Introduction

The rise of the so-called platform economy is reflected in the now ubiquitous presence in our cities of one of its most characteristic figures: the delivery rider (Lin et al. 2020; Seghezzi et al. 2021). The term rider is used to identify food delivery workers employed by digital platforms such as Glovo, Uber Eats and Deliveroo. The day-to-day work of these riders is controlled by a variety of algorithmic procedures that order the stages of the delivery—from order allocation through pick-up at the restaurant to final delivery to the customer—and evaluate the riders' performance (Griesbach et al. 2019). At the same time, these algorithmic processes match the courses of action of those involved in the business (riders, suppliers, and customers), ensuring that they coordinate effectively, with a common interest in minimising the duration of the service (Möhlmann et al. 2021). The processes are usually organised via mobile applications, on which users can check the progress of their delivery, and riders log the completion of a sequence of tasks assigned to them by the algorithms. This digital mediation leads to the characteristic image of delivery riders constantly checking their mobile devices while waiting outside restaurants to receive the next order, as well as during delivery, as they check for recommended routes.

The relationship between riders and algorithmic mediation processes has been studied in the literature in the context of the organisation of work in the platform economy (Basukie et al. 2020; Galière 2020; Griesbach et al. 2019; Möhlmann et al. 2021). This link has been explored from two main perspectives. First, authors analysing the dynamics of control and governmentality have examined the control effects that algorithmic mediation exerts on workers (Anderson 2017; Revilla and Blázquez Martín 2021). These studies indicate that these effects take on certain ‘panoptic’ and ‘dramatic’ overtones, with the worker ultimately becoming subordinate to the control logics predefined by the algorithm (Neyland 2016; Seaver 2018; Ziewitz 2016). The second field in which the subject is of interest is the computerisation of the workplace and the progressive shift of certain workers towards functions that are less valued or have less added value in the production process (Dachs 2018). In such a context, delivery workers may be seen as the means whereby the sequence of tasks prescribed by the algorithm is effectively fulfilled (López-Martínez et al. 2021). These two common approaches to algorithms and riders in the literature converge to give a common image of delivery workers that always seems to swing between two extremes: they are either passive subjects, without the power to make decisions and whose participation is reduced to completing a sequence that is external to them; or they are resistant subjects, who rebel against the prescriptions that the algorithm tries to impose on them (cf. Cameron and Rahman 2021; Fernàndez and Soliña Barreiro 2020; Revilla and Blázquez 2020; Veen et al. 2019; Woodcock 2020).

Although algorithms are social constructs (Seaver 2019) and therefore need to be addressed in the light of their ‘entanglement’ with the practical actions of users (Gillespie 2014), the two approaches described above structure the subjectivities of riders around an assumption that there is a clear separation between the algorithm and the delivery workers. This distinction has a variety of consequences. Firstly, the ‘technical’ nature generally ascribed to algorithms means that they are taken to be autonomous, independent objects and that their effects must therefore be assumed to be impartial, rational and reliable (Barocas et al. 2013). Likewise, viewing algorithms as autonomous objects removes the margin of indeterminacy inherent to any technical object, thereby disassociating them from all the assumptions, evaluations, inferences and exclusions that have shaped them (Simondon 1980). Thirdly, both approaches ignore the clear performative dimension of rider-application entanglement, and fail to examine the way in which this agency is contextually shaped and re-specified (Ziewitz 2017).

To challenge these assumptions and highlight other ways of exploring the effects of algorithmic mediation on the work practice of delivery riders, we need only think, for example, of the many unforeseen events that may arise during a delivery. Imagine a rider arriving at her destination only to discover that the drop-off location is in a residential area where she has to circumvent a number of access control elements, and that when she tries to gain access, the customer does not answer. Given the limited time available to her to drop off the order, the rider is forced to employ a variety of strategies that will enable her to incorporate this eventuality into her work and effectively complete the sequence of tasks displayed in her application. In this example, she could use the functions on the application to take a screenshot on her device, in order to show that the customer is not responding despite having been notified of her arrival. In this (real) example, we can see how the operability of the algorithm does not reside solely in its capacity to set certain courses of action to be completed by an external agent, but also in its ability to be open to unforeseen situations, precisely to ensure that it continues to mediate when faced with changing conditions (Finn 2018). In this regard, we see how riders are continually obliged to readjust the gap between the formal result returned by the algorithm and the contingent situation in which the order is filled—which is affected by multiple variables that are impossible to control definitively.

Let us now imagine that the same rider, having completed her previous delivery, has left the area in which she usually operates. In order to return to this more familiar area, she employs a new strategy, consisting of successively logging in and out of the platform; she logs out after completing her delivery and starts to head back towards familiar territory, only logging back in once she is close to those familiar areas. In this way, she can modulate the frequency with which she is allocated orders as she gradually gets closer to more familiar locations. In this regard, the algorithmic mediation must be set in such a way that it can favourably combine contingencies with riders’ schedule of tasks, times and expectations (Griesbach et al. 2019; Shapiro 2018).

In view of these strategies, and given the algorithmic reasoning that permeates their calculations, workers can hardly be viewed as being either uniquely subject to the track-and-trace functions imposed by the algorithm or as resisting it (Callon and Law 2005). On the contrary, these examples show that riders must be able to articulate diverse logics based on a calculability (Thrift 2008) that involves incorporating the action of the algorithm itself into their working practices (Shapiro 2018).

