities also requires the work of connecting different levels of analysis and temporalities. This is because researchers have studied issues of algorithmic power and human agency as operating on different scales. Scholars interested in algorithmic power have mostly adopted a macro perspective, one that is consistent with their view of power as a universal principle that transcends micro factors. Alternatively, when they have been interested in human agency, scholars have espoused a micro approach that privileges people’s capacities to act through knowledge, affect, and practice.
This analytical divergence between macro and micro approaches is reminiscent of Misa’s (1988) account of how historians have made sense of causality in the study of technology. Misa’s (1988: 319) words could aptly describe the literature on agency in relation to algorithms: [Scholars] writing large-scale [...] accounts deploy the Machine to structure social change, while as soon as the [...] microscope is unveiled, the Machine as such dissolves. [...] From a shop-floor perspective, the Machine is an irrelevant abstraction, and what makes [agency] is individuals [...] in conflict or accommodation.
Building on this observation, our third sensibility emphasizes the need to further explore transversalities between various levels of analysis and temporalities. By transversalities, we refer to the iterations, intersections, and continuums between spaces and scales.
One transversal connection can be established between the study of individual and collective expressions of agency. Focusing on either the individual or the collective has made valuable contributions to the study of agency, but has also come at the price of making invisible agency enactments that cannot be reduced to either one of these poles, that evolve over time, or that are both. Considering the collective as an aggregation of individual actions, as has often been the case, can limit our understanding of collective agency by downplaying larger dynamics that cannot be accounted for by individuals through their discrete relationship with algorithms. Moreover, reducing the study of collective agency exclusively to how social movements or ‘gig workers’ act together with or against algorithms also runs the risk of making invisible more organic practices that lack political articulation but that are collective in nature.
Cases that cannot be reduced to either individual or collective agency (or that blend both) afford an ideal possibility for implementing a transversal view of agency. As we noted, several scholars have emphasized autonomy and choice as core dimensions of practical-evaluative enactments of agency. These are typically envisioned as individual rational choices or as goals shared by a group. A transversal understanding of agency can account for practices that are not strictly individual choices nor have common goals. For example, people share their passwords or use streaming services individually (and thus collectively alter the work of algorithms that are optimized for individual ‘profiles’ on certain platforms); family members share devices where their accounts overlap; and individuals often employ Virtual Private Networks (VPNs) to obtain certain algorithmic recommendations that they couldn’t receive otherwise. These practices, which are common in the Global South, could be analyzed as instances of ‘affective’ or ‘intimate’ publics that form through the intersections of algorithms and collective expressions of affect in daily life (Papacharissi, 2015; Siles et al., 2019).
Klinger and Svensson (2018) have argued for a similar transversal process in their analysis of the analytical levels involved in claims about the agency of algorithms (they prefer this notion over that of power). They suggest exploring the meso level of organizations rather than locating the agency of algorithms at the macro level of society or the micro level of individual practices. In their words, ‘Instead of fearing that machines will take over the world [...], we should start to focus on the humans and organizations behind the machines and what structures and conditions shape their actions’ (Klinger and Svensson, 2018: 4666).
Mechanisms have provided a valuable means to advance theory about a range of social phenomena, including the micro-macro distinction (Hedström and Swedberg, 1998). Thus, one way to implement a sensibility towards transversality is to consider the mechanisms that connect micro, meso, and macro levels. In one of the few studies that have implemented such an approach, Kellogg et al. (2020) developed a typology of mechanisms that accounted for both algorithmic control and worker’s experiences in three areas of organizations: direction (recommending and restricting vs manipulation and disempowerment); evaluation (recording and rating vs surveillance and discrimination); and discipline (replacing and rewarding vs precarity and stress). In this way, they approached the issue of control by considering the tensions between algorithmic power and human agency that formed between the micro level of individuals and the meso level of organizations (Cellard, 2022).
