(Re)assembling algorithmic resistance: An integrative review of research-machines

Assembling the review-machine

The machine assembled for the goal of understanding productions of AR is an integrative literature review (e.g., Feldman, 1971) that analyzes how AR is produced across disciplines, designs (e.g., experimental, conceptual, empirical), types of data, methodologies, research interests and questions (Whittemore and Knafl, 2005).

By gathering, synthesizing, and analyzing literature on a topic, this review approach seeks to integrate, map, and bridge in order to generate new perspectives (Torraco, 2005: 356). Reflecting the disciplinary breadth and dispersal of discussions surrounding AR, it attempts to span the diverse conversations across the various communities of practice involved. Examining how and along what research-assemblages a specific understanding of AR is brought into being, how and why it has been studied within a certain problem space can then address needs for ‘review, critique, and the potential reconceptualization’ (Torraco, 2005: 357).

This integrative approach, however, does not claim to be complete, representative, or entirely comprehensive. Rather, the research-machine designed here – comprising databases, search terms, data annotation, and, crucially, the researcher – attempts to provide cartographic richness, to integrate and cluster various approaches to producing AR, and, on that basis, interrogate how and along what dimensions such production occurs.

As all research-machines territorialize knowledges, this work, too, mediates micropolitics and is complicit in generating certain realities at the expense of others. Territorializing the productions of AR is thus in itself reflexive, recursive, rhizomatic, an enactment that is both productive and reductive; it is a specific reassemblage by disssembling with notable limitations: For example, reviewing academic texts as outputs of research-machines allows for tracing territorializations of AR along their identifiable components: fields/disciplines, theories, or methods. Other defining elements, for example, institutional barriers, material constraints, personal motivations of researchers, or political economies of scholarly reception, cannot be accounted for in this way. However, what the integrative review can provide is a comprehensive mapping of how different studies produce a diverse landscape of AR, making the constitutive role of research-machines visible and demonstrable. Conversely, the research-machines lens is what transforms the review from a summary of findings into an interrogation of how those findings came to exist.

Data collection

Initial data were retrieved in January 2025 by identifying relevant publications through searching the ‘Web of Science’ (WoS) database for peer-reviewed publications containing a combination of the words ‘algorithm’ (including variants like ‘algorithmic’) and a signifier for resistance (‘resistance’, ‘refusal’, ‘antagonism’), either in their title or abstract. Since resistances against ‘algorithms’ and ‘artificial intelligence’ are often used interchangeably, a subsequent search refined the results to include the latter term. After screening and filtering, 65 publications were included in the corpus. ¹

Analysis of the initial corpus was accompanied by a second step of data collection consisting of forward and backward searches for further resources in the corpus. If an understanding of (algorithmic) resistance was specifically stated or referred to, the corresponding source was selected for screening (backward search). Forward searches using the platforms Google Scholar, ResearchGate, and litmaps.com were conducted if items presented their own understanding of AR to identify papers referencing it. This additional corpus particularly included publications that did not discuss AR as their primary concept, used alternative terminology, or were not listed in the WoS database. Additionally, four articles published during the research process were identified via publication alerts and included in the corpus after screening.

Following the exclusion of duplicates, items that did not adhere to the inclusion criteria, and items that were not retrievable, this yielded 41 additional publications, increasing the total corpus size to 106 (see Figure 1).

OPEN IN VIEWER

Figure 1.

Corpus, PRISMA flow diagram.

Arguably, this approach is flawed in a number of ways. It fails to account for varying terminologies in both technologies (e.g., automated systems) and resistances (e.g., defiance, opposition) and excludes practices that conceptualize regimes of domination (datafication, visibility, labour, capitalism) and their contestations differently (e.g., as ‘People's Practices in the Face of Data Power’, Crooks et al., 2024; or as resisting the ‘data-driven society’, Milan, 2024). However, in attempting to gain a broad (rather than comprehensive) corpus of various ARs, the search strategy provided a potent starting point that was effectively supplemented by the extended screening methods.

