Abstract
Scholarship on “algorithmic management” focuses on how employers use algorithms and digital devices to routinely control workers. It also looks at how workers resist this control. But how does algorithmic management affect the ability of employers to respond to collective action? The answer is important because algorithmic management is the “new contested terrain” of labor struggles. Drawing on 42 interviews with workers and court documents obtained through a Freedom of Information Act request, the author examines Amazon’s antiunion campaign in Bessemer, Alabama. The findings reveal that employers can weaponize elements or effects of algorithmic management against unions via repurposing devices that algorithmically control workers, engaging in “algorithmic slack-cutting,” and exploiting patterns of social media activity encouraged by algorithmic management. These findings demonstrate that the labor process can shape counter-organizing opportunities for employers, not just organizing opportunities for workers. They also reveal that algorithmic management has upgraded the antiunion arsenal, while shedding light on other aspects of algorithmic management that have escaped notice. The discussion section presents a framework for researching how other workplace variables shape counter-organizing. The conclusion discusses implications for our understanding of workplace regimes and the future of labor control.
Much of the world has entered an age of “surveillance capitalism” (Zuboff 2019), in which a “society of algorithms” (Burrell and Fourcade 2021) increasingly shapes social, political and economic life. Nowhere is this trend more pronounced than in a growing number of virtual and physical workplaces. These are the laboratories in which employers can turn de facto “captive populations” (Zuboff 2019:157) of workers into, as Amazon’s algorithmically powered warehouses have been described, “unwilling subjects in a sophisticated encompassing experiment in digital surveillance” that is having “disastrous impacts on their bodies and lives” (Delfanti, Radovac, and Walker 2021:4)
Following in the tradition of labor process theory (Braverman 1998; Burawoy 1985; Edwards 1979; Thompson and Smith 2010; Wood 2021b), scholars have tracked the solidification of such efforts into a control technique that involves the use of algorithms and supporting surveillance and communication technologies to automate managerial tasks and control workers (Kellogg, Valentine, and Christin 2020; Lee et al. 2015; Wood 2021a).
Recent research has focused on how employers use this technique, known as “algorithmic management,” to direct, evaluate and discipline workers on a day-to-day basis (Kellogg et al. 2020; Lee et al. 2015; Vallas, Johnston, and Mommadova 2022; Wood 2021a). It has also looked at how workers experience and resist algorithmic management (Kellogg et al. 2020; Lei 2021). But what happens after workers take collective action against employers that use algorithmic management? Can algorithmic management also affect the ability of employers to repel organizing by workers, not just control their routine activity?
The answers are important because algorithmic management is the ascendant control technique and the “new contested terrain” of intensifying labor struggles (Edwards 1979; Kellogg et al. 2020; Schaupp 2023), as well as a hothouse for the development of new surveillance and algorithmic manipulation techniques more broadly (Zuboff 2019). Forms of algorithmic management are rapidly expanding far beyond their birthplace in the gig economy, reaching into logistics, retail, manufacturing, hotels, customer assistance, banking, law enforcement and many other industries (Wood 2021a:1). The outcome of the escalating struggles over algorithmic management therefore could shape the future of work, capitalism and political systems by helping revive or further erode one of the most equalizing forces in history: the labor movement (Burawoy 1985; Edwards 1979; Reuschemeyer, Stephens, and Stephens 1992; Rosenfeld 2014; Thelen 2019:16).
This article reveals how algorithmic management can shape employer resistance to collective action by examining a key early battle in the new labor upsurge: the first large union election at an Amazon warehouse, which took place in Bessemer, Alabama, in early 2021 (Brook 2021). Drawing primarily on 42 in-depth interviews with Amazon workers and analysis of court records obtained through a Freedom of Information Act (FOIA) request, I do so by highlighting how Amazon exploited constitutive elements or effects of algorithmic management to resist the union drive. Specifically, I show that Amazon (1) weaponized workplace devices that algorithmically direct and discipline workers, (2) engaged in what I call “algorithmic slack-cutting” to partially restore consent, (3) leveraged an app that automates human resources (HR) management to disseminate antiunion messages, and (4) exploited distinctive patterns of social media activity encouraged by algorithmic management. Taken together, these findings suggest that algorithmic management generally enhances the antiunion arsenal of employers by guaranteeing the presence of additional capacity to intimidate workers, shape their preferences and rapidly restore consent.
These findings contribute to literature on labor studies, labor process theory, algorithmic management and the sociology of work. The empirical contribution of this research is to show that algorithmic management strengthens the antiunion arsenal—the weapons used by employers to repel a union drive after one has begun—and expands the power of employers to influence their workers more generally, including through penetrating and flexible communication capacities and through “algorithmic slack-cutting.” The main theoretical contribution of this research is to provide a framework through which scholars might investigate how other control techniques or dimensions of firm architecture (Lei 2021) can shape counter-organizing opportunities for employers, not just organizing opportunities for workers. In doing so, I demonstrate the value of bridging labor process theory with the employer counter-organizing literature. The findings also have implications regarding analysis of workplace regimes, “flexible despotism” as a macro regime and future uses of algorithmic management, which could include influencing the political and social behavior of workers beyond the workplace, not just their technical and political behavior within the workplace. Overall, these contributions recommend further inquiry into how employers scan weaponize elements or effects of the labor process to undermine union organizing and shape economic, social and political behavior more generally.
Labor Process Theory and Antiunion Arsenals: Missing Links
Labor process theory has identified different workplace regimes and control techniques that use varying degrees of coercion and consent to control workers, secured through the labor process: the organization of tasks, technology and rules in the (virtual or physical) workplace (Braverman 1998; Burawoy 1985; Edwards 1979; Wood 2021b). Although early theorists emphasized coercive control mechanisms including deskilling and surveillance (Braverman 1998), later thinkers noted the growing need for employers to obtain consent (Vallas et al. 2022). Burawoy introduced the concept of workplace regimes. He posited that a hegemonic regime established control primarily by providing workers with rights and limited autonomy, in large part through unions, and that this regime had replaced “market despotism” as the predominant workplace regime after World War II (Burawoy 1982). Edwards (1979), meanwhile, identified “simple control,” “technical control,” and “bureaucratic control” as the progression of the most salient control techniques, if not holistic regimes, during the same period. Other theorized control techniques include “normative” control and “neonormative control,” which rely on shaping and/or celebrating the values and identities of workers (Barley and Kunda 1992; Fleming and Sturdy 2009).
With the turn to “flexible production” in recent decades (Chun 2001; Moody 2017: 3, 13), evidence has mounted that many employers blend a range of coercive and consent-oriented techniques, even as many scholars have detected an overall shift toward coercion (Thompson and van den Broek 2010; Wood 2020). This multiplicity of control techniques at employers, as well as recognition of varying structural conditions affecting their presence or articulation, has led some scholars to question the value of totalizing workplace regime archetypes (Vallas et al. 2022:6–7).
Wood (2021b) has countered that regime archetypes remain useful for capturing broad historical tendencies in the use of coercion and consent by employers. Among defenders of the concept of archetypal workplace regimes, “flexible despotism” has received perhaps the most acceptance as a modern macro regime. Scholars have shown how different configurations of control techniques—most of which are coercive but some of which also rely on consent—all can combine to produce flexible despotism (Chun 2001; Wood 2020). Although scholars initially emphasized the role of numerical flexibility in flexible despotism, such as through the use of temporary workers and instant layoffs (Chun 2001), more recent scholarship has highlighted the role that scheduling flexibility as well as digital direction, monitoring and discipline can play in underpinning flexible despotism (Ikeler 2016, 2019; Wood 2020), sometimes culminating in forms of “algorithmic management” (Nguyen 2021; Noponen et al. 2023; Wood 2020).
Regardless of where one falls in the debate over workplace regimes, the discussion has highlighted the value of remaining closely attuned to the variety of architectural dimensions and control techniques, involving both coercion and consent, that can shape labor control in any given workplace. To that end, Lei (2021) recently proffered organization, legality and technology as the primary architectural dimensions of platforms, while Wood offers a broader conceptual menu for understanding control, including elements in the workplace (e.g., technological and cultural) and their embeddedness in external networks (e.g., markets) and institutions (e.g., social programs and race) (Wood 2021b).
Despite calls for attention to the multidimensionality of workplace regimes, labor process theory tends to overlook perhaps the most coercive component of labor control, what I call the “antiunion arsenal.” I conceive of the antiunion arsenal as the collection of the weapons and tactics used by employers to repel a union drive after one has begun. I follow antiunion consultants, organizers, and some scholars in using the term arsenal to convey the coercive nature of employer resistance, rather than obscure this reality through the use of a less descriptive term such as repertoire (DeMaria 2019; Fantasia 1989:36; Logan 2002:213).
Labor process theory research sometimes examines how the labor process can prevent union drives (e.g., Ikeler 2019). It also sometimes explores how the labor process can shape resistance and organizing opportunities (Edwards 1979; Hodson 1995; Moody 2017; Rojas 2017). But analysis of weapons used against workers after a union drive has begun—the antiunion arsenal—has been studied in isolation from the labor process (Burawoy 2008). The vast majority of the literature on labor struggles, moreover, centers worker and union organizing, not employer counter-organizing (Brookes 2018:255; Lepie 2016:147–48). This literature frequently spotlights innovations in organizing tactics (e.g., De Lara, Reese, and Struna 2016; Marvin 2014). Yet the more limited research on the antiunion arsenal has found that its weapons have remained strikingly formulaic over the past five decades (Bronfenbrenner 2022:17; Lafer 2007; Logan 2002; Warner 2013).
The generic antiunion arsenal contains both carrots and sticks. The “carrots” include improving working conditions and rewarding antiunion workers (Brodkin and Strathmann 2004:17). The “sticks” typically include isolating, threatening or firing workers, with penalties for this illegal activity regarded as a negligible cost of doing business (Lafer 2021:10–11). Employers use surveillance and various communication channels to deploy these carrots and sticks, leveraging what one antiunion consultant has called their “complete access to the minds of voters during working hours” (Lafer 2005:12). These communication channels include one-on-one meetings, campaign materials, and mandatory antiunion meetings known as “captive audience meetings” (Brodkin and Strathmann 2004; Lafer 2007; Logan 2002; Warner 2013).
