It’s the Algorithm, It Decides: An Autoethnographic Exploration of Algorithmic Systems of Management In On-Demand Food Delivery Work in Amsterdam

(1)

It’s the Algorithm, It Decides:

An Autoethnographic Exploration of Algorithmic Systems of Management In

On-Demand Food Delivery Work in Amsterdam

In Partial Fulfillment of: Master of Arts in Media Studies New Media and Digital Culture

Written by: Emma Knight ID: 12149888

Under the Supervision of: Dr. Niels van Doorn

Second Reader: Dr. Thomas Poell

Date of Submission: June 28, 2019

(2)

Abstract

Companies that operate in the ‘on-demand’ platform or ‘gig’ economy rely upon machine-learning algorithms to facilitate interactions between service providers and customers. This use of algorithms has come under scrutiny within the burgeoning field of platform labor studies, particularly with regard to the ways in which platform companies utilize algorithmic systems to manage and coordinate their semi-autonomous and disaggregated workforces. However, much of this existing scholarship has avoided critical analysis of the highly subjective and individualized experiences of platform service workers who conduct their work at the intersection of digital and urban space. In turn, this thesis contributes to this growing body of knowledge by investigating the following question from an autoethnographic perspective: how has the redistribution of managerial duties to algorithmic systems impacted the experience of work for on-demand food delivery riders in Amsterdam? Specifically, this thesis investigates the lived experiences of ‘riders’ (who conduct their work primarily on bicycles) and explores how riders are impacted by the algorithms they come into contact with while working, as well as how they make sense of and develop strategic responses to these algorithmic workforce management systems. Through the use of autoethnographic research methods and in-depth, semi-structured interviews with Deliveroo and Uber Eats riders, I argue that the platforms’ redistribution of managerial duties to algorithmic systems has negatively impacted the working experiences of Amsterdam-based riders. By using algorithmic systems to govern riders’ labor, Deliveroo and Uber Eats have successfully conditioned their allegedly free marketplaces for their own profit-maximization purposes and at the detriment of riders. Furthermore, riders are negatively impacted when these algorithmic systems fail to account for the complexity of their work, and are unduly punished as a result.

Keywords_{: platform labor, on-demand food delivery, algorithmic management, Deliveroo,} Uber Eats, autoethnography, workforce management

(4)

Acknowledgements

First and foremost, I am thankful to the riders who indulged me by answering my many questions and opened up a wealth of knowledge by sharing their insights. I am extremely appreciative of their collaboration in this research and thank them for allowing me to learn from them.

I am also indebted to my thesis supervisor, Dr. Niels van Doorn. His guidance, expertise, and constructive feedback were extremely helpful throughout the thesis writing process, and I thank him for his candor and support.

I would also like to thank my parents and siblings for their love and reassurance, and for always motivating me to reach my academic goals throughout my life.

Finally, I am thankful to my partner, Max, who spent countless hours reading my work and encouraging me when I doubted myself. Thank you for always believing in me.

(5)

Chapter 1 | Introduction

It is a warm Friday night and throughout Amsterdam, cyclists sporting teal, lime green, and neon orange backpacks dart in and out of restaurants, their faces illuminated by the hazy blue glow of their smartphones. As I pull up to a restaurant for my next order, I see another rider I know, smoking a cigarette, his head bobbing to the music playing through his headphones. We chat, asking how the evening rush is going, are you getting good orders? Before long, my phone vibrates and I navigate through the sea of waiting riders to collect my order inside the restaurant. Once I get back outside, I realize my friend has left; must have gotten another order request. Such is the life of a Deliveroo rider in Amsterdam.

So-called ‘on-demand work,’ in which workers sell ad hoc services to nearby customers via digital platforms, has become an increasingly popular and visible form of labor in recent years (Hunt and Samman 7; Ticona et al. 21). While accurate estimates of the platform labor economy are hard to come by, some scholars posit that around one and a half percent of the global workforce is engaged in some form of platform-mediated work (Hunt and Samman 11). This trend has prompted researchers to study the conditions faced by people engaged in platform-based ridehailing work (Lee et al.; Möhlmann and Zalmanson; Rosenblat and Stark; Ticona et al.), food delivery (Ivanova et al.; Shapiro; Sun; Veen et al.), as well as domestic labor and care work (Hunt and Machingura; Ticona et al.). In particular, these scholars have critiqued the role algorithms play in platform-based work, as platform companies increasingly wield these allegedly neutral machine-learning systems as managerial tools. For example, on-demand ridehailing platforms such as Uber and Lyft use algorithms to match passengers with drivers and determine the fees drivers receive for providing their services. Thus, in using these algorithmic systems, platform companies do more than neutrally mediate interactions on their platforms, as they claim to. Rather, they “shape the experience of work itself,” (Ticona et al. 20).

My curiosity about the nature of platform labor reached new heights in February 2019 when I became a ‘rider’ for the on-demand food delivery platform Deliveroo. Soon after, I began riding for Uber Eats and started recording my personal reflections and thoughts on the work. In doing so, I realized that much of the existing literature that explores how algorithms shape the experience of platform service work utilizes qualitative research methods, such as ethnographic interviews with platform workers. However, these studies avoid any discussion

(6)

of the affective, highly subjective and individualized nature of platform work, and particularly on-demand food delivery work, from an autoethnographic perspective. Autoethnography, as opposed to traditional ethnography, situates the researcher as a self-reflexive actor in a particular social setting or group and uses the personal experiences of the researcher to “illustrate facets of cultural experience” (Ellis et al., 276). Thus, I seek to contribute to this growing body of knowledge by combining both autoethnographic and interview-based research methods, all in an effort to answer the following question:

How has the redistribution of managerial duties to algorithmic systems impacted the work experiences of on-demand food delivery riders in Amsterdam?

In understanding how algorithms impact riders’ work, I also seek to explore how food delivery riders for these two platforms make sense of the algorithmic systems that influence their jobs. In this sense, my aim is to critically investigate how riders produce working knowledges and competencies, and how this generation of knowledge influences their interactions with both Deliveroo and Uber Eats’ algorithmic systems. This component is worthy of study because algorithms are not one-dimensional, authoritarian objects that solely act upon platform service workers. Rather, they are “embedded within local, contextual, and multi-layered sociotechnical relations,” meaning workers’ interactions with algorithms contribute to their remaking and reshaping (Sun 14). Thus, a study of the effect of algorithms on platform service workers must also include a rigorous examination of workers’ resilience and resourcefulness as they navigate algorithmically-mediated working environments.

The remainder of this thesis is divided into four chapters. Chapter 2 establishes a theoretical framework by reviewing existing literature about the nature of platform service work. In Chapter 3, I outline my research design and methodological approach. I present the findings and analysis of my research in Chapter 4 by delving deep into my personal working experiences and the experiences of other Deliveroo and Uber Eats riders. In Chapter 5, I summarize my work and connect my findings to broader themes in platform labor studies, as well as identify avenues for future research.