Coding and analysis

Items were systematically classified according to their context (based on the disciplinary background of authors and publishing venue), approach (conceptual vs. empirical), methodological choices, and domain of algorithmic application (e.g., recommender systems). Text passages were identified that referred to AR theoretically or described empirical examples. Subsequently, these became subject to an in-depth analysis guided by the research questions and using open coding, identifying practices of resistance, their characteristics, and conditions. Despite this sequentiality, especially during early steps, the coding phase was highly iterative, continuously moving between reviewing items and supplementing or revising codes on the basis of new findings.

However, as conceptual territorializations result from the totality of assembled components rather than singular choices and conditions, analyzing research-machines poses the problem of their comparison. Even if all elements of research-machines are known (which, as described above, was not the case), their influences can only be understood in relation to and through their interactions with all other elements assembled. Yet, adopting the integrative perspective described revealed a kinship across patterns, rhythms, and styles of assemblage. It showcased how related manufacturing processes of AR produce similar outcomes, even despite terminological differences. For example, initial codes such as workers’ tactics and platform gaming were later consolidated into Workers’ Counter-Conduct as shared patterns across labour contexts emerged. Similarly, repair, care, and maintenance initially appeared as separate practices before being recognized as expressions of what was later named Emergent Agencies.

Thus, the analysis revealed genealogical connections that allowed for gathering productions of AR not along terminologies, but along their definition of problem spaces: similar locations, problematizations, agencies, and respective understandings of resistance. These led to the identification of seven clusters of research-machines, each representing an ideal-type of specific territorializations of AR (‘Algorithmic research-machines: Contrasting cases’ section). For example, studies referring to Scott's (1985) notion of everyday resistance, using ethnographic methods, and focusing on creators on social platforms consistently produced AR as embedded opposition rather than overt refusal. The interactions of these elements assembled in the research process defined similar problem spaces that are generative of particular understandings of AR, across vocabularies and without essentializing or limiting practices to specific labels.

Such aggregation, while attempting to account for clusters’ inherent heterogeneity, is necessarily reductive. Items are never entirely representative of one cluster, but frequently span and transgress multiple. Thus, their exemplary classification is to be understood as tentative, as emerging from commonalities in their productions of AR while sustaining difference.

Particular territorializations of problem spaces define the properties of the resistance they produce: whether it is open or clandestine, ephemeral or sustained, seen as unintended or articulated as political intervention. To understand how research-machines produce particular characteristics of AR, subsequent coding was focused on dissecting axes along which these properties manifest (‘Axes of ARs’ section).

Algorithm Aversion

Aversion-machines focus on AR as users’ expressions of preferences for decision-making in the context of automated systems. Studies in this cluster (e.g., Bankuoru Egala and Liang, 2024; Isaac et al., 2024; Mahmud et al., 2023) construct resistance as a reluctance to adopt algorithmic technologies due to concerns over autonomy, accountability, or unfamiliarity (Dietvorst et al., 2015) as a form of maladaptive coping. Originating in behavioural psychology and technology acceptance research (Davis et al., 1989), these studies follow users’ avoidance of algorithmic recommendations because they perceive them as inaccurate, biased, unfair, or unreliable. They span fields such as (health) information systems research, medical AI, or managerial algorithm aversion, producing resistance mainly as individual and cognitive, structured around questions of attitude and willingness to adopt. AR is thus framed as subtle, cognitive, and affective disengagement from algorithmically mediated processes, positioning agency at the intersection of skepticism, experienced opacity, and a preference for human judgment. While these machines mostly see aversion as a form of resistance, it is sometimes also framed as a passive reaction, contrasting with active forms of AR (Zhang et al., 2024).

For Aversion-machines, resistance is a friction to be understood, factorized (Mahmud et al., 2023), cross-culturally compared (Liu et al., 2023), and ultimately mitigated. Foregrounding professional judgment, risk perception, and trust calibration, they construct AR as an individualized reaction while largely excluding political or collective dimensions. They emphasize human values within specific institutional settings, use cases, or interfaces; however, they remain limited to parameters of usability and acceptance. Grounded in a perspective that seeks to increase efficiency and to allow for ‘better’ decision-making, they enact resistance as human-centric, as a problem to be solved for achieving mutually beneficial human–machine relations.