This research has done little to explore potential diversity in the type, caliber and use of these weapons across management regimes. Brookes (2018) duly called for shifting more attention to employer resources and strategy in order to sharpen the “analytical and practical utility” of work on labor struggles (p. 255). Indeed, there is ample reason to believe that the control techniques and architectural dimensions of a firm might shape its antiunion arsenal. We know generic antiunion arsenals have shifted in the past (Fantasia 1989:25–72; Logan 2006). We know technologies “enable and constrain the range of options available to employers and thus shape the balance of coercion and legitimation utilized by a firm” (Wood 2021b:125). We know employers have weaponized workplace culture and relationships during some union elections (Penney 2006). And we know the labor process shapes the organizing opportunities available to workers (Edwards 1979; Rojas 2017).
We thus have good reason to believe that the labor process also shapes counter-organizing opportunities available to employers. One way to explore this proposition is to select a particular control technique or dimension of firm architecture and then probe whether the technique or dimension’s constitutive elements or effects appear to shape the antiunion arsenal. In this article, I take such an approach to assess whether the control technique of algorithmic management can shape the antiunion arsenal.
Algorithmic Management
Scholars have defined algorithmic management in various ways (Nguyen 2021:1; Noponen et al. 2023; Schaupp 2023:2). I adopt the “first and perhaps most used” definition (Noponen et al. 2023) of algorithmic management, which Lee et al. (2015) described as “software algorithms that assume managerial functions and surrounding institutional devices that support algorithms in practice” (Noponen et al. 2023). I also understand algorithmic management as a discrete control technique, not a totalizing regime or comprehensive system of control. From this perspective, any employer that uses algorithms and supporting devices to partially or fully automate a managerial role, such as HR management, is using a form of algorithmic management.
Algorithmic management can “greatly enhance employers’ capacities to direct, evaluate and discipline employees” through the use of computer code and devices including barcode scanners, motion sensors and smartphones to (Kellogg et al. 2020; Lee et al. 2015; Nguyen 2021:15). But, notably, even as algorithmic management automates forms of management, applications of the technique to date have always left room for some human intervention (Wood 2021a:3–4). In Wood’s typology, the most advanced forms of algorithmic management only offer “conditional automation,” not “full” or even “high” managerial automation (Wood 2021a:12). And even as algorithmic management sometimes helps employers tighten control over workers, scholars have also noted that it can sometimes simultaneously offer a sense of autonomy, such as by allowing them to choose when and where to work. This contradiction has been labeled a “paradox of autonomy” (Möhlmann and Zalmanson 2017; Wood 2021a:13). Examining algorithmic management within the context of a union drive may offer further insight into how and why the technique heightens and automates control even while preserving room for intervention and purportedly sometimes providing workers with autonomy.
Research on algorithmic management has tended to focus on how it controls and effects workers on a day-to-day basis (Duggan et al. 2020; Nguyen 2021; Tassinari and Maccarrone 2020:40; Veen, Barratt, and Goods 2019). Moreover, most focuses on gig platforms, such as Uber, TaskRabbit, and Instacart (Griesbach et al. 2019; Lee et al. 2015; Rosenblat 2018). This has left a gap in research on how firms in other sectors are deploying algorithmic management (Schaupp 2023:2). A subset of the literature on algorithmic management has specifically looked at how algorithmic management can both help prevent unionization and spark forms of resistance (Griesbach et al. 2019; Kellogg 2020; Lei 2021; Rogers 2023; Rosenblat 2018; Schaupp 2022). Research on major retailers, particularly Walmart, has noted that flexible scheduling, which can be algorithmically assisted or automated, has offered employers a more subtle and potent capacity to resist worker organizing than overt retaliation (Bank Muñoz 2017; Reich and Bearman 2018; Wood 2020:14, 97–98). And Rogers (2023) recently surveyed a range of ways in which employers could exploit forms of algorithmic management to thwart collective action. But no case studies have empirically investigated whether algorithmic management affects how employers respond during a formal union election, when we would expect employers to fully exploit all counter-organizing opportunities at their disposal.
This research helps fill in these gaps through analysis of how a pioneer of algorithmic management beyond the gig economy resisted a union drive. In the following section I summarize certain constitutive elements and effects of algorithmic management that might be expected to generate counter-organizing opportunities for employers. By constitutive elements, I refer primarily to the “software algorithms that assume managerial functions” and/or “surrounding institutional devices that support algorithms in practice” (e.g., workstation displays) referenced in Lee et al.’s (2015) definition. By effects, I refer to effects on workers resulting from the application of algorithmic management.
Constitutive Elements and Effects of Algorithmic Management
Communication and Surveillance Devices
Past research has noted the key role that digital communication and surveillance devices play in algorithmic management. This research has shown how employers can use these devices to algorithmically track, direct, evaluate and discipline workers (e.g., Griesbach et al. 2019). Ride-hailing firms uses apps to direct drivers to destinations and collect customer ratings that can trigger automatic termination (Cameron 2022; Lee et al. 2015; Rosenblat and Stark 2016), while hotels increasingly track the work speed of housekeepers and update their assignments through tablets they are required to use (Mateescu and Nguyen 2019:9). Less examined is the capacity for employers to leverage these devices to communicate not just work instructions, but to try to shape the opinions of their workers. These devices include mobile apps, workstation displays and scanners. Digital communication and surveillance devices are constitutive elements of algorithmic management that employers conceivably could leverage for antiunion purposes.
“Algorithmic Imaginaries.”
Scholars have also examined how algorithmic systems can generate particular beliefs about how algorithms work, and that these beliefs can shape behavior (Bucher 2017; Sharone 2017). This research mobilizes Bucher’s (2017) concept of the “algorithmic imaginary,” defined as “ways of thinking about what algorithms are, what they should be and how they function” (p. 30). Algorithmic imaginaries are effects of algorithmic management and likely to be salient and potentially manipulable during union elections.
The “Electronic Whip,” Automated HR, and Dehumanization
Research on algorithmic management has also documented the tendency of algorithmic management to drive work at a frantic pace through quotas, tracking and automated discipline (Delfanti 2019; Wood 2021a:8–11). McCallum (2020) described this algorithmic drive system as the “electronic whip” (pp. 85–107). Another form of algorithmic management involves the automation of HR-related tasks. This approach tends to rely on using employee mobile apps that provide information like pay records, use chat bots to answer HR questions, and administer (sometimes automated) scheduling (Duggan et al. 2020). Both the electronic whip and HR automation produce distinct dehumanizing effects and corresponding grievances (Lei 2021). The electronic whip causes anxiety and exhaustion through opaque surveillance, high work intensity, and the threat of automatic termination (Delfanti 2019; Mateescu and Nguyen 2019:14). And HR management automation can cause confusion and frustration by removing the “more interpersonal and empathetic aspects of people management” (Duggan et al. 2020:128). The electronic whip and automated HR management are common constitutive elements of algorithmic management and the distinctive resulting dehumanization of workers is a distinctive effect of the technique. We might expect to employers to manipulate these elements and effects of algorithmic management in a counter-organizing campaign.
The Effect of Algorithmic Management on Social Media Activity
Scholars have also noted how the grievances and confusion arising from algorithmic management have driven workers to flock to social media to vent and try to “decipher and keep up with the opaque processes and policies” of the workplace regime (Mateescu and Nguyen 2019). Given the isolating nature of the work, this has turned these social media communities into key potential sites for workers subject to algorithmic management to plan collective action (Kellogg 2020:392; Maffie 2020:125; Tassinari and Maccarrone 2020). But this effect of algorithmic management would also seem to create surveillance and influence opportunities for employers. The point here is not that only employers that use algorithmic management can exploit social media to resist unions. It is that such employers tend to have more potential to do so because of the particular social media behavior that algorithmic management encourages.
Amazon Warehouses: “The Perfect Candidate” to Understand the Future of Work
Amazon has rightly been dubbed the “perfect candidate to understand changes in the relation between labor and machinery” (Delfanti 2019). It is the second largest employer in the United States (Vallas et al. 2022:7) and employs more than 1.5 million workers across the world. And it is a quintessential pioneer of “one of the most intrusive and pervasive systems of workplace surveillance the world has ever known,” one that is shaping the future of work (Delfanti et al. 2021:3–4). Amazon’s size and use of employees in physical workplaces—not just independent contractors without physical workplaces (the preferred workforce of gig platforms)—means the firm’s labor practices may offer more insight into algorithmic management and the future of work than research on gig platforms
Scholarship on Amazon has mostly examined the employer’s day-to-day managerial strategies and the roadblocks they pose to worker organizing, often based on analysis of warehouses outside the United States (Delfanti et al. 2021; Massimo 2020; Rogers 2023; Vallas et al. 2022; Vgontzas 2020). Recent research has underscored the employer’s use of a multiplicity of control techniques, shaped in part by the demographic composition of its workers. Vallas et al. (2022) found that Amazon used not just “techno-despotism” (algorithmic management applied despotically, as it usually is), but also consent-oriented techniques. The latter included fostering a sense of indebtedness (“relational control”) and frequently offering workers voluntary unpaid personal time (UPT) (part of “governmental control”), a novel way to achieve temporal flexibility and simultaneously provide workers with a sense of choice. Lee et al. (2024), meanwhile, uncovered racialized control techniques, including “plantation-style management” and “public and private policing,” by examining the same warehouse in Bessemer, Alabama, as this study, though without a specific focus on the union election there.
The first opportunity to see how Amazon would respond to a full-fledged union drive in the United States arose in late 2020 (Brook 2021) when workers at the Bessemer facility triggered the first large union election at an Amazon warehouse in the United States. During the election workers, voted against affiliating with the Retail, Wholesale and Department Store Union (RWDSU). Only one of at least six organizing campaigns that have petitioned for union elections have emerged victorious since then, and Amazon has refused to recognize the one union that won. An independent union seeking to represent workers at a warehouse in North Carolina recently successfully petitioned for the first election at an Amazon warehouse in more than two years. The election is scheduled to take place from February 10 to 15, 2025. 1
Some of the control techniques uncovered by Vallas et al. (2022), Lee et al. (2024), and others no doubt conditioned Amazon’s response to the Bessemer union drive (some evidence of which is briefly mentioned in the discussion). But given Amazon’s exceptionally advanced application of algorithmic management in a physical workplace (Wood 2021a:13), the case offered perhaps an unprecedented opportunity to investigate how employers might leverage the specific control technique of algorithmic management to repel (not just prevent) collective action by worker. My study takes advantage of that opportunity. But I also specify a framework that future researchers might use to examine the impact of other control techniques on employer counter-organizing, including other techniques other scholars have found operating at Amazon.