(7)

Chapter 2 | The Origins of Platform Labor & Algorithmic Management

This chapter comprises the theoretical framework of my research. I first explore what platforms are, and analyze the current landscape of on-demand platform service work. I then draw from Philip Agre’s capture model of privacy and explain how this theoretical model is applicable to the ways in which corporations attempt to optimize the activities of industrial and service workers (Agre 744). Next, I analyze the capture model in the context of platform labor and introduce the concept of algorithmic management. With this established, I review literature that explores how platform workers make sense of algorithms and negotiate the constant capture of their on-the-job activities. I show that much of the existing literature which explores the effects of algorithmic systems on platform workers lacks methodological diversity, and thus outline my intended contributions to the broader research area of platform labor studies.

2.1 Surveying the Platform Landscape

The rise of the modern technology industry has heralded an era in which companies increasingly label themselves as ‘platforms.’ What are platforms, and how do they function? While the etymology of the word platform stems from its architectural definition — a raised, level surface on which people can stand — its computational meaning as “an infrastructure that supports the design and use of particular applications” has become a mainstream connotation in recent years (Gillespie, “The Politics of ‘Platforms’” 349). Not only do platforms offer a technological infrastructure for designing and building other digital products, such as software applications, but they facilitate exchanges and interactions between users, thereby orchestrating multi-sided marketplaces (Gawer 1240). Yet in facilitating interactions, platforms also shape such exchanges in important ways. As Benjamin Bratton notes, platforms “set the terms of participation according to fixed protocols,” meaning that interactions between users, and between a user and the platform itself, are governed by the technical and infrastructural choices made by the platform owners (44).

Due to this capacity to mediate interactions and serve as a technological infrastructure, platforms have gained prominence as _{the “dominant infrastructural and} economic model of the social web,” (Helmond 5). In this sense, by interacting on and with a

(8)

platform, every action taken by users generates valuable data (Srnicek 99). For example, Facebook users generate valuable data every time they click the ‘like’ button, as Facebook can use these aggregated likes to serve users targeted advertisements (Helmond 4). By owning the marketplace in which users operate, platform companies _{occupy a prime position} to accumulate, analyze, and monetize data that are produced from users’ interactions (Srnicek 44). These data may be used to train and improve machine-learning algorithms, optimize production processes, sell advertisements, or support a variety of profitable activities for platform companies (Srnicek 40).

In his book _{Platform Capitalism, Nick Srnicek identifies five varieties of platforms} and discusses the corresponding business model for each type. These include _advertising platforms_{, such as Google and Facebook, which gather and analyze data on their users and} sell these data to advertisers; _{cloud platforms like Amazon Web Services which rent their} hardware and software infrastructure to other businesses; _{industrial platforms, such as} Siemens and Intel, which sell hardware and software that enable traditional manufacturing production to incorporate Internet connectivity (i.e, ‘the Internet of Things’); _product platforms, such as Zipcar, which rent their physical assets to customers; and finally, _lean platforms_{, such as Uber, Airbnb and TaskRabbit, which mediate the exchange of goods and} services without actually owning any of these goods or services themselves (Srnicek 49). For the purposes of brevity and applicability to my object of study, on-demand food delivery platforms, I focus on the lean platform business model.

Lean platform companies first emerged in the late 2000s during an era of pivotal technological, societal, and economic change (Srnicek 85). To begin, mobile computing technology was rapidly improving; the first iPhone was released in 2007, and with its subsidized price point, more consumers than ever before could own a smartphone (Manjoo; “Demographics of Mobile Device Ownership”). At the same time, the 2008 global financial crisis was in full swing and taking its toll on working class Americans who quickly found themselves with few opportunities for secure employment (Srnicek 81). Further still, the post-crisis years were marked by widespread financial speculation, as venture capitalists were hungry for new investment vessels that could offer higher rates of return than those yielded by traditional financial investments (Srnicek 86). Thus, this confluence of factors – burgeoning technological advancements, a lack of employment opportunities for working

(9)

class folks, and available venture capital for innovative yet high-risk companies – created the conditions for lean platform companies to emerge.

Lean platform companies operate by outsourcing a majority of their business-related operating expenses (Srnicek 76). These include fixed capital assets, like software and hardware, which lean platform companies rent from other technology companies (Srnicek 83). For example, Uber uses Google’s mapping software as the backbone of its navigation system and Amazon’s Web Services for its cloud hosting (Srnicek 83). Lean platform companies also outsource their need for labor by utilizing self-employed contractors (Srnicek 76). By labeling themselves as ‘platforms,’ these companies argue they are simply intermediaries or marketplaces that connect independent providers of a service with a customer base (van Doorn, “Platform Labor” 902; Shapiro 2954). In turn, workers can be hired as self-employed contractors who are paid for each gig or task they complete rather than as employees who receive an hourly or salary wage (Srnicek 76). This designation frees platform companies from paying costly employee benefits and taxes required in traditional employment relationships, as well as offloads the responsibility of equipment supply and maintenance to contractors themselves (Aloisi 653). Indeed, classifying workers as independent contractors has stoked significant debate over the legal and ethical obligations platform companies have to individuals who use their platforms to obtain work (De Stefano 5). While this debate continues, it is essential to recognize the tenuous economic conditions American workers were facing when lean platform companies such as Uber and TaskRabbit began operating in the United States in the late 2000s. The precarious labor market conditions during and after the 2008 financial crisis allowed lean platform companies to emerge at a time when American workers were perhaps more willing to enter nontraditional employment contracts in order to make ends meet (Srnicek 81).

In addition to their hyper-outsourcing of labor and business assets, lean platform companies rely on substantial up-front capital investments to sustain their operations. Indeed, the lean platform business model is premised on the assumption that a monopoly position in the market will eventually generate substantial profits (Srnicek 87). In turn, lean platform companies use venture capital financing to subsidize the cost of attempting to achieve a monopoly position by offering customers artificially low prices and recruiting workers with large sign-on incentives and earnings, all while promising investors this allegedly temporary loss will later be made up in volume (Griswold). As the potential returns for venture

(10)

capitalists who invest in lean platform companies are significantly larger than the returns offered by traditional stock or corporate investments, there has been widespread speculation in on-demand platforms in recent years (Srnicek 86).