Technopolitics

Technopolitics-machines explore AR through collective mobilization aimed at contesting digital (platform) power and algorithmic governance. Often embedded within broader political mobilizations (e.g., anti-surveillance campaigns, platform labour organizing, ‘data activism’, Milan and Velden, 2016), such resistance manifests in organized campaigns and legal advocacy through civil society organizations or grassroots movements. It aims to reshape laws, policies, or organizations and emphasizes issues such as data rights, algorithmic bias, digital sovereignty, privacy, and autonomy. Technopolitics-machines construct resistance as a strategic negotiation of power dynamics, wherein various (human and nonhuman) actors disrupt or contest algorithmic decision-making. They assemble agencies as collective and hybrid along the use of technological tools, as embedded in organized actions, and as encompassing institutional actors, legal structures, and corporate and governmental agencies. Localized in movements that span digital platforms and offline actions, they enact AR as political, as practices that challenge conditions sustaining or allowing for algorithmic control.

While Technopolitics encompasses a wide array of practices, their enactments of AR can be further distinguished into subcategories: emphasizing collective struggles against platform power, machines of Algorithmic Solidarity are more likely to be rooted in research on Latin American digital activism or transnational/feminist STS (e.g., Milan, 2015; Yu et al., 2022). These are reflected within conceptions of resistance that centre on mutual aid, networked protests, the formation of horizontal infrastructures of dissent, care, and solidarity as broader movements against extractivism or platform colonialism. By assembling research methods such as participatory design or activist ethnographies, they de-emphasize the role of (formal, hierarchical) organizations. Instead, they highlight mutualism, Indigenous perspectives, intersectional analysis, and communal infrastructures of resistance.

By contrast, Professional Technopolitics can be found in the example of grassroots auditing collectives, networks of activists exposing algorithmic logics (e.g., in hiring, welfare, policing) or arts-based interventions (e.g., Pereira et al., 2022; Wiehn, 2022). While configuring resistances largely similarly, professional approaches often require technical access or literacy and thus preferentially assemble agencies of tech-savvy activists at the cost of broad inclusivity.

Workers’ Counter-Conduct

Algorithmic technologies colonizing workplaces transform the exercise of workplace control (Kellogg et al., 2020). Research-machines of Workers’ Counter-Conduct thus construct resistance as opposition to algorithmic management and the automated enforcement and interpretation of computational metrics. Applied within environments marked by precarity and power asymmetries, algorithmic systems are positioned as means of control and are thus countered with workers’ resistance, for example, through practices of withdrawal, non-compliance, and small acts of sabotage. Examples include Chan's (2021) work on labour control or Arubayi's (2021) study of ride-hailing drivers.

Situated primarily in studies of work, labour, and organizations, these research-machines construct resistances mostly as covert, continuous, and highly adaptive. However, and broadly understood as ‘workers’ attempts to maintain work autonomy and re-claim a sense of control over the labor process’ (Chan, 2021: 7), resistances may also manifest as collective contestations (Kellogg et al., 2020; Novianto, 2024).

Theoretically grounded, for example, in Braverman's (1974) labour process theory and Edwards’ (1979) concept of contested terrain, these studies analyze algorithmic management as the latest iteration of workplace control and position AR as a continuation and reinterpretation of labour struggles. Methodologically, such assemblages rely on ethnographies and (digital) fieldwork, with researchers participating in the work process or within workers’ chat groups (Yu et al., 2022). They foreground precarity and surveillance, and emphasize how workers develop situated knowledges around algorithmic infrastructures and enact tactical agency in the absence of formal protections.

Materially, such resistances are entangled with organizational, physical, and digital infrastructures of workplaces, with workers resisting by manipulating interfaces, slowing down processes, or gaming metrics and systems. As managerial control through algorithms is most clearly and commonly exercised in the gig- or platform economy, numerous studies particularly engage with practices of gig workers, food delivery drivers, or users of freelance platforms (e.g., Arubayi, 2021; Bulut and Yeşilyurt, 2024; Guerra and d’Andréa, 2022; Vasudevan and Chan, 2022). Resistant practices such as refusing certain jobs, altering delivery routes, or using multiple identities are thus intricately entangled with the material affordances of the respective platforms.