Methods
This article is based on an analysis of dozens of interviews with Amazon workers and review of transcripts of hearings held by the National Labor Relations Board (NLRB) on the union election. I conducted semistructured interviews with 42 workers who said they worked at the warehouse in Bessemer, Alabama, where the union election took place from early February to late March 2021. Thirty-four workers said they were eligible to vote in the union election, with 18 identifying as prounion and 16 identifying as antiunion during the election.
The final vote tally of the election was 1,798 to 738 votes against RWDSU, out of about 5,867 workers eligible to vote (NLRB 2021c; Thorbecke 2021). The decision-making trajectories of the voters I interviewed suggest that Amazon was overwhelmingly successful at persuading (or intimidating) originally prounion or neutral workers to vote no: 14 of the 16 antiunion workers started as prounion or undecided but subsequently turned against the union. Notably, within six months of the election, four of these yes-to-no voters said they regretted their decision and wish they had voted yes (they said they were deceived or intimidated by Amazon into voting no). None of the 18 prounion workers switched from undecided or antiunion to prounion.
Workers employed during the union election were asked questions primarily about their working conditions, Amazon’s response to the union drive, and how they decided to vote yes, no or not at all. Although the eight worker interviewees who said they were ineligible to vote in the election could not speak to Amazon’s counter-organizing tactics, they provided information on working conditions and surveillance practices that was helpful for further contextualizing Amazon’s response to the union drive.
Recruitment for interviews took place overwhelmingly at convenience stores near the warehouse during a three-week field visit during July 2021. Amazon workers would visit the stores before shift, during lunch breaks or after shifts. I would ask them if they worked at Amazon outside the store entrance, present my university ID to prove my identity, offer anonymity and collect their contact information. Then we would schedule an interview and conduct the interview by phone when they were off work. A handful of interviews were also conducted on site. I also recruited a minority of worker interviewees through other means. Several came through referrals from other interviewees. Several more came from striking up conversation with locals in various settings across Birmingham and Bessemer and asking if they knew someone who work at Amazon. For example, an employee at an airport rental car station whom I chatted with turned out to work at Amazon and agreed to an interview. In another case, a bar patron I explained my project to revealed he had been a former Amazon worker and referred me to a friend who was employed there during the election and agreed to an interview. A few interviewees were also sourced from social media groups and social media keyword searches for posts related to the campaign after my visit. Dozens of conversations with Amazon workers and community members outside convenience stores and across Bessemer and Birmingham, as well as with labor experts and a former manager of another Amazon warehouse, also provided context and insights. My experience as a labor reporter, recruitment advice from labor reporters who had covered the campaign (whom I contacted before my field visit) and tips from interviewees led me to select my recruitment tactics and to target specific sites. No interviewees were referrals by union staff or Amazon. Interviews typically lasted between one and three hours.
I obtained the transcripts of the election hearings, which were held to determine if Amazon violated labor law, 2 by filing a FOIA request through a request for “transcripts for the post-election hearings” in January 2022, six months after my field visit (NLRB 2021a). In addition to the transcripts, the FOIA deliverable included NLRB filings related to the setup of the election and the unfair labor practice charges filed by RWDSU and Amazon that were the basis of the hearings. Also included in the FOIA deliverable were the court exhibits, such as photos of campaign materials and messages, as well as company guides for Amazon managers and workers, that RWDSU used to argue their case (NLRB 2021b). The full FOIA deliverable was more than 1,800 pages.
I also examined hundreds of pages of other documents relating to the election and unfair labor practice charges that were filed with the NLRB (2021c, 2021d, 2021e), including decisions by NLRB judges. These documents can be obtained through a case number search of 10-RC-269250 on the NLRB’s Web site. The NLRB documents, especially the court transcripts and exhibits obtained through the FOIA request, confirmed and supplemented reports from interviewees, including by providing testimony from Amazon workers, managers and antiunion consultants on the firm’s employee mobile app, captive audience meetings, and antiunion messages, as well as company guides and documents that further elucidated them. They also include some images used in this article. In addition, review of several articles on the basis of media leaks informed this article’s analysis of how Amazon weaponized social media against the union. After my field visit, I analyzed the interviews, NLRB documents, 3 and news articles using abduction to produce new propositions on the basis of surprising research evidence (Tavory and Timmermans 2014, 5). All interviews and NLRB documents were extensively coded through an iterative process. Initial coding focused on themes from literature on antiunion tactics and union election decision making by workers. When the unusual role of technology during Amazon’s antiunion campaign emerged as a theme, engagement with the labor process theory literature suggested connections between these tactics and the workplace regime of algorithmic management. Following Charmaz (2006), focused coding was then used to highlight and merge relevant codes into higher level codes. These higher level codes, such as automated human-resources management, mobile apps and the social media behavior of algorithmically employed workers, were deductively drawn from the extent literature on algorithmic management. Through this exercise, I was able to clarify links between antiunion tactics and some constitutive elements and effects of algorithmic management.
Setting: Warehouse Location and Working Conditions
Bessemer is a former steel town that has suffered some of the worst effects of deindustrialization. The city is a lower-income and predominantly Black community that reportedly had the highest violent crime rate in the country in 2019, the year before the Amazon warehouse opened (Schiller 2019). It provides the ideal conditions that logistics firms look for in locating their distribution facilities: economic malaise; racial enclosure; shrinking manufacturing and public sector employment; and eroded social spending (The Economist 2022; Moody 2017:61)
Most Amazon workers interviewed for this study stowed, picked or packed packages in isolated workstations through interactions with robots and conveyor belts that would transport items between stations. They reported being normally subject to unforgiving productivity standards, though these standards were largely suspended during the campaign. Workers reported sometimes receiving automatic write-ups or even termination for failing to meet productivity standards. They often felt stressed and frustrated by the working conditions for a range of reasons, including constant surveillance, pressure to work faster, confusion about grounds for discipline and trouble with resolving work-related problems. These conditions mirror those documented in other literature on Amazon (e.g., Struna and Reese 2020; Vallas et al. 2022).
Amazon used many common antiunion techniques during the campaign. These included one-on-one conversations with supervisors and consultants, captive audience meetings and antiunion media. Amazon also issued typical warnings to sow fear and doubt about unionization through these communication channels, including claims that unionization might reduce their pay and benefits and make it more difficult to resolve problems (Brodkin and Strathmann 2004; Bronfenbrenner 2009; Lafer 2007). But the devices and software needed to power Amazon’s algorithmic management, and their effects on workers, also offered Amazon opportunities to amplify or supplement these traditional tactics.
Findings
In this section I show how different constitutive elements or effects of Amazon’s algorithmic management shaped and amplified the employer’s capacity to intimidate workers, influence their beliefs and rapidly restore consent. The section is organized into four subsections. Each covers a different general tactic, beginning with an explanation of the relevant constitutive elements and effects of algorithmic management and then proceeding to explain how Amazon weaponized those elements and/or effects. In the first subsection I examine how the employer weaponized surveillance and communication devices that algorithmically monitor and direct employees to intimidate and persuade them. In the second subsection I show how the “electronic whip” and automated HR management, two elements of Amazon’s algorithmic management, and their dehumanizing effects on workers empowered the employer to engage in “algorithmic slack-cutting” to restore consent. In the third subsection I look at how Amazon leveraged a mobile app that automates HR management to distribute antiunion messages. In the fourth subsection I explore how Amazon exploited distinctive patterns of social media activity encouraged by algorithmic management to surveil, intimidate and propagandize workers.
Weaponizing Workplace Surveillance and Communication Devices
Algorithmically Haunted Captive Audience Meetings
Research has detailed how employers use captive audience meetings to sow fear, doubt and misconceptions about unionization (Brodkin and Strathmann 2004; Masson 2004). In this section I show how devices needed to support algorithmic management in the workplace can evidently empower employers to magnify their intimidating character. Amazon accomplished this through the theatrical deployment of scanners and productivity-monitoring computers, both of which workers are conditioned to associate with discipline.
Algorithmic management can often involve the use of digital devices to monitor and discipline workers. Ride-hailing apps, for example, use one mobile app to direct drivers to passengers and to destinations and another app to collect customer ratings that can trigger automation termination of drivers if they fall below a certain level (Cameron 2022; Lee et al. 2015; Rosenblat and Stark 2016).
In Amazon’s case, scanners and computers are constitutive elements of its algorithmic tracking and disciplinary apparatus (Delfanti 2019; Struna and Reese 2020; Vallas et al. 2022). At the Bessemer warehouse, Amazon workers used scanners to register items that they were directed to pick, sort or pack through “algorithmic instructions,” as part of Amazon’s use of “chaotic storage” (Wood 2021:3). Supervisors also used scanners to audit and discipline workers. In addition, supervisors visibly used computers to monitor the performances of workers based on indicators generated by the scanners (NLRB 2021a:1150). In all, interviewees confirmed Vallas et al.’s (2022) description of scanners as enabling management to generate real-time data on the productivity of workers, to publicly rank them, and to algorithmically monitor their compliance with quotas, thereby imposing “the constant threat of discipline and termination on workers who fail to make rate” (p. 23). Scanners and computers consequently occupied a central place in the “algorithmic imaginaries” of workers. Both were associated with the threat of algorithmically assisted discipline. James, an early-20s yes-to-no voter who says Amazon manipulated him into voting no, recalls facing this threat as soon as he started working there, after returning one minute late from his lunch break. “One of the managers was like, ‘Let me scan your badge,’ and I was like, ‘Hey man, I’m new . . . it’s my first day, I could have had a verbal warning not necessarily having my badge scanned.’” Roger, an early-20s undecided-to-no voter who sometimes audits workers, pointed to awareness among workers of the disciplinary function of laptops and tablets used by supervisors to monitor workers’ productivity: “[Managers] have a laptop there that can see a floor’s whole performance being like, ‘OK, this person is going a little slow. . . . We’ll check on that and make sure they understand how many items we want per tote.’”
Amazon supervisors and human resource officers weaponized these associations at captive audience meetings, in which consultants and supervisors warned workers against unionizing (see Figure 1 for a presentation slide shown during one such meeting). They did so by conspicuously scanning meeting attendees and watching workers with their laptops open as the meetings unfolded (e.g., NLRB 2021a:492, 528, 1150). Matt, a mid-20s union activist, described how Amazon supervisors leveraged the devices to magnify the intimidating character of captive audience meetings and to publicly single out workers whose questions or comments at the meetings appeared to indicate support of the union.