Uber is widely recognized as the archetype of the lean platform model, and its prominence has inspired a wave of other technology companies to develop ‘gig work’ platforms on which customers hire service providers for one-off or repeat jobs (Madrigal; Alkhatib et al. 4599). Uber’s ridehailing competitors include Didi Chuxing, Grab, and Lyft operate by connecting drivers with customers at the tap of a button (Schelifer). Similarly, on-demand courier and food delivery platforms such as Caviar, Deliveroo, DoorDash, Eleme, Grubhub, Meituan, Postmates, and Uber Eats operate multi-sided marketplaces in which fleets of service workers transport restaurant meals to customers. Domestic work has also been subject to disruption, as platforms like Handy, Care.com, and TaskRabbit have established digital marketplaces where clients can find workers who will clean their homes, care for their children and complete household repairs and tasks (Mateescu and Nguyen 8). Finally, so-called ‘crowdwork’ platforms such as Upwork, Fiverr and Amazon Mechanical Turk connect clients with geographically dispersed workers who typically complete “skilled knowledge work” tasks such as copywriting, graphic design and data entry (Jarrahi and Sutherland 2). In this vein, ridehailing and food delivery are considered ‘on-demand’ platforms in that they give customers the ability to buy services immediately from undifferentiated workers, whereas domestic and crowdwork platforms require customers to pre-schedule and select the worker who fulfills a service (Ticona et al. 21).

Moreover, this divergence in platform type influences the nature of platform service work in the sense that on-demand workers are more fungible than pre-scheduled workers (van Doorn, “Platform Labor” 904). As van Doorn aptly summarizes, lean platform companies must constantly recruit new workers, thereby creating a “‘surplus population’ of underemployed gig workers whose _{fungibility and superfluity is orchestrated through digital} platform architectures,” (“Platform Labor” 904). In this sense, some forms of platform service work are treated as plentiful and easily substitutable (van Doorn, “Platform Labor” 904). For example, on-demand food delivery workers and ridehailing drivers are highly fungible, as it does not matter much for either the platform companies nor the customers who the worker is. Domestic service workers who obtain work through platforms, on the other hand, are less fungible as they must attain high levels of trust from clients (van Doorn 904). It

(11)

is important to keep in mind the varying degrees of fungibility across different forms of platform service work, as it also impacts the way platform workers interact with and are impacted by algorithms, which will be discussed in greater depth in Chapter 4.

In this section, I have established the origins of modern lean platforms and explored the varying nature of on-demand platform service work. In the following section, I expand upon how platform companies are able to capture valuable data from workers, and how this data capture supports the redistribution of management functions to non-human management systems.

2.2 The Rise of Workforce Capture

For on-demand platform companies, the need to manage dispersed, semi-autonomous labor forces necessitates various methods of monitoring and controlling work activities efficiently. In this section, I historicize the workforce management practices of on-demand platforms by examining precursory managerial philosophies such as scientific management and Fordism. I then introduce Philip Agre’s capture model of privacy and describe how contemporary companies capture and aggregate data about workers’ activities to remotely manage labor processes. In doing so, I attempt to illustrate the evolution of capitalist workforce management practices and establish a theoretical grounding for a deeper investigation into how platform companies use algorithmic systems to remotely manage workers.

Within capitalist societies such as the United States, efforts to centralize managerial control and optimize workers’ activities have dominated the organizational calculus of industrial companies since the late 19th century (Rosenblat et al. 1). An early means of achieving these goals was introduced by Frederick Winslow Taylor, who believed that managers should scientifically study production processes (Alkhatib et al. 4965). Such scrutiny of worker performance and formalization of tasks, popularly deemed ‘scientific management,’ thus allowed companies to mitigate inefficiencies and control workers’ activities for maximum productivity (Agre 752). Industrialist Henry Ford’s approach to workforce management, now known as Fordism, similarly sought to standardize and optimize manufacturing work (Alkhatib et al. 4965). In this sense, Taylorism and Fordism not only presented novel approaches to managing production processes, but also codified bureaucratic systems of managing labor power under capitalist conditions (Braverman 90).

(12)

Deconstructing, formalizing and monitoring the activities of blue-collar, industrial workers paved the way for the use of piecework in mass manufacturing settings. Piecework reconfigured manufacturing by segmenting work activities into discrete tasks whereby workers were paid for output produced rather than time spent working (Alkhatib et al. 4961). Critics of the Taylorization of the labor process, such as Harry Braverman, argued that the standardization of piecework in manufacturing led to the subordination and exploitation of workers, whose knowledge over the production process as a whole was now reduced to his or her knowledge over one regimented and repetitive task deemed by his or her employer as the most efficient use of labor (Braverman 80). Others argued that this shift towards piecework manufacturing required managers to exercise significant oversight of workers’ activities to ensure compliance, maximum productivity, and uniformity of output (Sewell and Wilkinson 275). Without such oversight, according to a Taylorist perspective, workers would purposefully attempt to reduce their productivity and thereby output (Rosenblat et al. 2).

However, other critics of capitalist labor processes suggest that Taylorist perspectives do not consider the “organization of consent,” or the need for companies to obtain blue-collar workers’ cooperation throughout the labor process (Burawoy 27). In this sense, as Michael Burawoy argues, modern corporations do less to coerce manual laborers into achieving production targets through intensive monitoring than they do to compel workers into producing surplus labor by organizing their consent to the “legitimacy of the capitalist profit-making system,” (Zabala 282). To organize consent, firms configure labor processes such that manufacturing workers feel a certain degree of autonomy and can receive rewards for their individual effort (Burawoy 81). This configuration creates the conditions for the game of “making out,” whereby workers engaged in piecework production are financially incentivized to produce more than the required quota and employ production-maximizing strategies such as stockpiling finished goods (Burawoy 82). In other words, by organizing labor activities such that workers believe they are being presented with real choices and opportunities to ‘win,’ corporations can generate workers’ consent to production processes in which workers produce surplus labor, and thereby value, for the company (Burawoy 27). In turn, the need for companies to actively monitor workers’ production diminishes.

Yet while the need for intensive monitoring may actually diminish in industrial production work through the organization of workers’ consent, corporations must still quantify and analyze the work employees do. Particularly with the evolution of

(13)

manufacturing work in the 1980s and 1990s, in which ‘just-in-time’ production processes and lean manufacturing became the norm, the need to make workflows (and manufacturing workers) more flexible and adaptive led companies to adopt information management systems that enabled the real-time monitoring of quality and productivity (Sewell and Wilkinson 280). As the computing capabilities of such information management systems have improved, the use of software and automated systems in modern production and service industries has achieved greater ubiquity. For example, software has revolutionized modern warehousing work, as companies like Amazon have adopted complex technical systems to track workers’ activities and increase their overall efficiency (Rosenblat et al. 6). In turn, the influx of automated management systems, facilitated by the permeation of software through all facets of contemporary life, has led to the creation of “capture systems that actively reshape behavior by altering the performance of a task,” (Kitchin and Dodge 85).