Another possibility for implementing this transversal approach to agency is temporality. The ch’ixi condition argues for multiple temporalities where past, present, and future coexist. In a similar manner, a key in Emirbayer and Mische’s (1998) temporal approach to agency is how its three dimensions – iterational, practical-evaluative, and projective – relate to one another. Neither Rivera Cusicanqui (2018) nor Emirbayer and Mische (1998) conceive of these dimensions as successive but rather as simultaneous. As shown above, the focus in the scholarly literature on algorithms has been on the present. Moreover, although scholars have paid attention to all three dimensions, they have focused on a single dimension at a time rather than accounting for their interplay and evolution. Alternatively, studies could show the connections between past, present, and future of various kinds of algorithmic use. This would make it possible to identify instances of beliefs, emotions, and practices that have a longer history than anticipated and those that are more novel. Historical approaches could also help develop genealogies or trajectories of the relationship between technologies, imaginaries, structures of feeling, and practices (Gómez-Cruz and Harindranath, 2020; Papacharissi, 2015).
Concluding Remarks
The perspective we have developed in this article avoids both deterministic accounts of technological power and approaches to agency that depict it as an all-or-nothing condition. In many ways, the so-called ‘algorithmic age’ or ‘algorithmic turn’ has meant a return to deterministic perspectives that posit algorithms as technologies that are external to society and have intrinsic capacities to overpower humans. Examples of this tendency can be found in both academic and journalistic discourses. These accounts go against decades of research that has challenged technological determinism on both theoretical and empirical grounds. The prevalence of deterministic accounts rests on the premise that algorithmic power is essentially different from previous forms of technological power. Yet, as many authors have recognized, providing systematic, empirical evidence for testing the historical singularity of algorithmic power has proved a difficult task (Cardon, 2018; Schwarz, 2021). To be sure, we do not seek to deny the consequences of algorithmic mediations for marginalized populations. We think it is imperative to keep ‘auditing’ algorithms and revealing their inherent biases. Instead, we have argued for broadening the type of dynamics through which the actions and consequences of both algorithms and people unfold and can be studied.
A common analytical response against deterministic approaches has been to empirically examine human agency. Focusing on users as a reaction to claims of technological power has numerous historical precedents in various fields of knowledge. In communication studies, the analysis of what audiences do with the media (rather than what media do to people) followed years of discussion about ‘hypodermic needles’ and the media’s ‘powerful effects’ in society (Katz, 1980). This resulted in some of the most productive theories in the field: the so-called ‘two-step flow’ of communication, the ‘uses and gratifications’ framework, work on the varieties of audience interpretation, and domestication theory, to name just a few. In science and technology studies, the recognition of the role of ‘users as technological agents’ nuanced well-established accounts of the role of technology producers (Kline and Pinch, 1996). Misa (1988) followed a parallel transition between historical studies of technology that attributed a deterministic role to artifacts and a broader consideration of the role of individuals in the making of history. The turn to human agency has been equally productive in critical algorithm studies. It has helped to account for the lived experiences of people in the context of algorithmization.
However, despite its many contributions, focusing on either algorithmic power or human agency is limited in its capacity to explain the intersections, entanglements, and entrapments of people and algorithms through which agency emerges and is enacted in everyday life. By foregrounding issues of tensions (as a supplement to thinking in terms of either/or binaries), mediations (to reverse the tendency to treat algorithms as an ahistorical and universal force), and transversalities (to explore various temporalities and scales of analysis), we have hoped to provide various means to situate the study of agency in relation to algorithms within the larger history of research that prioritizes the interweaving of technologies and humans in embodied, situated, and dynamic perspectives.
Footnotes
Acknowledgements
We wish to extend heartfelt thanks to Dominique Cardon and Angèle Christin for their most helpful comments on previous versions of this manuscript. We also thank the editors of this special issue, Stephan Görland and Andreas Hepp, and the anonymous reviewers for their excellent suggestions.
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 Vicerrectoría de Investigación, Universidad de Costa Rica.