A slide that was shown at a captive audience meeting during the Bessemer campaign. Amazon theatrically wielded scanners and computers—constitutive elements of Amazon’s algorithmic management that workers associated with discipline—to increase their intimidating nature. This slide emphasizes that workers could lose their benefits as a result of unionization, in part, because Amazon has “no obligation” to “contract to continue all existing benefits.”
They would scan your badge to let you know that they’re keeping track of you, and that you’re at the meeting. They would have HR reps sit alongside the walls in meetings. At certain points if they thought you were asking too many questions, they would say, “we can talk about that after the meeting,” and they would say, “you stay behind.” If you stayed behind, they would scan your badge. Workers were in there watching you [get your badge scanned].
As this comment and others revealed, surveillance devices that are constitutive elements of Amazon’s algorithmic management and the paranoid algorithmic imaginary they give rise to offered the employer an opportunity to increase the potency of a traditional antiunion tactic, specifically, captive audience meetings.
Digital Polling and Communication
With algorithmic management, the same devices that employers use to direct, monitor and discipline workers can be used to not just communicate job instructions to workers, such as to pick up passengers or make deliveries, but also to influence their beliefs (Hawkins 2020). In the Bessemer warehouse, workstation displays served this purpose. The displays instructed workers on how to do their jobs, such as what items to pick, stow or pack, in interaction with robots, a fusion of human and machine that is core to Amazon’s algorithmic “chaotic storage” (Delfanti 2019:45; Wood 2019:3). But they also beamed messages and questions to workers (see Figures 2 and 3). Thus, a device that is key to allowing Amazon to use algorithmic “chaotic storage” empowered Amazon to layer communication unrelated to that system into the labor process.
Amazon sends questions to workers through their workstation displays. During the union drive, some workers believed that these questions were adjusted and/or used to gauge union sympathies or purge disliked managers en masse. This is from a display in the Bessemer warehouse a few months after the election.
Another question sent to a workstation display in the Bessemer warehouse a few months after the election. During the union drive, Amazon also sent messages through the displays that warned workers against voting for the union, not just questions.
Amazon leveraged this capacity by sending messages to workers to campaign directly against the union. At different points, displays at workstations encouraged workers to “vote ASAP and vote No” (NLRB 2021b, Employer Exhibit 104), to “not let anyone fill out your ballot for you” (part of string of claims implying that the union was ballot harvesting), and to use the “easy, safe and convenient” mailbox that it had installed outside the entrance” (NLRB 2021b, Employer Exhibit 104). Amazon also used the digital displays to frame its workplace as desirable in new ways after the union drive began. The company publicized its career training program, the purportedly generous compensation that workers already received and the number of workers that Amazon had recently promoted, a messaging tactic that exploited the “relational control” of Amazon (Vallas et al. 2022) that fosters a sense of indebtedness among vulnerable workers for receiving any employment at all (NLRB 2021b, Employer Exhibit 104).
The workstation displays also enhanced the ability of Amazon to strategically identify and intimidate (or court) workers. During the campaign, the employer beamed questions that workers were prompted to answer in order to begin or continue work that included how satisfied they were with their jobs, whether they felt recognized “for work well done,” and whether they were satisfied with their career development (NLRB 2021b, Employer Exhibit 104; NLRB 2021e:12). Illiad was among workers who believe that Amazon leveraged such polling specifically during its antiunion campaign, calling it a “way of finding out the folks who are more likely to question things.” One worker suggested before the NLRB that it did so in part by manipulating the questions during the campaign. Survey questions beamed through their displays “changed dramatically” during the campaign, they said. “That’s when the questions of, ‘Is your supervisor doing his job?’” began to appear (NLRB 2021a:335). The concern that Amazon may use polling results to punish workers caused many to chronically lie and provide the most positive answers. This offered Amazon another antiunion capacity: the ability for the employer to churn out and publicize inflated approval ratings for antiunion public relations purposes. Amazon’s claim that 94 percent of Amazon workers would recommend Amazon to a friend as a place to work, for example, is reportedly based on such polling (Del Rey 2021).
Many workers believed that Amazon used workstation polls to gauge their union sympathies and to neutralize union supporters. Many also said Amazon purged disliked managers en masse during the campaign, a task that the digital polls would have made easier, given that they can gauge the popularity of managers in real time. I did not uncover any direct evidence to support these contentions. But what matters most is that workers thought Amazon used workstation polls to engage in these practices. Combined with the theatrical use of badge scanning, this perception increased the salience of an already paranoid “algorithmic imaginary” that, as one worker put it, held that “everything is monitored and Amazon kind of knows everything, hears everything, sees everything.” Several workers said they voted no or abstained from voting because they believed that Amazon could find out if they supported or voted for the union and then fire them. It is highly plausible that the aforementioned ways in which workplace devices powering Amazon’s algorithmic management were weaponized or perceived during the antiunion campaign contributed to this fear.
In this section I have shown that surveillance devices—scanners and computers—that collect data used in algorithmic management allowed Amazon to amplify the intimidating character of captive audience meetings. Likewise, communication devices baked into the algorithmic labor process—workstation displays—offered the firm an additional communication channel that the firm used to try to turn workers against the union, to ask questions that workers thought were designed to discern their union attitudes, and to produce approval ratings for antiunion purposes. These techniques likely produced a more intense chilling effect on prounion discussion (along with a potential public relations bump) than if Amazon had exclusively employed antiunion tactics in their traditional form.
“Algorithmic Slack-Cutting”: Rapidly Restoring Consent
Research shows that employers have long improved working conditions during union election campaigns to try to persuade workers to vote no, such as by raising wages and firing unpopular managers (Lawler 1990:145–46). But in this section I demonstrate how algorithmic management empowers employers to provide an even sweeter “carrot” through what I call “algorithmic slack-cutting.” Algorithmic slack-cutting involves softening the “electronic whip” (McCallum 2020:80) of algorithmic management and lavishing workers with human care previously denied to them by automated HR management (Lei 2021:285–86). The technique has the potential to restore more consent than the traditional antiunion carrot because employers that use algorithmic management have more room to improve working conditions, especially more room to humanize them (Lei 2021:285–86; Vallas et al. 2022:15). It also highlights how leaving room for managers to adjust or intervene in algorithmic management, as they did to engage in algorithmic slack-cutting, likely strengthens both the potency and durability of the control technique.
Amazon’s Electronic Whip
Algorithmic management often intensifies work by increasing the pace and monitoring of tasks to reduce gaps in the workflow and sometimes by even automating disciplinary decisions based on “capricious and opaque ratings” (Wood 2021a:10). Lyft and Uber, for instance, automatically deactivate drivers from their platforms for falling below a certain rating or declining too many rides (Cameron 2022:235, 244; Lee et al. 2015; Rosenblat and Stark 2016). The opacity, relentless pace, and harsh discipline of this system generates higher rates of stress, injury, and dissatisfaction (Noponen et al. 2023; Wood 2021a:8–11). Pressure from algorithmic direction, monitoring, and discipline on housekeepers means they have “no opportunity for recovery,” for example (Wood 2021a:10). Following McCallum (2020), I call this algorithmic drive system the “electronic whip” (p. 80).
Amazon’s electronic whip has three primary elements, Time Off Task, UPT, and quotas, all of which are algorithmically tracked and enforced. Amazon’s Time-Off-Task system tracks the number of minutes that workers are not actively working (Vallas et al. 2022) by requiring workers to log in and out through digital devices when they take breaks or go to the bathroom. Workers can be disciplined up to termination for accumulating too much of time off task. Kate, a mid-20s yes-to-no voter who said she voted against the union because she was “misled” by Amazon, explained the dehumanizing effects of this system: “I’ve seen so many people I work with get fired for [time off task] time, and you know they have families and stuff that depend on them working, and they get fired for going to the bathroom.”
A second aspect of Amazon’s electronic whip was the employer’s UPT system. According to interviewees, this system automatically terminates workers after they use up a certain amount of allotted time, which they can deplete through being late or leaving work early. Several cited the reported death of a worker after the campaign ended as emblematic of this system’s inhumanity (Soper and Eidelson 2021). The worker was denied a request to go to the hospital after complaining of chest pains, interviewees said. He then faced a stark choice: he could leave and be automatically fired, as he did not have any UPT to spare, or he could work through the pain. He reportedly chose the second option and died of a heart attack later that day in the bathroom. “They more or less killed the man ’cause he didn’t have any VTO,” said Derry, a late-50s union activist, referring to “voluntary time off” that workers are sometimes offered (Vallas et al. 2022:21).
The third component of Amazon’s “electronic whip” was production quotas. As Vallas et al. (2022:15) found, workers are pressured to comply with such quotas through real-time productivity rates generated by scanners and displayed on monitors, and workers can be automatically flagged for discipline if they underperform. Employees often felt these quotas were extremely taxing. “It’s impossible. It’s impossible,” said Jasmine, a late-20s yes-voter, echoing comments from other workers who felt dehumanized by these quotas. “To keep people from moving up to get more pay, I feel like they use that to get them out.”
Amazon’s Automated HR Management
Another form of algorithmic management is the partial or full automation of HR management. This typically involves the use of a mobile app and “moves work into an inhuman form” by removing the “more interpersonal and empathetic aspects of people management” (Duggan et al. 2020:128; Lei 2021:286).
Amazon used an app called A to Z for this purpose (NLRB 2021a:1534), causing significant frustration for workers. “My biggest thing is like the communication,” Kate said. “There’s none. It’s all through the app. At other places I worked, communication was always so easy. I could go up and tell my manager, ‘How do I do this?’” (Figure 4 shows the app’s chatbot feature.) Workers sometimes described believing they had successfully used Amazon’s app to obtain time off, only to receive termination e-mails. Niya described experiencing this shortly after she thought she had obtained time off following a car crash. “I broke a fucking bone, and how the hell are you going to stand there and fire me, and I can’t even stand?” She later regained her job, but only after much back and forth with HR employees whom she was eventually able to reach.
Amazon’s A to Z mobile app partially automates human resources management. Features include a chatbot that workers are encouraged to use to get answers to questions and make requests. Workers expressed frustration with getting the information they needed and with filing requests, in part, because “It’s all through the app,” as one worker said.