Kitchin and Dodge’s mention of “capture systems” references Philip Agre’s seminal concept of the capture model of privacy. In introducing this theory, Agre characterizes ‘capture’ in terms of its use in computing vernacular, referring to the act of acquiring data to be used as an input for a computational system, such as when an order is captured by an employee at a restaurant’s point of sale station (Agre 744). In turn, he defines the capture model as a linguistic metaphor that facilitates the “parsing of human activities” for representation and practical application in computational systems (Agre 744). In this sense, by capturing human activities and translating them to fit within the confines of a computer’s representational language, these activities can be analyzed and reorganized by various actors, such as companies (Agre 744). Agre juxtaposes the capture model of privacy with the surveillance model, traditionally expressed through visual metaphors of omnipresent ‘watching’ by nefarious state actors such as the Nazi-era Gestapo or the Soviet Union’s KGB (Agre 743). Thus, while both the surveillance and capture models serve as metaphorical systems for understanding information gathering practices, the capture model offers a contemporary conceptualization that takes into account modern technology’s capacity to track, represent, and modify the activities of individuals with computers (Agre 744).

In theorizing the capture model, Agre argues that when human activity is captured for the purposes of computational expression, it must adhere to a specific ‘grammar of action,’ (Agre 746). We can think of grammars of action as linguistic rules that assign words to human activities, which in turn allows computers to express and connect these words or

(14)

activities like a sentence (Agre 746). Agre identifies a cycle of five stages that occur, often concurrently, when a grammar of action is imposed upon an activity for the purposes of capturing and representing it computationally (Agre 746). The first stage of this cycle is analysis_{, in which an existing activity is analyzed and broken down which may or may not be the same ontology participants of the activity use (Agre 746). Following this is _{articulation, in which a grammar of action is defined connected to create “sensible stretches of activity,” (Agre 746). The third phase is during which the grammar of action is imposed upon the participants of the activity. In this imposition, participants are compelled to organize themselves so that their actions can be parsed in terms of the articulated grammar (Agre 747). Fourth, the _{instrumentation takes place, in which social and technical normative forces are instituted to ensure the ongoing activity can be continuously captured and parsed (Agre 747). Finally, in the elaboration _{phase, these captured activity records can be statistically en masse for quality control, performance tracking, and error detection purposes (Agre 747). The capture model cycle reveals a truly cybernetic process in that both the grammar of action and the activity system that has been grammatized undergo constant revision 743). Such revision is necessary, because even when highly technical systems and sophisticated grammars of action are employed to capture human activities, humans will continuously interpret, develop workarounds, and circumvent such systems (Agre 748). More importantly, the capacity for people to interpret and circumvent capture is a desired effect by those who make use of this captured data, as it allows for the creation of new efficiencies. In this vein, the capture model is reminiscent of Nikolas Rose’s concept of ‘governing through freedom,’ in that the self-governing and self-enterprising capabilities of individuals upon whom a grammar of action is imposed actually support the consolidation of power and control for those who implement the grammar (Rose 147).

into ontological units,}

so the units may be} _imposition,

phase}

analyzed and monitored} (Agre

If we view the capture model in the context of workforce management, a clear technological evolution emerges from the Taylorist and Fordist managerial philosophies that preceded it. The segmentation of tasks, as well as the introduction of piecework and worker monitoring systems in the early and mid-twentieth century, paved the way for more sophisticated computer-aided methods of capturing and reorganizing blue collar workers’ activities through technical means such as location tracking devices or identity cards (Agre 749). However, where the capture model clearly diverges from traditional Taylorist and

(15)

Fordist managerial practices is in its reflexiveness; under Taylorism, workers had little to no flexibility to negotiate the imposed task, whereas workers who engage in captured activities can take an “infinite variety of sequences of action,” so long as these sequences fit within the prescribed grammar of action (Agre 752). This difference is important, because it concerns the ways workers are disciplined and their activities reorganized. As Agre notes,

“Capture does not require that control be exercised through the fragmentation of jobs and the _{a priori specification of their forms. Instead,} capture permits work activities to be disciplined through aggregate measures derived from captured information,” (752).

In a labor context, then, a worker is not disciplined by their manager for individual instances of non-compliance with a prescribed work method, but rather by the computationally assigned meaning to the aggregated measurements of their work (Agre 752). It is in the elaboration phase that workers’ captured activities are scrutinized, which supports the development of ‘tighter’ capture systems overall. As will be discussed in Chapter 4, the capacity for software powered by algorithms to carry out the elaboration phase of capture is crucial for on-demand food delivery companies in their efforts to manage riders from afar. Furthermore, Chapter 4 will analyze the capture model in the specific context of Deliveroo and Uber Eats, and explore how each platform grammaticizes their respective on-demand food delivery activity systems.

In this section, I have established the historical evolution of workforce management practices that aim to reorganize workers’ activities for the purposes of efficiency and control. Furthermore, I have introduced Agre’s capture model of privacy and differentiated its reflexive nature from earlier workforce management philosophies in an effort to demonstrate how the introduction of software has revolutionized modern workforce management. In the following section, I analyze capture in the context of platform labor generally and explore the ways in which platform companies use capture systems and algorithms to track and manage platform service workers.

(16)

2.3 Algorithmic Management in Platform Labor

As the capture of industrial and service work activities has become increasingly common, platform companies have embraced algorithms as tools to make sense of and act upon this captured data. In this section, I define what algorithms are, introduce the concept of algorithmic management, and review literature that explores how platform companies operationalize control of their fragmented workforces through algorithms. In turn, I discuss how algorithmic systems help platform companies monitor workers, evaluate performance, enforce automated decisions, and leverage indirect communication channels in their efforts to manage platform workers.

In a basic computational sense, an algorithm is “a series of steps undertaken in order to solve a particular problem or accomplish a defined outcome,” (Diakopoulos 3). When considered in a broad sociological perspective, algorithms are computational objects that can condition how humans consume information and make decisions (Kitchin and Dodge 109). Yet for the most part, both an algorithm’s presence and its effects are largely invisible to human users despite near daily interaction with algorithms, such as in one’s Facebook newsfeed (Eslami et al. 153). Algorithms, like other highly sophisticated computational technologies, are frequently described as ‘black boxes,’ because their technical complexity prevents the average user from understanding how they operate (Diakopoulos 13). Moreover, companies whose business models depend on algorithms, such as on-demand platform companies, typically refuse to disclose the ‘rules’ their algorithms follow, claiming such information constitutes a proprietary trade secret (Möhlmann and Zalmanson 5).

When thinking critically about algorithms and their effects in society, it is crucial to avoid a technologically deterministic view. While algorithms may be discussed as abstract or black box-like, the “warm human and institutional choices that lie behind these cold mechanisms” must always be taken into account (Gillespie, “The Relevance of Algorithms” 169). Furthermore, when attempting to determine the power algorithms possess in governing and ordering the social world, one cannot divorce the technical properties of algorithms from their social processes (Beer 4). In other words, a critical study of algorithms requires nuanced consideration of how algorithms are intertwined with the social world they are coded in, and the ways in which human social power is exerted through algorithms (Beer 4). Particularly within the growing body of research that examines the role of algorithms in shaping labor

(17)

processes, scholars have correctly called for a reconceptualization of algorithms as “algorithms in everyday labor,” which take into consideration the “human and non-human, as well as technical and social” elements that form algorithms (Sun 3).