Mundane Opposition

Mundane Opposition-machines are documented in the corpus through examples of subtle, non-institutionalized acts of petty defiance on social platforms, within matchmaking systems, or along other forms of algorithmic recommendation. Examples include works on visibility management (Bishop, 2019), algorithmic knowledge practices (Cotter, 2021), or resistance against recommendation algorithms (DeVito et al., 2017). Here, AR results from ordinary users and captures how they challenge or reroute algorithmic influences on their experiences. These machines focus on AR not primarily as politically motivated activism, but as rejection of algorithmic results out of unease, habit, or irrelevance, for example, resulting from a perceived suppression of certain social identities (Karizat et al., 2021). Such resistances are scattered, rhythmic, and deeply entangled with the mundanity of digital life, thus decentering (yet not abandoning) notions of revolutionary or strategic opposition. This frame highlights practices like selective misinterpretation, frictional use, disengagement (Magalhães, 2022), micro-acts of non-compliance, and negotiations to retain self-expression (Chartrand and Duguay, 2025). Here, AR is produced as creatively navigating, managing, and (alternatively) appropriating algorithmically mediated environments, often by making sense of algorithms through concepts such as folk theories (DeVito et al., 2017; Siles et al., 2020) or imaginaries (Bucher, 2017). By capturing AR as an ambient force, woven into the routine of digital life, frequently dismissed or overlooked, yet widespread and sometimes subtly transformative, these machines draw from notions of everyday resistance (Scott, 1985) or mundane tactics of quiet encroachment (Bayat, 2000; de Certeau, 1984).

Mundane Opposition-machines tend to enact inclusion and exclusion fluidly, as nearly all users are seen to engage in some form of everyday resistance. Yet, the meanings, recognitions, and impacts of these acts vary by context. However, and contrary to machines of Workers’ Counter-Conduct, the regimes to which their resistances relate are usually less oppressive, determined less by external necessity, and thus characterized by a somewhat voluntary subordination. Typically constructed as individual or only loosely collective, as uncoordinated and lacking self-perception as resistance, these resistances' visibility is extremely low or covert. However, their material outcomes may cumulatively distort infrastructures and steadily alter the constantly evolving relations between users and platforms. While rarely framed as resistance by users themselves, AR is produced when mundane practices become assembled through research. Researchers thus affect the production of resistance in this cluster even more than within the machines discussed so far.

Alongside everyday users of platforms, it is especially influencers, creators, and marginalized communities whose actions are rendered resistant through this machine (Bishop, 2019; Cotter, 2021). Their tactics particularly include ‘playing the algorithm’, gossiping over functions, ironic subversion, or resigning from platforms. In these studies, often conducted as ethnographies of communities, platforms, or interfaces, the affective labour and the identity work required to counteract soft governance and regimes of visibility are highlighted and entangled with the specific logics of online communities and platform affordances.

Emergent Agencies

Rooted in STS, feminist HCI, and design justice approaches, these research-machines conceptualize AR not as a matter of overt refusal or politically articulated opposition, but as situationally emergent, relational, and material (e.g., Ettlinger, 2018; Qadri and D’Ignazio, 2022; Velkova and Kaun, 2021). Critiquing peripheral and detached notions of resistance, these studies see AR as entangled, as co-produced within heterogeneous assemblages of humans, algorithms, devices, and social structures (Ettlinger, 2018). Within a vocabulary of resistance as practices of care, maintenance, and repair, these machines foreground the labour of fixing broken systems by addressing and mitigating algorithmic harms (Velkova and Kaun, 2021). AR is emphasized as sustained engagement, through and along infrastructures, that results in mutual recalibrations of all elements involved, even when unintentional. Drawing particularly from Foucault's (e.g., 1982) late scholarship, these machines highlight a making-use-of rather than the rejection of digital environments and emphasize productive, cyborg forms of resistance that blur classifications through their simultaneous complicity and opposition (Ettlinger, 2018; Siles et al., 2023).