Algorithmic Slack-Cutting
These dehumanizing working conditions distinctly associated with algorithmic management offered Amazon the capacity to improve working conditions faster and to a greater degree than traditional employers. During the campaign, Amazon leveraged this capacity, demonstrating how employers can engage in “algorithmic slack-cutting” to rapidly restore the consent of some workers. The employer softened the “electronic whip” (McCallum 2020:80), and it also provided workers with some of the “interpersonal and empathetic aspects of people management” missing from automated HR management. It accomplished the latter by expanding workers’ assistance to human care and assistance.
Softening the Electronic Whip
Jasmine reflected on how Amazon softened its electronic whip:
The whole time the union was campaigning they didn’t care about our Time Off Task or none of that. They was telling us to go to the bathroom as we please. They just didn’t say nothing like they normally would, but now they ring you up for any little thing just to get you out of there.
Tyrone’s experience shows how this helped curry favor with workers. “Some rates that they wanted us to meet — it was just like we were robots, like only a robot can meet that rate every hour,” he said about conditions when he started working there just as the union drive began. But after the campaign began, on top of installing fans, providing hydration and popsicles, and granting wage increases, “they became very lenient about rates,” he said. A late-20s undecided-to-no voter, Tyrone was made to feel less like a robot and more like a human. He indicated that these improvements are part of why he said he voted against the union.
Illiad found Amazon’s shift away from automated termination of workers for UPT depletion especially notable. UPT “was Amazon’s main thing of terminating people. If you go negative, you’re gone. But during the first election they actually didn’t do that,” he said. He also claims that Amazon temporarily slashed package volume to the warehouse during the campaign.
I was asking around, and they did this solely because of the union election because they’re like, “OK, the workers are tired of having so much volume come down so let’s cut the volume in half by 50 or 75 percent and hopefully they’ll shut up about it.”
Supporting interviewees’ claims about loosened enforcement, the number of appeals filed by workers to contest disciplinary action dropped dramatically just before and during the voting period (NLRB 2021b:Employer Exhibit 106).
Expanding Human Assistance
Amazon also humanized working conditions by providing workers with more assistance and empathy from actual humans. The company set up a second HR desk right after the union campaign began (NLRB 2021a, 1526). “Until shortly after the union drive began the only way to contact HR was to either use the app or schedule a meeting through a manager,” Clyde said. Moreover, Amazon flooded the floor with human resource officers, out-of-town managers, and consultants. They would ask workers if they had any issues and offer to help solve them, according to interviewees and union witnesses (e.g., NLRB 2021a:805). Derry, a union partisan, said he couldn’t help but feel a bit more appreciated by all the Amazon supervisors visiting from other facilities across the United States. “They was bringing in sunshine. ‘Hey, how you doing. Good morning.’ Just really speaking to you and talking to you, and it was a like a different atmosphere with them around than it was with the original management.” The empathic supervisors and consultants that materialized during the campaign promised to report problems of mismanagement, interviewees said.
Some workers said their jobs remained easier after the campaign ended. But more said Amazon reverted to unforgiving (and obscure) rules and enforcement. Joanne, an early-30s yes-to-no voter, was among interviewees who described feeling tricked.
At first, I was going to vote yes. Then when they kept taking us to meetings every week . . . about how bad the union was I . . . I got confused. Eventually, I’d vote for no because I was confused. I didn’t understand about the union or whatever. I thought things were going to be better at Amazon, but it seems like it’s not getting any better. That’s the biggest mistake I made. But this time I have my mind made up. I’m going to vote yes.
Thus, we see how two elements of Amazon’s algorithmic management—the electronic whip and automated HR management—and the distinct dehumanizing effects they had on workers offered the firm greater opportunity to improve working conditions, and consequently, greater capacity to restore consent. Although algorithmic management generally increases workers’ everyday frustrations, the strategic use of algorithmic slack-cutting allows employers to undercut the consequences of this frustration for worker organization during critical periods. Furthermore, the strategy is notably premised on the ability of human management to intervene in or adjust algorithmic management, a point that will be further discussed later.
Weaponizing a Mobile App
Previous research shows employers often deliver antiunion messages to workers in their homes during campaigns, including through antiunion letters, phone calls, and text messages (e.g., Logan 2002). But Amazon’s antiunion campaign illustrates how a common constitutive component of algorithmic management—mobile apps needed to automate managerial tasks—expands the capacity of employers to impress their antiunion values on workers when they are off the clock. Although Bronfenbrenner (2022) recently noted the emergent use of such apps in antiunion campaigns, I shed new light on the potency and effects of this capacity.
Platforms and some conventional employers that have embraced algorithmic management use apps to communicate, monitor and discipline workers, including by directing and tracking the performance of actual tasks. Many more employers use such apps in a more limited form: to automate HR-related duties and functions (Duggan et al. 2019). In the case of retailers, for example, employee apps are used to administer automated scheduling generated by algorithms that process foot traffic data collected by surveillance devices (Levy and Barocas 2018: 7–9; Mateescu and Nguyen 2019:11). Workers must pay close to attention to communications sent through these apps because their jobs depend on heeding their content.
At Amazon, workers use the A to Z app to clock in and out, correct time punches, request time off, file HR reports, try to get answers to questions (including from chatbots, as shown in Figure 4), apply for transfers, check their pay records, and perform other tasks (NLRB 2021a:1529). Amazon can also use the app to send notifications to workers through three communication channels: push notifications, text messages, and e-mails (see Figure 5). This information includes vital items, such as schedule-change alerts and offers of overtime or VTO. Signing up for the notifications is technically optional, but supervisors are provided with detailed instructions for how to pressure workers to opt into them and to visibly record their responses if they decline to (NLRB 2021b, Employer Exhibits 100 and 101; NLRB 2021a:1611). Unlike communications from the union, “you [practically] can’t opt out of these [app] notifications,” said Clyde, a prounion activist. “The building had to shut down one day a few months ago, and we were notified about that through the notification.” As Clyde suggests, workers are effectively captive to communications sent through the app. (Clyde offered an intriguing Marxist proposition related to this point that I mention in the conclusion.)
Amazon’s A to Z app can send push notifications, text messages, and e-mail alerts with important information, such as schedule changes and offers of voluntary time off (VTO). The image above includes push notifications sent to a Bessemer warehouse worker a few months after the campaign ended. During the union drive, Amazon leveraged both A to Z and a separate text-messaging tool to send antiunion messages to workers.
Amazon took advantage of this capacity (in addition to also using a separate text-messaging tool; see Figure 7) to reinforce the warnings that it delivered to workers through captive audience meetings and other communication channels at work (see Figure 6). (Figure 8 also shows a text message a worker received.) One message, for example, warned workers that they would have to pay $500 in dues and could “receive less, more, or the same as what you already have” because of unionizing. “Protect what you have. Vote no. Visit www.doitwithoutdues.com,” the message concluded (NLRB 2021b: Employer Exhibit 59). Workers could not get around receiving messages through the app. “They were constantly sending us anti-union propaganda like in the middle of the night, super early in the morning, while we were working,” said John, a mid-20s father of two who said he is avowedly prounion but still voted no. One union witness at the NLRB hearings described trying to ignore antiunion messages in her A to Z inbox but apparently still having to at least briefly confront them because “you have to clear those out on your messages so then they go away” (NLRB 2021a:356).
An antiunion message sent through A to Z during the campaign that warned workers not to allow union organizers to “trick you into thinking you do not have to return your ballot.”
In addition to using A to Z to send antiunion messages (which Amazon has the capability to do through push notifications, text messages, and e-mails), Amazon also used a mass-texting tool to send such messages. This is a record from the mass-texting tool used by Amazon that shows a message the tool sent out. It warns that the union might trade away workers’ benefits for “dues check-off” because “it makes it easier for them to take your money!” (NLRB 2021b, Employer Exhibit 66).
During the union drive, Amazon sent this text message warning workers to “Protect what you have” and reminding them of their current benefits. Amazon used both A to Z—which can deliver messages by push notification, text message and e-mail—and a separate text-messaging tool to send messages to workers.
Workers appear to have internalized messages they received from the app, which offered extra reinforcement to the messages they received through a mass-texting tool and captive audience meetings. Or at least their reservations about the union matched the content of the A to Z messages. For example, John complained about the volume of Amazon’s antiunion A to Z notifications and text messages, and he recognized that Amazon’s strategy was “overwhelmingly” rooted in “fear.” Yet John still voted against RWDSU. His explanation corresponded with propaganda delivered through the app (as well as through traditional and digital communication channels at work). “I didn’t want to be represented by people who are known for losing wages, for having worse benefits after they came.”
The A to Z messages were especially potent given that workers were leery to speak with union organizers and had few other sources of information beyond what Amazon told them. “A lot of people don’t read the news often,” Clyde said about some of his coworkers. “[And] they had a much stronger incentive to read and process this information [sent through A to Z] than they did any other news source in their life.” Joanne said she avoided speaking to a union organizer out of fear of retaliation, and that she never learned of Joe Biden’s high-profile endorsement of the union or “about the [the union] rallies and stuff” on the news. She did, however, feel “harassed” by the volume of texts and phone calls she said she received from Amazon. Yet claims made in those texts, such as warnings that she might lose her benefits, are part of what she said led her to vote no.
Workers were wary of speaking to union organizers because Amazon directly warned them that organizers might visit their homes and try to seal or falsify their ballots. A to Z played a major role in stoking this fear. Workers received at least four such warnings through the app. Valentine, a yes-to-no voter, said he decided to mail his ballot through the mailbox that Amazon had illegally installed at the warehouse because, “If I got union people coming to my house, and they know which way I’m voting, because they talked to me,” they could steal the ballot out of his mailbox. “Boom, my vote don’t count,” he said. “There was nobody guarding [my] mailbox.”
This might seem paranoid, but one A to Z message counseled workers not to be “tricked” by the organizers and to do exactly what Valentine did. “If your mailbox is not secure, we suggest placing your ballot envelope in a secure USPS mailbox,” such as the one that Amazon had installed. Reflecting on such communications, Valentine began to ponder for the first time whether they played a role in flipping him from a yes to no voter: “Was my vote swayed?”
Employers have long leveraged what one antiunion consultant calls their “virtually complete access to the minds of voters during working hours” to defeat unions (Lafer 2005:12). In this section I have shown how another constitutive element of Amazon’s algorithmic management, a mobile app that automates HR management, offered Amazon the capacity to powerfully extend this access beyond the workplace, and with it, the employer’s capacity to influence the views of its workers toward unions.