In the context of platform labor, the capacity for computer software powered by algorithms to enforce managerial decisions on gig workers has gained significant attention from policymakers and scholars alike. Scholars at Carnegie Mellon University’s Human-Computer Interaction Institute have introduced the phrase ‘algorithmic management’ to describe the ways in which gig work platforms deploy data capture technologies and computer software to systematically monitor and evaluate workers ( _{Lee et al. 1603). In this} sense, by designing and redistributing management functions to algorithmic systems, platform companies are changing the experience of work itself (Ticona et al. 20).

Multiple factors characterize algorithmic management in the context of platform labor. To begin, algorithmic systems that carry out managerial decisions depend upon the constant input of information to function, which in turn requires platform companies to implement technical data capture infrastructures to track and analyze workers’ activities (Mateescu and Nguyen 13; _{Möhlmann and Zalmanson 4). As Agre notes, in the age of} captured work, tracking occurs when an individual completes a causal chain between the tracked entity — in this case, the worker providing a service — and a centralized computational system (Agre 742). In turn, as platform companies seek to remotely manage and coordinate their workforces, they must collect enormous amounts of data on workers’ activities, which can run the gamut from an Uber driver’s acceleration or braking patterns (Ticona et al. 22) to an UpWork freelancer’s keyboard presses and mouse movements (Wood et al. 64). Tracking workers with the use of GPS technology is a quintessential way platform companies collect data on workers’ activities. For example, on-demand food delivery workers have their location constantly tracked while working, thus providing the company with data about workers’ mobility patterns and time spent on each step of the grammaticized delivery process (Ivanova et al. 23). The computational parsing of these captured activities subsequently allows platform companies to catalog them, extrapolate meaning from them with the use of algorithms, and restructure work tasks in ways that benefit a company’s bottom line.

By capturing and collecting data on a worker’s activity in any given situation, platform companies can also constantly evaluate a worker’s performance ( _{Möhlmann and}

(18)

Zalmanson 4_{). With some on-demand platforms, customers take a central role in evaluating} worker performance. Rosenblat and Stark note that Uber passengers assume the role of quasi-managers by rating drivers’ performance out of five stars (Rosenblat and Stark 3774). With this rating as an input, platforms can then utilize algorithms which take into account a driver’s cumulative rating and, if necessary, discipline them based on this aggregated information (Agre 752). To be clear, human managers within platform companies determine the performance thresholds upon which a platform’s algorithms will act, such as Uber’s decision to remove drivers from its platform if they maintain less than a 4.6 out of 5 stars rating (Rosenblat and Stark 3774). Moreover, if these thresholds are communicated to platform workers, which they often are not, their enforcement varies across regional markets, as platform companies depend upon adequate labor ‘reserves’ to meet demand (Veen et al. 12). By obscuring the parameters and consequences of their performance management systems, platform companies can utilize captured customer ratings as “a bureaucratic control lever” to “elicit particular behaviors” from workers (Veen et al. 12). Thus, by using technical systems to impose a grammar of action, such as a five-star rating system, and capture the rating, platforms can then use algorithms to automate workers’ performance evaluations and enforce disciplinary actions at will.

A platform worker’s performance is similarly evaluated through the use of aggregated ratings of a worker’s compliance with a platform’s requests or requirements. For example, Lyft and Uber drivers are encouraged to maintain high acceptance rates of incoming ride requests (Lee et al. 1605). While both companies have been known to deactivate riders with low acceptance rates (without informing riders up front of any acceptance rate requirements), recent reports reveal that Uber and Lyft have stopped enforcing acceptance rate deactivations to avoid appearing as true employers (“Have You Been Sidelined by Uber?”). Instead, Lyft drivers with acceptance rates below ninety percent over a week-long period can be disqualified from receiving certain bonuses and pricing incentives, and Uber drivers who decline multiple rides in a row may be temporarily logged out of the app (Campbell). So-called ‘Taskers,’ or service workers who complete home repair and household tasks on the platform TaskRabbit must maintain an acceptance rate of at least seventy five percent, and falling below this threshold reduces a Tasker’s ability to appear in the recommendations section of the client-facing app (“Acceptance Rate”). In this manner, the algorithmic ranking

(19)

and rating of workers based on their adherence to a platform’s desired behaviors can significantly influence how platform workers conduct themselves while working.

Certainly, all service jobs require workers to adhere to rules and standards or otherwise risk punitive consequences. However, the use of algorithms to automate such punishments enables platform companies to enforce highly subjective judgement calls that promote platform companies’ profit maximization at the detriment of the worker. For example, platforms such as Handy frequently charge workers with steep fines for cancelling appointments less than forty-eight hours in advance, even if the reason for cancellation was no fault of their own (Ticona et al. 32; van Doorn, “Late for a Job”). In turn, situations in which human managers may be more lenient with punishments become opportunities to “identify the weakest link [...] and cull workforces,” as the decision to punish is no longer in the hands of a human, but rather an algorithm (Mateescu and Nguyen 17). Indeed, this culling potential does not impact all platform service workers equally. For example, on-demand food delivery riders whose labor is considered abundant are more easily replaced than in-home care workers, who are considered less fungible due to the highly personal nature of their work (van Doorn, “Platform Labor” 904). Platform companies do permit workers to appeal these decisions, but routinely make use of indirect communication channels, automated responses, and outsourced customer service representatives (van Doorn, “Platform Labor” 903). This allows platform companies to avoid any direct interaction with workers, as well as puts the onus on workers to spend additional unpaid labor hours advocating for themselves (van Doorn, “Platform Labor” 903). In this manner, platform companies often attempt to absolve themselves of responsibility for harsh decisions by blaming their own algorithmic systems, essentially dissolving “their authority into the disinterested medium of a software program,” (Tomassetti 46, qtd. in van Doorn, “Platform Labor” 903). Although human managers at platform companies are the ones delineating the rules that govern workers’ activities, they make use of algorithms within their management decisions to obfuscate their role and justify the ways in which subjective managerial rules are enforced.

In this section, I have outlined the ways in which algorithmic management tactics enable platforms to track and monitor workers’ activities, evaluate performance, and enforce disciplinary actions en masse. At the same time, however, platform workers maintain a significant degree of autonomy and freedom within this capture system. In the next section, I

(20)

review literature that discusses how platform workers make sense of and negotiate algorithmic management tactics.

2.4 Developing Algorithmic Competencies

While platform workers may not know the inner workings of the algorithmic systems which govern their work processes, they do develop nuanced perceptions of how these algorithms affect their working lives. In this section, I review existing literature that explores how platform workers attempt to make sense of the algorithms they come into contact with while working. In doing so, I establish the research context of my study and outline my intended contribution.