Machines of Emergent Agencies depart from binaries of power and opposition and offer a perspective on resistance where agency is not portrayed as something that is possessed in advance, but as arising through ongoing, situated practice. Such agencies are highly relational, as they do not reside in users or systems, but emerge through contextual human–nonhuman entanglements. Drawing from notions of inherently broken systems and a continuous demand for their stabilization (Graham, 2010; Jackson, 2014), these machines construct resistances as a creative, reparative practice. As affirmative politics of sustaining alternatives and rethinking infrastructure as a site of struggle, AR emerges as modest, slow, and long-term oriented. Primarily territorialized as ambiguous or ambient, as occurring through mundane interfaces and within daily routines, it departs from a focus on visibility and intentionality strongly emphasized by other research-machines.

Professional Autonomy Defense

Professional Autonomy Defense-machines produce AR as centred on individuals in expert roles and as resulting from frictional relations between institutional and organizational logics, professional ethics, accountability norms, and discretionary judgment (e.g., Hanemaayer, 2021; Jussupow et al., 2022).

AR is portrayed as a struggle over professional epistemic authority that is challenged by automated systems and their ethical implications. It is framed as an attempt to maintain discretion often found within scientific, educational, medical, legal, managerial, and other domains, as the protection of expert legitimacy from algorithmic disruptions (e.g., Christin, 2017). Thus, such machines produce resistances as highly intentional and grounded in (privileged) socio-organizational roles, often leveraging bureaucratic opacity or processual friction. Practices involve subtly ignoring, modifying, or overriding recommendations, with resisters drawing from collective professional norms and ethics to justify their actions. Resistances are inherently embedded, enabled, and constrained within their organizational context and thus assembled in relation to the respective organizational or professional communities.

Because they configure agencies along professional identities and their institutional positions, they privilege actors with access to decision-making positions or credibility within their fields. Thus, while addressing resistance in workplaces, these research-machines neglect workers lacking the standing or opportunities to challenge algorithmic systems and their outputs (e.g., gig- or frontline workers described in by Workers’ Counter-Conduct-machines).

Epistemic and Ontological Refusal

Prominent within critical data studies, post/decolonial media theory, and feminist technoscience, these machines conceptualize resistance not merely as opposition, but as outright refusal: as rejection of the epistemic and ontological terms upon which systems operate (e.g., Ganesh and Moss, 2022; McQuillan, 2022). Here, resistance is situated within a critique of data capitalism, with practices ranging from critical stances toward algorithmic systems to their abolition: non-participating, challenging systems’ terms of legibility, rejecting algorithmic epistemologies, and advocating for a return to pre-algorithmic, human-centred forms of interaction and governance. Such resistance rejects the co-optation into algorithmic logics and instead seeks to unsettle dominant algorithmic knowledge regimes (Jarke et al., 2024) and their capitalist embedding. These machines emphasize struggles not merely against individual applications, but instead oppose wider algorithmic knowledge projects (Ganesh and Moss, 2022): the material and epistemic foundations that support present-day algorithmic infrastructures, their classifications, truths, and legitimacies.

Invoked through theoretical analysis, critique, and speculative advocacy for action, the AR these machines speak of is produced primarily through scholarly debate rather than empirical studies. While providing blueprints for practices, interrogating and disrupting demands for their visibility, legibility, and thus what counts as resistance, their practical examples remain largely conceptual. Moreover, such resistance presupposes the ability to say ‘no’, a capacity particularly constrained for marginalized or vulnerable groups, who often lack both meaningful choice among technologies and any real option to refuse them.

Intentionality

Intentionality describes the degree to which resistance is produced either as a conscious, deliberate act of opposition or as a habitual, affective reaction. Capturing whether resistance is rendered as an active choice or as a byproduct of routine engagement in algorithmic systems, it fundamentally relates to the ways in which (and for whom) research-machines assemble agencies. Intentionality thus varies across machines that privilege strategic actors and those acknowledging ambivalence and unconscious tactics. Frequently, it relates to requirements of knowledges, literacies, and expertise to distinguish high- from low-intentional resistances. For example, Mundane Opposition highlights resistance primarily as tactical coping rather than premeditated action (as found in Professional Autonomy Defense).