Haunting the “New Public Square”: The Weaponization of Social Media
Labor scholars have looked at how social media can serve as a resource for employers to monitor workers (Kellogg 2020:389; Taylor and Dobbins 2021; Wood 2015:264), while also providing a space for workers to cope with the effects of algorithmic management and organize collectively (Lei 2021; Tassinari and Maccarrone 2020). Although Bronfenbrenner (2022) recently noted the use of social media by employers to resist unions in general terms, in this section I provide the first case study–based analysis on the matter and highlight how algorithmic management can amplify its potency. The analysis specifically shows how employers can deter unionization by haunting the “new public square” (Hamilton 2021:117) using three tactics: social media surveillance, paid employee promotion on social media and social media advertising. I find that algorithmic management expands and strengthens the opportunity for employers to use these tactics because it tends to generate distinct patterns of social media activity and to activate algorithmic imaginaries especially corrosive to union discussion.
Social Media as a Coping Mechanism for Algorithmic Management
By automating traditional managerial tasks, algorithmic management encourages workers to flock to social media to try to “decipher and keep up with” the workplace regime’s “opaque processes and policies” (Mateescu and Nguyen 2019). Workers also use social media to engage in broader communications with other workers. Uber drivers, for example, use social media groups “to learn tricks and tips for success on Uber’s platform; compare and share practices and screenshots; complain socially about passengers and the company; and debate Uber’s practices, including discrepancies between the passenger and driver apps” (Rosenblat and Stark 2016). Given the increased difficulty of communication, social media also serves as a key setting where workers controlled through algorithmic management can plan collective action (Kellogg 2020:392; Maffie 2020:125; Tassinari and Maccarrone 2020).
Social Media Groups for Amazon Workers
Amazon workers use social media extensively, as revealed by their countless Facebook groups. In one group with nearly 200,000 members, workers ask questions that often reflect uncertainty and confusion about how to use the A to Z app. Other questions reveal the fear generated by Amazon’s electronic whip. For example, one member asked whether they would be fired for stowing less than 250 units per hour and another asked if they could avoid termination even if they deplete their UPT. Conflicting answers to such questions underline the complexity and unstable character of Amazon’s algorithmic management, and the greater need for a setting in which to hash out nuances. These groups also make clear the role of social media as a setting in which to commiserate over grievances. One member complained about never receiving a promotion even though they had “busted my ass” for 13 years. Chatter within the groups suggests how such griping could foster a collective understanding of their problems at work. “To me reading most of these comments let’s me know that . . . every location seems to have the same problems and everyone’s frustration is REAL.” Although exchanges in these groups gesture toward how social media could seemingly foster union organizing at Amazon, the employer engaged in three tactics that rendered it more of a hindrance than help to the union drive in Bessemer.
“Capture” and Categorize: Social Media Surveillance
The first way in which Amazon leveraged social media to deter union activity was to monitor workers beyond the workplace and to magnify a paranoid “algorithmic imaginary” that they had developed at work. The employer ran a social media surveillance program that monitored more than 43 Facebook groups, most of which were nominally private, as well as numerous Web sites, subreddits, and a Twitter keyword across the world. The program’s described aim was to “capture” and categorize posts of interest for potential investigation, including those mentioning complaints from warehouse workers and planned strikes or protests. Vice News discovered and reported on the program in September 2020, right around when Bessemer workers had begun organizing. The publication reported days before that Amazon was looking to hire two intelligence analysts to track “labor organizing threats” against the company (Gurley and Cox 2020).
Such efforts could obviously help Amazon discover labor activists and counter organizing efforts. But revelations of their existence have also contributed to a fear among some members of Amazon Facebook groups that Amazon may be monitoring their discussions (Palmer 2020). In May 2021, a member of a major Facebook group for Amazon warehouse workers, to which a number of workers at the Bessemer warehouse belonged, posted the Vice article on the surveillance program after another worker expressed skepticism that Amazon could identify them there. The worker said they had made it “abundantly clear to ya’ll for months” that the firm had teams devoted to discovering posts about union activity or “anything critical of Amazon.”
Amazon showed workers at the Bessemer warehouse shortly after the campaign ended that they had good reason to fear discovery and termination for making sensitive comments on social media. One worker at the warehouse said they were terminated for comments they made about the aforementioned man who had died of a heart attack. “He just put the phone in my face, and he’s like is this you?” the worker said, describing a confrontation with a manager. The worker responded, “yeah, that’s me, but how did you get on my Facebook page?” The worker said they were fired without receiving an explanation. Several interviewees mentioned this incident unprompted, viewing it as an attempt by Amazon to make an example of someone.
Paying Employees to “Leave No Lie Unchallenged.”
A second element of Amazon’s social media strategy was an “ambassador program,” launched in 2018. This program involved paying warehouse workers to counter criticism about working conditions at Amazon. According to an internal document leaked to the Intercept, workers were recruited and trained to counter “all posts and comments from customers, influencers (including policymakers), and media questioning the FC [fulfillment center] associate experience . . . leaving no lie unchallenged and showing that people who actually know what it’s like to work in our FCs love their jobs” (Klippenstein 2021). One former manager at another Amazon warehouse told this author their friend would compose tweets for the program in a dedicated room at the warehouse. “It was considered a really cushy thing to do . . . you would turn in your tweets and someone would look over them, and then they would batch schedule them.”
Social Media Advertising
The third element of Amazon’s social media strategy involved using a high volume of social media ads and communications. During the campaign, the company ran videos on Twitch with testimonials from workers employed at the Bessemer warehouse saying why they were voting against the union. Amazon also created a procompany Facebook group that promoted its Black Employee Network affinity group, which according to several interviewees, management tried to convert into a de facto antiunion committee during the campaign. Some workers said they found it difficult to escape a bombardment of digital antiunion ads, including Facebook and Instagram ads that linked to Amazon’s #doitwithoutdues antiunion Web site. “I saw that on dude everything, not only my social media,” Illiad said. “When I was playing my Xbox and go through my store, I saw the ad pop up again, even on my TV. It was wild. It was like they had a monopoly on like everything.”
Given this environment, it is not surprising that many workers encountered almost entirely antiunion views on social media. “I . . . couldn’t find someone to tell me [on social media] why they were for a union,” said Valentine, who cited this as one reason he went from prounion to voting no. Cathy, a supervisor who advised workers to vote no, noticed a similar absence of prounion comments from workers. But, meanwhile, she was happy to “just be talking shit [against the union]” on Facebook herself.
Some interviewees said that many workers did not express support for the union on social media because they were worried their comments would be detected by Amazon. Illiad articulated this belief:
There are folks who believe, and they’re not wrong to believe this, that Amazon is everywhere. They think that somehow management is going to find out, and so they feel like they shouldn’t make a post about it, or like one of their coworkers, especially those who are vehemently anti-union, will rat them out.
One worker said such concerns are why they only used Reddit and Snapchat to make prounion statements, given that the former is anonymous and messages posted on the second disappear. The algorithmic imaginaries of some workers already include a vague suspicion that Amazon might somehow be surveilling workers beyond work. Disclosures that Amazon actually has done so through social media can bring this fear to full life.
Social media discussion about the union drive, inflected by the aforementioned incentives and beliefs, encouraged workers not to vote for the union. Tamika, a late-30s worker who said she abstained as a result of negative comments she saw on social media, described how even positive news stories posted on Facebook could actually backfire for the union:
Every time they post like a [prounion] news thing about the union, and then people put comments saying that the union won’t do nothing; that the union is a fraud. It was more comments of the union—“it won’t help”—than comments that said, “the union will help.” So that’s what had me feeling like maybe I shouldn’t vote.
In line with the tendency of social media to promote “conspiracy, belief and gossip,” Facebook evidently also served as a breeding ground for rumors that voting for a union could bring dire consequences (Burrell and Fourcade 2021:230). Niya said she wanted to vote for the union but abstained after seeing social media posts that reported that many workers were terminated in a mass firing for voting for the union. “I think if I would have followed everybody else and voted for the union, I probably wouldn’t be at Amazon right now,” she said. “That’s like the truth.”
This analysis reveals how employers can weaponize social media using three tactics and points toward how algorithmic management expands their potency. Employers can easily plant or incentivize antiunion workers in social media groups to bash a union, and through known or imagined surveillance, intimidate prounion workers into remaining silent. Algorithmic management expands the opportunity to engage in these tactics because workers employed through algorithmic management tend to flock to social media groups in especially high numbers to grapple with the intense complexities of their working conditions. They bring a lingering concern with omni-present surveillance backed by the threat of automated discipline: the algorithmic imaginary that they develop at work. The employer-seeded antiunion rhetoric they encounter in social media groups is consequently less likely to be counterbalanced by prounion comments from workers.
Discussion
Labor control scholarship generally looks at how employers control workers on a day-to-day basis through coercive and consent-based techniques (Burawoy 1985; Edwards 1979; Wood 2021b). It also occasionally delves into how these techniques shape union prevention and organizing opportunities for workers (Edwards 1979; Ikeler 2019, 501–503; Rojas 2017; Lei 2021). But the tactics that employers use to resist union drives after they have begun—the antiunion arsenal—are usually analyzed separately and portrayed as fixed (Bronfenbrenner 2022; Lafer 2007; Logan 2002). In this research I bridge labor process theory with union resistance scholarship to demonstrate how control techniques, their apparatus and effects on workers can shape counter-organizing opportunities in distinctive ways. I specifically do so by looking at how the technique of algorithmic management applied in a physical workplace shapes counter-organizing opportunities.
In recognition of the multidimensionality and frequent multiplicity of control techniques in workplace regimes (Lei 2021; Vallas et al. 2022), in this section I review my findings about algorithmic management by slotting them into a framework that future scholars might use to explore how other control techniques and dimensions of firm architecture condition the antiunion arsenals of employers. This framework may be most clearly described as a progression, even though actually deploying it will likely involve shuffling back and forth between steps.
The first step is to select an architectural dimension or control technique of a workplace regime to drill into. In this article I delved into the dimension of technology and fixated on the control technique of algorithmic management, defined as “software algorithms that assume managerial functions and surrounding institutional devices that support algorithms in practice” (Lee et al. 2015). The second step is to zoom in on the constitutive elements and effects (e.g., on workers) of the dimension and/or control technique. The third step is to examine how these elements and effects may shape the counter-organizing capacities of the employer.