The ways in which platform workers experience, engage with and negotiate algorithms while working has received increased attention in recent years. Much of the academic research in this area has focused on how drivers for ridehailing platforms, namely Uber and Lyft, are impacted by algorithmically-assigned work (Lee et al.; Rosenblat and Stark; Ticona et al.; Möhlmann and Zalmanson). Similar studies of so-called “crowdworkers” on platforms such as UpWork and Fiverr have sought to evaluate the extent to which algorithmic control influences remote platform workers’ behaviors (Wood et al.; Jarrahi and Sutherland). Within the scope of on-demand food delivery work, recent publications have explored the influence algorithms exert on the experiences of on-demand food delivery workers in China (Sun), Australia (Veen et al.), Philadelphia, USA (Shapiro), and Berlin, Germany (Ivanova et al.).

As these aforementioned studies make quite clear, the experience of working under algorithmic management is not a passive one and workers do not simply acquiesce to control operationalized through algorithmic systems (Shapiro 2965; Veen et al. 12). Platform workers adapt and leverage algorithmic working environments based on their functional understandings, whether consciously or unconsciously formed, of how a platform’s algorithms work (Jarrahi and Sutherland 2). Agre reminds us that individuals who participate in captured activities are likely to adjust their conduct “based on their understanding of what will become of the data and what this entails for their own lives,” (Agre 748). In this sense, platform workers who develop these understandings not only calculatively adjust their behaviors to maximize their own financial earnings while working, but also qualitatively weigh past experiences and their own “sense of moral economy” when deciding to accept a

(21)

gig (Shapiro 2967). In this manner, workers engage in “on-the-job bodily and affective” sensemaking of a platform’s algorithms based on accumulated work experiences (Shapiro 2966).

Multiple studies of platform labor have analyzed workers’ sensemaking practices as they negotiate a platform’s use of algorithmic management tactics. Since platform workers operate in working environments in which “few work rules are communicated outright” but rather “enforced through indirect and automated means,” they often infer how a platform’s algorithms function by experiencing the effects such algorithms have on their work (Mateescu and Nguyen 12). Such sensemaking practices are evident among Uber drivers who were found to prioritize maintaining a high acceptance rate after learning one’s acceptance rate substantially influences the quality of future rides in terms of price and distance (Lee et al.; Rosenblat and Stark). Similarly, Jarrahi and Sutherland identified sensemaking activities among UpWork freelancers who reported registering for the client-facing side of the site in an effort to better understand and improve their ranking positions on the platform (5). In the context of food delivery, Sun noted the communal sensemaking practices of food delivery workers in China who utilized the social network WeChat to share platform-specific information and strategies with one another (13). By attempting to make sense of the constantly changing algorithmic systems that platform companies use, workers can develop functional understandings of the effects a platform’s algorithms have on their work.

Moreover, platform workers’ sensemaking practices allow them to develop methods of manipulating and appropriating a platform’s algorithmic systems to suit their own needs (Jarrahi and Sutherland 8). These “algorithmic competencies” result from a person’s repeated interactions with algorithms and subsequent construction of a “data infrastructure literacy,” (Jarrahi and Sutherland 9). For example, an Uber driver may notify a passenger that they have arrived before actually arriving in order to encourage the passenger to walk outside earlier, thereby reducing the driver’s unpaid waiting time (Ticona et al. 31). Similarly, an UpWork freelancer may split a large project into multiple small projects in an effort to accumulate more ratings and rank higher in the platform’s algorithmically-determined search results (Jarrahi and Sutherland 8). More practically, workers may learn to ignore certain algorithmically-mediated cues, as illustrated by delivery workers in China who used their own knowledge to navigate from restaurants to customers rather than follow the route recommended by the app (Sun 12). Thus, as platform companies modify their algorithmic

(22)

management systems based on the feedback loops supported by the platforms’ capture of workers’ activities, so do workers, who continually adjust their competencies to cope with these changes. In this vein, the reflexive manner in which competencies are formed represents the continuous work done in the elaboration phase of Agre’s capture model (Agre 747). Platform workers’ development of algorithmic competencies resembles an iterative learning process in which workers continuously adjust their activities depending on the modifications platform companies make to their algorithmic management systems.

Many researchers have utilized ethnographic research methods, such as semi-structured interviews, to uncover how platform service workers engage in sensemaking and develop algorithmic competencies. Indeed, such a methodological approach allows researchers to gather qualitative data and compare workers’ competencies across platforms, as Ivanova et al., Lee et al., Shapiro, Sun, Veen et al., and Wood et al. have done. However, to the best of my knowledge, there have been no studies to date that purposefully utilize autoethnography in combination with semi-structured interviews to compare workers’ resilience and resourcefulness across platforms. To be sure, Shapiro did participate in food delivery work in his study of on-demand food couriers in Philadelphia (2958). However, Shapiro did not highlight or emphasize his subjective experiences as a courier, and instead characterized his participation in food delivery work as means of collecting “ethnographic observations,” (2958). Thus, I seek to add to contribute to the growing body of research that analyzes the experiences of on-demand food delivery workers by triangulating the perspectives of workers with my own autoethnographic reflections and existing theoretical frameworks, such as Agre’s capture model. Incorporating autoethnography within platform labor studies is valuable because it allows researchers to experience for themselves the working environments they seek to understand.

In this section, I have established the theoretical basis of my study and situated my intended contribution to the growing body of literature that seeks to understand the dynamics of algorithmic managerial control in the context of platform labor. In the section that follows, I outline my methodology and qualitative research design.

(23)

Chapter 3 | Methodological Framework

To understand how the redistribution of management functions to algorithmic systems has impacted the work experiences of on-demand food delivery riders in Amsterdam, I employ a multimethod qualitative research design. In this chapter, I first present an overview of the platform-mediated food delivery industry and situate Deliveroo and Uber Eats within this landscape. Following this, I discuss my research methods, which include the use of autoethnography, semi-structured interviews with riders, and secondary research. I also review my method of analysis and coding procedure. Afterwards, I explain how I began working with Deliveroo and Uber Eats, outline my interview protocol, review participant recruitment strategies, provide an overview of the participants, and discuss steps taken to ensure the ethical treatment of participants. Finally, I review the limitations of my research design.

3.1 The Current Landscape of On-Demand Food Delivery Platforms

On-demand food delivery platforms have emerged quite recently within the broader industry of delivering restaurant meals to customers. According to McKinsey & Company, the traditional model of delivery, in which customers place their order directly with a restaurant and the restaurant manages the delivery process, is still the most common globally (Hirschberg et al 1). However, shifting consumer preferences and technological developments have facilitated the platformization of the food delivery industry (Hirschberg et al. 1). Online ‘aggregator’ platforms such as Grubhub and Just Eat developed first in this space by creating platforms that allow customers to order from a variety of restaurants through a single online portal or app (Blumtritt 2). In this manner, aggregator platforms offer restaurants access to more customers and automate the order process, but the restaurants still manage the logistics of delivering the food (Blumtritt 2).