Scale

Resistance is produced at different levels, from individual actions to global movements. This axis considers whether practices are localized in individual refusal, group-based coordination, or systemic/infrastructural transformations and whether they are spontaneous, synchronized, tactical or strategic. Examples of Algorithm Aversion and Mundane Opposition occur primarily at the individual level, as acts of avoidance, rejection, or reinterpretation within personal decision-making spaces. In contrast, Technopolitics moves resistances into collective realms through coordinated campaigns, policy interventions, and global networks of activists. Professional Autonomy Defense typically scales to the organizational level, with pockets of coordination across professional networks or institutions. Scale is prefigured by the assembled agencies of the respective resisters and reflects what constellation of elements is given the capacity to resist (individuals, groups, movements).

Visibility

Resistance can be spectacular and overt (e.g., protests, strikes, campaigns), visible but subtle, or covert, taking forms that escape platform and public detection. While the repair work described in Emergent Agencies is often only produced through careful in-depth investigation, visibility is central to Technopolitics and part of its repertoire to amplify and network resistances. On the other hand, resistances in Professional Autonomy Defense are usually discreet, enacted through bureaucratic maneuvers or interpretive latitude, rather than overt confrontation. Visibility thus also reveals machines’ micropolitics of recognition: What counts as real resistance is not just determined by who is resisting, but also depends on who is watching and what means of making visible (e.g., research methods) are available to them. Furthermore, visibility can be described as a strategic trait: platform workers’ resistances often necessarily remain illegible to platform metrics. While being plotted and shared within clandestine solidarity networks or off-platform spaces, they intentionally reject visibility due to vulnerability.

Materiality

Spanning from discursive interventions to infrastructural dismantling, this axis distinguishes material, embodied doings from symbolic critique, though these often overlap into hybrid assemblages; it refers to the extent and scale to which resistances are assembled from symbols, discourses, physical acts, bodies, technologies, and infrastructures. Emergent Agencies-machines foreground the embodied, infrastructural labour of repair and care, the material practice of correcting flawed systems through interventions or the creation of alternatives. Mundane Opposition engages with tools, apps, and algorithmic materialities as sites of struggle, but can also take symbolic forms (e.g., aesthetics and discursive framings). By contrast, Epistemic and Ontological Refusals operate mainly in symbolic or discursive terrains, resisting the classification of systems and logics of algorithmic governance (at least when not calling for their outright destruction). Materiality is a relevant vantage point for the effects of resistances: As interactions shape, construct, and modify algorithmic systems, resistance can become materially inscribed into them, making them sites of reconfiguration rather than mere negation. This is not restricted to material practices: While resistances can be predominantly symbolic (e.g., advocating for fairer systems), they can shape technological trajectories and thus manifest tangible material consequences.

Temporality

AR is a matter of temporal dynamics, of rhythm and duration. Mundane Opposition and Algorithm Aversion tend to be episodic or ephemeral, linked to specific contexts or moments of friction, while the aversion itself may be sustained as ambient rejection. Professionals’ resistances and those produced within machines of Emergent Agencies are embedded in ongoing negotiations and recurring cycles of decision-making (rather than singular acts), enacting resistances as rhythmic. These modes reveal how resistance becomes part of routine interactions with systems (and their operations) without necessarily being fully oppositional. However, effects may be cumulative, result in outright breakdown or refusal, and thus transform resistances over time. At the other end of the spectrum, Technopolitics supports resistance sometimes as rhythmic (occasionally ignited), but primarily as sustained, sometimes spanning years of campaigning and lobbying.

Relationality

While Materiality reflects material-semiotic questions of resistances, relationality delineates how resistance is configured in relation to (human, technical, institutional) actors and their agencies. For example, Algorithm Aversion-machines explore the technological factors that spark AR, but their resistances remain centred on human experiences and situated in individual cognition. In contrast, Emergent Agencies’ resistances mobilize nonhuman actors as co-constitutive entities, including sensors, interfaces, apps, data, and algorithms themselves. Technopolitics highlights the interdependencies between activists, policy makers, infrastructures, and technological affordances. Both Workers’ Counter-Conduct and Professional Autonomy Defense-machines are relational in a hierarchical sense: their human-nonhuman entanglements occur within power asymmetries structured by their institutional/organizational context. However, while the former focuses on AR in relation to affordances of technologies and work regimes, the latter emphasizes professional ethics and identities.