In this article I drew four of these connections. (See Table 1 for a detailed breakdown.) First, I looked at two constitutive elements that were, to draw on part of Lee et al.’s (2015) definition of algorithmic management, “institutional devices that support algorithms.” These were the scanners and workstation displays that direct “algorithmic instructions” as part of “Amazon’s chaotic storage” system (Delfanti 2019:45). I also noted an associated effect of these two elements on workers, the production of a paranoid “algorithmic imaginary” that haunted workers. I then showed how these two elements and this effect enhanced Amazon’s antiunion communication and intimidation capacities. Second, I drew attention to two other constitutive elements and effects of Amazon’s algorithmic management. The two elements were Amazon’s “electronic whip” (McCallum 2020) and semiautomated HR management, both powered by “software algorithms that assume managerial functions” (to draw on another part of Lee et al.’s (2015) definition). Their relevant effect was a distinct form of dehumanization. I then showed how these two elements and this effect enhanced the capacity of Amazon to restore consent by rapidly improving working conditions through “algorithmic slack-cutting.” Third, I examined Amazon’s employee mobile app. This was another constitutive element of Amazon’s algorithmic management. The app was powered by “software algorithms that assumed managerial functions” (Lee et al. 2015)—in this case, software algorithms that partially automate HR management—and it ran on a surrounding institutional device that supports these algorithms. I then showed how the mobile app enhanced Amazon’s capacity to communicate antiunion messages to workers beyond the workplace. Fourth, I highlighted well-known effect of algorithmic management: social media clustering encouraged by the control technique’s isolating, confusing and frustrating nature. I then illustrated how this effect boosted Amazon’s capacity to deter union support through surveillance, paid employee promotion and advertising.
Weaponizations of Elements or Effects of Control Techniques.
Note: DEI = diversity, equity, and inclusion; PR = public relations; VTO = voluntary time off.
See “Weaponizing Workplace Surveillance and Communication Devices” under “Findings.”
See “‘Algorithmic Slack-Cutting’: Rapidly Restoring Consent” under “Findings.”
See “Weaponizing a Mobile App” under “Findings.”
See “Haunting the ‘New Public Square’: The Weaponization of Social Media” under “Findings.”
See “Discussion.”
I don’t mean to claim that this framework will only reveal counter-organizing opportunities exclusively associated with a given control technique or dimension of firm architecture. Some constituent elements and effects of a control technique, such as scanners and digital polls, may help power multiple (older, newer, or adjacent) control techniques, not just one. But by directing attention to the apparatus and effects of a control technique, or those implicit in a particular architectural dimension, use of the framework should reveal how a technique or piece of architecture promises a certain set of counter-organizing opportunities. To apply this caveat to my findings, I am not claiming to have described a cohesive strategy meriting the label “algorithmic counter-organizing.” I have only highlighted a package of tactics that employers that use various forms of algorithmic management can expect the capacity to deploy, by virtue of the apparatus and effects of algorithmic management. Some of the summarized tactics, such as monitoring social media or theatrically wielding certain surveillance devices, may be available to employers that do not use algorithmic management, albeit likely in diluted form.
In this article I examine how the specific technique of algorithmic management influenced Amazon’s antiunion arsenal. But as the framework above is designed to address, multiple architectural dimensions and/or control techniques, such as those rooted in more cultural, institutional and organizational elements, also shape the counter-organizing opportunities of employers.
In Amazon’s case, one new article provides evidence that Amazon tried to ease its “plantation-system management” to win support during the Bessemer campaign by replacing predominantly white supervisors with Black supervisors, thereby “weaponizing Black leaders against subordinate Black workers” (Lee et al. 2024:26). In addition to corroborating the use of this tactic, some of my interviewees suggested that Amazon also tried to convert its Black Employee Network into an antiunion committee while engaging in conspicuously energetic celebration of its workers different identities, particularly black identity. Whether these alleged tactics were effective is a separate question. Lee et al. (2024:24–27) provided evidence that such efforts offended some workers.
Similarly, what Vallas et al. (2022) called Amazon’s “governmental control,” the rules and provisions that offer limited formal rights and rules to workers that can obscure exploitation, also likely shaped Amazon’s counter-organizing opportunities in various ways. (Burawoy’s [1982] concept of “internal state” and Edwards’s [1979] concept of “bureaucratic control” largely encapsulate “governmental control.”) One interviewee speculated that Amazon boosted and manipulated offers of VTO—an aspect of Amazon’s governmental control that periodically offers workers the ability to miss or leave work early—to steer workers away from a union rally that took place outside the warehouse during the campaign (Vallas et al. 2022). I uncovered no proof of this, but it may be worth noting that a former Amazon manager said such a tactic sounded “exactly” what Amazon would do. Regardless, the point is that the use of the framework outlined above is likely to yield new insights into how the labor process can animate, constrain or expand counter-organizing organizing opportunities for employers.
Conclusions
Although some past research has analyzed organizing campaigns at employers that use forms of algorithmic management (Bank Muñoz 2017; Reich and Bearman 2018), this article is the first to probe the impact of the algorithmic management during an American union election at an employer that applies an advanced form of the technique within (but also beyond) a physical workplace.
I focused specifically on the first large union election at an Amazon warehouse in Bessemer, Alabama. Drawing primarily on interviews with 42 workers at the warehouse and analysis of court records obtained through a FOIA request, I found that Amazon exploited constitutive elements and effects of algorithmic management on workers to engage in the following four counter-organizing techniques: (1) weaponizing disciplinary and communication technology that algorithmically directs and disciplines workers (within a physical workplace), (2) engaging in “algorithmic slack-cutting” to partially restore consent by improving working conditions faster and to a greater degree than traditional employers can, (3) leveraging a mobile app that automates HR management to distribute antiunion messages (beyond a physical workplace), and (4) exploiting distinctive patterns of social media activity encouraged by algorithmic management through monitoring and manipulation. These findings suggest that algorithmic management generally enhances the antiunion arsenal of employers by promising additional capacity to intimidate workers, shape their beliefs and rapidly restore consent. I will now specify the contributions, limitations and implications of this article in separate subsections.
Contributions
This study makes contributions to the sociology of work, labor process theory, labor studies, and the literature on algorithmic management and social media.
First, this study contributes to the sociology of work and labor studies by demonstrating the value of bridging labor process theory and the union resistance literature to enhance our understanding of labor control. It also provides an analytical framework to facilitate further engagement between these two literatures. Much labor process theory literature has looked at how employers control workers on a day-to-day basis through different techniques involving varying degrees of coercion and consent (Burawoy 1985; Edwards 1979; Wood 2021b). Some of this work has paid attention to how control techniques prevent union drives (e.g., Ikeler 2019) and shape resistance and organizing opportunities for workers (Edwards 1979; Hodson 1995; Ikeler 2019:501–503; Lei 2021; Rojas 2017). Meanwhile, a separate literature on union resistance has explored how employers use different counter-organizing techniques to suppress collection action (Bronfenbrenner 2022; Lafer 2007; Logan 2002). Validating calls from some labor scholars to pay closer attention to employer resources (Brookes 2018; Refslund and Arnholtz 2021), my study draws connections between concepts and findings from these two literatures to explore how the labor process and its effects on workers can also shape the counter-organizing opportunities available to employers after a union drive has begun, or what I call the antiunion arsenal. The framework I provide for undertaking such investigations involves selecting a control technique or architectural dimension and then examining how some constitutive elements and effects of the technique or dimension shape the antiunion arsenal.
Second, as the first study to focus squarely on the role of algorithmic management in a union election, this research contributes to labor studies and the algorithmic management literature by empirically showing that constitutive elements and effects of algorithmic management tend to strengthen the antiunion arsenal. This qualifies the perspective that the antiunion arsenal has remained remarkably consistent (Lafer 2007; Logan 2002) and validates an emerging concern about algorithmic management (Rogers 2023). Employers have long used one-on-one conversations, captive audience meetings, improvements in working conditions and letters to discourage workers to discourage workers from unionizing (Logan 2002). But although employers that use algorithmic management can continue to use these traditional tactics, they also can count on the capacity to leverage digital devices and their effects on workers to further influence, intimidate and restore the consent of workers.
Third, this study contributes to the algorithmic management literature by shining light on heretofore unrecognized capacities of algorithmic management that employers may exploit outside of a union election context, while also underscoring a compelling reason they may avoid applying the most advanced forms of algorithmic management in the future. Previous literature has looked at how algorithmic management enhances the control of workers through automated direction, monitoring and discipline, deskilling and data monopolization (Griesbach et al. 2019; Rosenblat and Stark 2016; Veen et al. 2019). I posit a latent capacity made available by such control enhancements: “algorithmic slack-cutting.” This technique involves rapidly alleviating the distinct dehumanizing effects of algorithmic management to restore consent, in Amazon’s case, by introducing more human managerial interaction and softening the “electronic whip.” This article also underscores that algorithmic management generally enhances the capacity of employers not just to direct the technical activity of workers, but also to shape their beliefs and preferences. Algorithmic management achieves this by embedding flexible communication channels into both the labor process and the private lives of workers, in Amazon’s case, through workstation displays and a mobile app that workers must pay attention to. Employers could rapidly address all manner of labor control issues by exploiting the flexibility of algorithmic management brought to light by these tactics to engage in a wide range of influence campaigns or more subtle, targeted or preemptive algorithmic slack-cutting (or its logical opposite, algorithmic tightening) than the form documented by this study.
Relatedly, algorithmic slack-cutting and the other tactics examined here confirm the continued role and importance of human intervention in algorithmic management (Wood 2020). They make clear that providing the flexibility for human managers to manipulate or intervene in algorithmic management offers employers more control over workers than ceding fully ceding it to computer code. Only because management could adjust or intervene in algorithmic management could Amazon manually send new messages, redeploy scanners to stoke fear in captive audience meetings, soften the electronic whip and humanize HR management to restore some consent. Indeed, the analysis confirms that “keeping humans in the decision loop” (Parent-Rocheleau and Parker 2022:11–12) can moderate the dissatisfaction caused by algorithmic management in a manner that stifles collective action (Lei 2021:297).
Fourth, this article contributes to the literature that has examined the use of social media by workers and employers. Scholars have debated whether social media benefits workers or employers more on balance (Blanc 2022; Taylor and Dobbins 2021; Wood 2015). This article lends weight to the view that employers, particularly those that use algorithmic management, have come out ahead on this front. But more research is needed to test Rogers’s (2023) hypothesis that “social media posts may invite employer retaliation more readily than offline conversations” (p. 93).
Limitations
I now highlight two limitations of this article before going on to discuss some of its implications.