More recently, on-demand platforms that manage the logistics of delivering food while also functioning as aggregator platforms have gained a small portion of the food delivery market. These on-demand food delivery platforms have enticed higher-end restaurants to utilize the flexible delivery workforces offered by the platforms and participate in on-demand delivery (Hirschberg et al. 2). While consumer trends indicate that customers prefer on-demand delivery, the substantial capital investments required to sustain lean

(24)

platform companies operating in this space has largely prevented the on-demand delivery model from capturing a larger share of the overall food delivery market (Vergauwen and Akkermans).

Within the Netherlands, Thuisbezorgd (the Dutch name for Takeaway), Deliveroo and Uber Eats are currently the only on-demand food delivery platforms. Thuisbezorgd, which was founded in the Netherlands in 2000, is the market leader and most established player (“Takeaway.com Annual Report 2017”). However, the company remains primarily an aggregator platform, as less than two percent of its total orders are completed by Thuizbezorgd’s own riders in the Dutch market (“Takeaway.com Annual Report 2017” 41). Additionally, Thuisbezorgd riders are employed directly by the company and are paid hourly wages, whereas Deliveroo and Uber Eats riders are self-employed contractors paid piece-rate wages for each order delivered. Thus, for the purposes of this research, I compare Deliveroo and Uber Eats due to their similar business models and operating practices in the Netherlands.

3.1.1 Deliveroo

Deliveroo was founded in London in 2013 and as of 2019, offers delivery in more than five hundred cities around the world (“Ambitieuze plannen”). The company entered the Dutch delivery market in September 2015 and reports to work with more than two thousand riders in the Netherlands (“Over Deliveroo”). As a lean platform company, Deliveroo relies on massive venture capital investments to subsidize the true cost of its operations in its efforts to monopolize markets. These venture capital investments allow Deliveroo to set artificially low delivery fees to attract new customers when it enters new geographic markets, as well as spend heavily on marketing and recruitment efforts to attract new riders. Since its founding, Deliveroo has raised more than one and a half billion dollars from venture capital investors, including a $575 million investment from Amazon in May 2019 (“Deliveroo - Crunchbase”). While Deliveroo’s revenues doubled in 2017, its overall financial losses before tax rose to nearly $240 million due to the company’s expansion into new markets and higher operating expenses (Ram and Hodgson). Despite this, Deliveroo is rumored to be exploring an initial public offering in 2020 (Boland).

Deliveroo’s business model resembles a multi-sided marketplace in which it must manage customers’ demand for food, riders’ supply of labor, and working relationships with

(25)

partner restaurants. Customers can order from a selection of “high-quality and diverse” restaurants via the Deliveroo mobile app or website (“Deliveroo FAQ”). Once the customer places an order with a restaurant, Deliveroo’s order algorithm, which the company calls “Frank,” matches the order with a rider who either accepts or rejects the order request (“Meet Frank!”). Throughout the delivery process, riders must notify the rider app when they complete certain actions, including when they arrive at the restaurant, when they collect the order, when they are approaching the customer’s address and when the delivery is completed (“Tech round-up”). These stages indicate Deliveroo’s grammatization of the delivery process and efforts to capture riders’ work activities for subsequent computational parsing, which will be expanded upon further in Chapter 4.

In addition, Deliveroo utilizes a shift booking system which requires riders to book one-hour shifts in a city’s various booking zones one week in advance (“Self-serve booking”). Riders who provide “the most consistent, quality service” are given early access to reserve shifts in the system. This early access is determined by a rider’s ‘statistics’ which are based on the percentage of booked sessions a rider attends (attendance rate), the number of sessions the rider has worked during the busiest periods (super-peak participation rate), and if a rider cancels a booked session less than twenty four hours before it starts (cancellation rate) (“Self-serve booking). In turn, Deliveroo utilizes an algorithm to rank riders based on these statistics, which will be further explored in Chapter 4.

Finally, the employment status of Netherlands-based riders has risen as a point of contention for Deliveroo in recent years. In the past, Deliveroo employed Netherlands-based riders directly, but has since reclassified riders as self-employed contractors _{(“Deliveroo} FAQ”). With this employment classification, riders must register with the Dutch Chamber of Commerce and pay business taxes if they earn more than €596 in a four week period (“Regular Riding”). As of August 2018, all Deliveroo riders earn money according to a distance fee model which compensates riders based on the distance and time needed to complete a delivery (“Distance Fees”).

3.1.2 Uber Eats

Uber, the on-demand car service platform based in San Francisco, launched Uber Eats in late 2015 in select cities in the United States and Canada (Hempel). The company’s existing international infrastructure allowed the service to quickly expand to cities around the

(26)

world, such as in Amsterdam, where Uber Eats began operating in 2016 (van den Outenaar). Currently, Uber Eats works with more than twelve hundred restaurants in the Netherlands, and aims to double this offering as it aggressively expands in large and medium-sized Dutch cities (“Uber Eats Plans”).

The experience of riding with Uber Eats is very similar to that of Deliveroo, with a few key differences. To begin, Uber does not require riders to reserve sessions in advance. Rather, riders can simply go online at any time and, depending on consumers’ demand, will receive order requests from nearby restaurants (“How do I receive delivery requests?”). Another difference concerns the information Uber Eats riders receive prior to accepting an incoming order request. When Deliveroo riders receive a new delivery request, they are informed of the drop-off location and the fee they will receive for delivering the order. Uber Eats riders, on the other hand, only learn the delivery location after they confirm they have collected the order from the restaurant, and payment information is displayed after the delivery is completed. These information asymmetries and their effects on riders’ work will be further investigated in Chapter 4.

Uber’s reliance on investment capital is very similar to Deliveroo’s, albeit at a much greater scale. In May 2019, Uber became a publicly traded company following a highly anticipated initial public offering that led to the company’s current valuation of $82 billion (Davies and Wong). Despite this valuation, however, Uber has only once turned a quarterly profit in its ten years of operations, and has sustained net losses of almost $6 billion in the past two years alone (Isaac and Conger). Uber has not revealed any losses associated with its Uber Eats business specifically, but has self-reported that Uber Eats’ revenues grew by nearly one hundred and fifty percent in 2018 and constitute thirteen percent of the company’s overall revenue (Newcomer).

In the competitive on-demand food delivery market, Deliveroo and Uber Eats face numerous challenges in terms of profitability, competition, and appeasing its myriad of stakeholders. Particularly within the Netherlands, where aggregator platform Thuisbezorgd maintains a dominant position, Uber Eats and Deliveroo face significant market pressure to expand their operations while keeping costs low for customers and earnings high for riders. This overview of the on-demand food delivery landscape in the Netherlands has sought to establish the technical and operational contexts in which Deliveroo and Uber Eats riders work. In the following section, I outline my qualitative research design.

(27)

3.2 Qualitative Research Design

This qualitative research design makes use of semi-structured, in-depth interviews in conjunction with autoethnography to gather information on the working experiences of Amsterdam-based food delivery riders. As Taina Bucher recognizes, “accessing people’s personal stories and experiences with data and algorithms can be tricky,” (32). Indeed, algorithms are intangible and largely invisible to those who come into contact with them (Bucher 31). Yet in this invisibility, algorithms hold tremendous power; as Eslami et al. note, algorithms influence how individuals receive information and can affect how they behave (153). This power is deserving of interrogation, especially considering the increasingly influential role algorithms play in fundamental parts of life, such as one’s work. To access the personal stories and perspectives of food delivery riders in Amsterdam, I follow in the methodological footsteps of platform labor scholars (see Ivanova et al., Jarrahi and Sutherland, Lee et al., Möhlmann and Zalmanson, Rosenblat and Stark, Shapiro, Sun, Ticona et al., Veen et al., and Wood et al.) who utilize qualitative, semi-structured interviews and ethnographic participant observation research methods.

Where I diverge from these previous studies, however, is in my intentional use of autoethnography as opposed to ethnographic and participant observation research methods. Participant observation, in its traditional scope, concerns a researcher’s presence in a social group or setting for the purposes of scientific investigation (Schwartz and Schwartz 344). In doing so, the intended outcome of both observing and participating is the objective and empirical generation of “human understanding” about that particular group’s behaviors (Tedlock 70). Autoethnography, on the other hand, subverts the notion of participant observation by emphasizing the researcher’s own “observation of participation,” (Berger 506). In this sense, autoethnography allows researchers to reflexively interrogate their own subjectivity, emotionality, and influence on their research, thus resulting in a more narrative than scientific assessment (Ellis et al. 274). Moreover, autoethnography as a research method uses the researcher’s own embodied experiences and perspectives to explore and problematize broader theoretical claims, as the researcher’s experiences must constantly be in dialogue with existing theoretical concepts (Ellis et al. 276).

In my view, the aforementioned studies provide important theoretical and empirical contributions with regard to the nature of platform work. Yet what these studies seemingly

(28)

avoid is discussion of the affective, highly subjective, and individualized nature of platform work – and particularly on-demand food delivery work – from a first-person perspective. From my experiences as a rider, I found food delivery work in Amsterdam to be isolating and solitary. Long stretches of cycling alone were dotted with brief interactions with restaurant staff, customers, and the occasional fellow rider. Thus, pursuing participant observation as a research method in the context of on-demand food delivery work is pragmatically challenging. Not only is it difficult to observe the experiences and behaviors of riders who are constantly in motion as they work, but it is nearly impossible to participate in any larger social group, as this collective work dynamic does not exist for riders in Amsterdam. Therefore, to explore how the redistribution of management functions to algorithmic systems has impacted the nature of work for food delivery riders in Amsterdam, it was necessary to experience the work for myself. In doing so, I could observe and analyze how I, as a rider, was impacted by this redistribution, as well as compare my experiences with those of other riders whom I interviewed. Finally, I could triangulate my experiences and those of other riders with existing theoretical concepts, all in an effort to most effectively investigate my research question.

Autoethnography as a social science research method is not without its own politics. Some scholars seek to distance autoethnography from the positivist tendencies of contemporary ethnographic research, arguing instead that autoethnography “was designed to be unruly, dangerous, vulnerable, rebellious and creative,” (Ellis and Bochner 433). In this view, autoethnography offers a highly aesthetic and intimate method of inquiry that allows researchers to acknowledge and give importance to their own personal and interpersonal experiences that result from their participation in a social group or setting (Ellis et al. 277). Another camp of scholars, however, believe that autoethnography should be considered as a method for gathering “empirical data to gain insights into some broader set of social phenomena,” (Anderson 387). In this sense, so-called analytic autoethnography requires researchers to demonstrate a “commitment to an analytic agenda,” (Anderson 387). Ultimately, the debate between these two epistemological views centers upon how the self should be positioned, and if this positioning should be expressed in an evocative, artistic narrative or more sociological, scientific manner (Ellis and Bochner 438).

My use of autoethnography falls somewhere in between these two schools of thought. While participating in food delivery work over the course of five months as a rider for

(29)

Deliveroo and Uber Eats, I recorded field notes and took many app screenshots. These field notes took into account quantifiable components of my work as a rider, such as the number of deliveries made, earnings per order, and restaurant waiting times, but also my own emotional responses and interpretations of personal interactions with other riders, customers and restaurant staff. In this sense, my intention was document moments that would provide insight into the broader working experiences of food delivery riders in Amsterdam, as well as produce “aesthetic and evocative thick descriptions of personal and interpersonal experiences,” (Ellis et al. 277). Thus, situating myself within my field notes was not a “decorative flourish” nor “exposure for its own sake,” but rather an essential component of my field research (Behar 13-14).

In addition to autoethnography, I conducted semi-structured, in-depth interviews with riders to learn how they experience their work and how, if at all, they perceive each platform’s use of algorithms. Semi-structured interviews allow researchers to gather “descriptions of the life world of the interviewee in order to interpret the meaning of the described phenomena,” (Kvale and Brinkmann 3). The semi-structured aspect of such an interview relates to both the format and flow of the interview, meaning that the interviewer prepares questions which target certain issues, as well as conducts the interview in a manner that is “flexible enough for interviewees to be able to raise questions and concerns in their own words and from their own experiences,” (Brinkmann 285). In this vein, semi-structured interviews can resemble a conversation between the researcher and interviewee, which requires the researcher to maintain a significant degree of reflexivity in their role as a co-constructor of meaning (Heyl 370). Semi-structured interviews were best suited for my purposes as they allowed the participants and I to engage in thoughtful, free-flowing dialogue while still ensuring specific topic areas within my broader research area were addressed. Furthermore, all interviews were conducted face-to-face, which allowed me to witness participants’ non-verbal cues and facial expressions, thus providing for a richer interview experience (Brinkmann 290).

Lastly, I conducted primary and secondary research by consulting online resources, conducting interface analyses of each platforms’ rider applications, and communicating directly with Uber Eats and Deliveroo rider support representatives. In particular, I consulted Deliveroo and Uber Eats’ rider ‘Frequently Asked Questions’ webpages to gather publicly available information about each companies’ policies, payment models, and algorithmic

It’s the Algorithm, It Decides: An Autoethnographic Exploration of Algorithmic Systems of Management In On-Demand Food Delivery Work in Amsterdam