One limitation concerns the generalizability of this article’s findings. This article offers a reference point for understanding how algorithmic management can shape the antiunion arsenal. But not all opportunities generated or amplified by algorithmic management at Amazon will be available to all employers that use some version of algorithmic management. For example, gig platforms can’t weaponize physical workstation displays because their workers only work remotely through a mobile app (which they certainly can weaponize) (Lei 2021:285; Rosenblat and Stark 2016). Meanwhile, employers that only use algorithmic management to automate HR management, but not to direct, surveil and discipline workers, will only be able to engage in a diluted form of algorithmic slack-cutting. Some forms of weaponization parallel to Amazon’s are easy to imagine at other employers, though. For example, hotel operators could leverage the tablets that algorithmically update room assignments for housekeepers to send antiunion communications (Mateescu and Nguyen 2019:9–12). Or retailers could intervene in algorithmically assisted or automated scheduling (Nguyen 2021:17–18) to strategically neutralize or reward workers during union campaigns. Looking forward, the more that schedules are normally determined by algorithms, the more employers could camouflage and plausibly deny such retaliation (Wood 2020:97–98). Furthermore, employers that do not use any forms of algorithmic management may still use some of the technologies necessary for the operation of algorithmic management, such as certain surveillance and communication devices, and may still monitor employee behavior on social media. They therefore also may have the capacity to weaponize such technologies against union drives. In this article I mean only to suggest that algorithmic management either necessitates or expands the availability of these tactics, by virtue of its constitutive elements and effects on workers.
A second limitation is that in this research I did not explore how algorithmic management shapes (or might shape in the future) organizing opportunities for workers, both for sparking and succeeding at these efforts. Future research can build on insights into the capacity for algorithmic management to fuel solidarity and collective contention (e.g., Lei 2021) by specifically searching for organizing opportunities that the workplace regime may generate or magnify. By shedding further light on challenges that algorithmic management can pose to worker organizing, I have sought to lay more of the groundwork for such analysis. Study of the first union election at an Amazon warehouse in more than two years, scheduled to begin February 10, 2025, at a facility in North Carolina, could yield new insights along these lines (Palmer 2025).
Implications
This work also has implications concerning the capacities of algorithmic management, the future of labor control analysis, and the study of algorithmic manipulation in general.
The first implication is the need for further research into how other workplace variables or control techniques shape the antiunion arsenal. Scholars could undertake such investigations by selecting control techniques or elements from control frameworks, such as those provided by Lei (2021) and Wood (2021b), and then drill into how their constitutive elements and effects of those elements or techniques shape counter-organizing opportunities. For example, In Amazon’s case, future research could examine how the “relational control,” “governmental control,” “normative control” (Vallas et al. 2022), “plantation-style management,” and “private policing” also used by Amazon (Lee et al. 2024) may animate, expand or constrain counter-organizing opportunities for the employer. As mentioned in the discussion, Lee et al. (2024:24–27) and my research provide some evidence that Amazon sought to weaponize plantation-style management during the campaign. Somewhat relatedly, recent news coverage invites research into how employers are leveraging diversity, equity, and inclusion initiatives, environmental, sustainability, and governance initiatives, and social justice discursive resources to engage in “woke union-busting” (Fang 2022).
The second implication concerns the “paradox of autonomy” of algorithmic management. Simply put, if we consider access to collective bargaining and union representation as an indicator or facilitator of worker autonomy (as Burawoy 1982 did), then this article highlights another way that algorithmic management, in practice, would seem to restrict rather than enhance autonomy (Noponen et al. 2023).
A third implication addresses the future of algorithmic management. Amazon was only able to exploit algorithmic management and its associated devices and effects because human managers had the ability to manually intervene. We therefore might expect employers to stop short of implementing full managerial automation (Wood 2020:12), in part so they can retain maximum flexibility to respond to worker unrest and organizing.
Taken in combination with recent work on Amazon, this article also has implications for debates about the most analytically useful ways for understanding labor control, Amazon’s workplace regime and the macro regime of “flexible despotism” more generally. The multiplicity of control mechanisms and evidence of their weaponization at Amazon affirms the value of conceptualizing the architectural dimensions, coexistence of multiple and potentially contradictory control mechanisms, and even the potential for differential application of those techniques, within workplaces regimes (Vallas et al. 2022). The fact that Amazon implements one of the most advanced forms of algorithmic management but uses a wide range of other techniques suggests algorithmic management is best understood as a control technique, rather than a holistic regime (Lee et al. 2024; Massimo 2020; Vallas et al. 2022; Wood 2021a). After all, even platforms, regarded as the pioneers of algorithmic management, use other recognizable control techniques, such as normative control (Noponen et al. 2023; Rosenblat and Stark 2016). At the same time, conceptualizing ideal types of workplace regimes retains merit because “historical tendencies in the use of legitimation and coercion” at the macro level—the core of the regime framework—are identifiable (Wood 2021a:132). Amazon, for example, may use a panoply of control techniques whose use and configuration might even vary by location. But in the United States at least, the evidence suggests its warehouse regimes broadly fit the archetype of flexible despotism (Chun 2001; Wood 2020). Amazon’s warehouse regimes merit the modifier despotism because they clearly have an overall coercive character. This character is the product of overtly coercive control techniques, such as algorithmic management. But coercion is also implicit in the firm’s consent-oriented techniques because the techniques are premised on underlying coercive conditions, including a despotic labor market, algorithmic discipline, and in the case of Bessemer at least, “plantation-style management” and “private and public policing” (Lee et al. 2024). The coercion of algorithmic management, and its dehumanizing effects, are the reason why Amazon workers deeply value the frequent offers of unpaid VTO by Amazon that Vallas et al. (2022) casts as a consent-oriented form of flexible scheduling. And the “economic whip of the market,” meager social protections, and few decent jobs explains why Amazon exercises the “relational control” that causes some workers to view their jobs there as a gift, another consent-oriented control mechanism that Vallas et al. (2022) identified in operation at Amazon. Meanwhile, Amazon’s warehouse regimes deserve flexible as a modifier because flexibility (needed above all to provide on-demand delivery) is the core engine of Amazon’s “despotism,” an engine that is configured by the aforementioned control techniques, including algorithmic management (Vallas et al. 2022:6, 22, 27).
By extension, this research gestures toward another potential contour of the flexibility in flexible despotism that has been overlooked or merely implied by the existing literature: the flexibility not just to numerically or temporally adjust a workforce but to rapidly alter or reconfigure the mix of coercive and legitimating techniques used to control said workforce, including for the purpose of quelling collective challenges to the previous configuration (that sparked the collective challenge in the first place). In other words, it points to the possibility that flexible despotism on a day-to-day basis, whether achieved partly through algorithmic management or not, may generally translate into more flexibility to use repression or cooptation to undermine collective action, not just to prevent collective action. Perhaps enhanced counter-organizing flexibility and capacity has played more of a role in the durability of flexible despotism as a macro workplace regime than has been fully appreciated.
A fifth implication relates to how employers could leverage algorithmic management to influence workers beyond a union election context, recalling pre–New Deal forms of employer control and inviting engagement with two core Marxist concepts, the “means of production” and the “means of communication” Employers could clearly use the same communication capacities latent in the apparatus of algorithmic management to seek to manipulate their political beliefs (beyond their union attitudes) and lifestyle preferences. Alex Hertel-Fernandez (2018:4, 111–12) has already shown that employers are using technologies such as e-mail, tracking software and online town halls to increasingly mobilize employees on behalf of political candidates and legislation. Uber and Lyft demonstrated how constitutive elements of algorithmic management can supplement such efforts when they helped persuade voters to overturn a California law designed to protect drivers by sending alerts to drivers and to customers warning that the law could raise prices and put drivers out of work (Hawkins 2020). This invites comparison between the power of contemporary employers to try to shape the habits and preferences of workers through invasive communication technologies today and the penetrating influence they wielded through “sociology departments” in the early 1900s, most famously that of Ford (Rogers 2023).
Such influence campaigns also suggest a Marxist proposition. Perhaps the capacity for the encompassing communication technologies embedded in algorithmic management to be used not just to direct technical activity but also to exert ideological influence may mark the beginning of a partial fusion between the “means of production” and the “means of communication,” and one that is distinct from a related fusion manifested by digital data production by consumers (Hebblewhite and Henning 2012). This is not my idea. It is an insight from Clyde, the late-20s Amazon union activist who is also, evidently, a Marxist organic intellectual. In his words,
New forms of production enable data and attentional captivity [of workers]. If [workers] weren’t as attentive to their monitors and scanners as they are, their workplace would be a lot less efficient and union drives would be a lot easier. . . . The means of production are the means of communication.
Finally, this research more generally gestures toward the value of further research on how other power-holders may leverage algorithmic administration and its supporting technologies beyond the workplace to manipulate nonelites or to advance objectives at their expense (see Esthappan 2024 for a particularly well-researched example in the judicial system). The assassination of United Healthcare’s CEO recently drew attention to what might be interpreted as a case of this: it highlighted how health insurers appear to have adopted algorithmic systems in part to deny health care claims to customers at higher rates, while also pointing to the extreme backlash such practices may sometimes help provoke. Scrutiny of this practice could induce insurers to engage in (consumer-targeted) algorithmic slack-cutting to appease customers, but to what extent and for how long may be difficult to discern (Gil 2024).
Footnotes
Acknowledgements
Thank you to all rank-and-file Amazon workers, especially those interviewed for this article. If you all stopped working, so would much of our economy. Thank you also specifically to my friends Devin Wiggs, Maura Fennelly, Kathy Copas, Karin Yndestad, and Ryan Fajardo for their essential guidance, critiques, and encouragement and more generally to my exceptionally kind cohort for their comradery and support. I also express my deep gratitude to Chris Rhomberg for his feedback and encouragement to attend graduate school, to Barry Eidlin for his general insight and advice, to Monica Prasad for shepherding this article from beginning to end, to Anthony Chen for his constructive feedback and moral support, to Wendy Espeland for her helpful comments, and to Jim Mahoney for his essential edits and encouragement that pushed this article over the finish line. Significant credit is also due to Alex Wood, Joseph van der Naald, and Peter Ikeler for pushing me to hone this article’s arguments and findings and to clarify its contributions. This article also benefited from presentations at the 2022 annual meeting of the American Sociological Association and the Labor and Employment Relations Association 2023 annual meeting. Thank you to the editor and reviewers for guiding me to engage more deeply with the existing literature and for offering many other suggestions that strengthened the article.