This ‘recursive loop’ between the calculations of the algorithm and the riders' own exercise of reflection gives rise to what some authors have defined as a ‘participatory subjectivity’ (Bucher and Helmond 2018; Gillespie 2014; Seaver 2022; Ziewitz 2017). Participatory subjectivity is defined in this case as being a disposition to action in which the users' practice takes into account and makes allowances for the algorithmic logic; just as the algorithm itself is constantly reworked depending on the input it receives, including information on user behaviour. A number of different fields refer to the development of participatory subjectivities. For example, authors such as Bucher (2012, 2017) link the category to users' own visibility on the platforms; by adopting the operating rules of the algorithm in their practice, users can augment their recognition. Bucher uses the concept to explore the way Facebook users relate to the algorithm that ranks their status updates in feeds. She notes that in making this adjustment, users adopt practices that take into account the way they assume the algorithm functions (i.e., commenting on a friend's photo, in the expectation that this will be a central input in the algorithmic model). Participatory subjectivity also emerges as engagement in a particular mode of reasoning. This is evident in the experiment performed by Ziewitz (2017) with colleagues and students, which consisted of taking a series of city walks in which the purpose was to create an algorithm. The participants agreed on the instructions they would use to decide what route to take (e.g. ‘At any junction, take the least familiar road’), as well as on the complementary criteria to be used in the event of unforeseen situations (in the example above: ‘If all roads are equally familiar, go straight on’) (Ziewitz 2017, p. 4). These walks link participatory subjectivity to the practical context and the way of thinking about courses of action, showing how the algorithm is a recursive process in need of continuous contextual re-specification. Placing himself in the tradition of ethnomethodology, the author pays particular attention to the process whereby the walkers become immersed in a routine as a result of participating in the algorithm's logic of practical reasoning—an algorithmic rhythm that shows how the logic deployed is an embodying logic that mobilises both reason and experience, and induces action. It is, so to speak, a logic that envelops, entangles, involves, and entraps those who interact with an algorithm.

Echoing the contributions of these authors, we believe that participatory subjectivity is a very useful starting point from which to develop other perspectives on algorithmic mediation in the practice of riders that go beyond merely characterising them as (passive or resistant) subjects external to the algorithm. In any case, two remarks should be made. Firstly, although the idea of participation may suggest a participatory intention, and while it is true that this article will show strategies or tactics through which riders intentionally engage with the algorithm, we acknowledge that participation, as highlighted by Kelty (2020) in tracing its genealogy, is a complex and sometimes elusive concept. Therefore, in this article, we understand participation not solely as an intention or conscious experience but also as an embodied, affective condition that arises from the entanglement with an algorithmically-organized environment. Secondly, although our work does not discuss the relationship with the algorithm in terms of resistance or co-optation, it is important to recognise the role of algorithms as tools for organising information and shaping patterns of action. Algorithms contribute to the (re) production of categories of objects, subjects, and collectives (Stevenson and Helmond 2020). Consequently, our argumentation will closely examine the ways in which algorithmic systems can contribute to generating (in) visibilities and exclusions.

Based on an ethnographic study carried out with riders from different home food delivery platforms in the City of Madrid (Spain) between September 2021 and February 2023, this paper therefore explores the algorithmically-mediated practice of riders in relation to the concept of participatory subjectivity. In our proposal, like other authors (Chen and Sun 2020; Lin et al. 2020; Sun 2019; Woodcock 2020, 2021), we consider essential to address the entanglement between the algorithm and the riders by analyzing the workers' experience. Additionally, we acknowledge that this relationship is subjected to labour regulatory frameworks within the platforms. In Madrid, the distinction between contract types (salaried and various self-employment arrangements) is crucial, as different employment models involve diverse strategies and possibilities of action enacted by the algorithm, which, in turn, impact the agency of the riders. For instance, as we will explore, salaried riders may be obligated not to reject orders and must adhere to specified working hours, while self-employed workers experience a distinct form of regulation or workforce control, often relying on gamification strategies to incentivize their participation. It is important to consider how the limits of algorithmic management (Woodcock 2021) may contribute to the reproduction of unequal labour structures related to origin and social class. Thus, in the case of Madrid, certain population profiles (the new undocumented migrants, usually of Venezuelan origin) tend to insert in the self-employment regimes offered by the platforms.

Considering the rider's experience and how it varies according to the diverse conditions that affect their work, we take the particularities of the specific practices analysed in Madrid to list three categories in which this participation and this participatory subjectivity are expressed and developed. Specifically, we describe how subjects themselves seek to be recognised by the algorithm; how, on other occasions, they seek to be ignored or passed over; and how, sometimes, it is the design of the algorithm itself that fosters participation or immersion in the proposed practical logic.

This study followed an intensive format based on two focus groups and eight semi-structured interviews with both riders and engineers involved in designing geolocation applications using algorithmic procedures. Based on the socio-demographic profile of the majority of riders in Spain (Adigital 2020), the riders were selected on the basis of their gender, country of birth, time working as delivery workers and form of labour relationship (salaried, subcontracted salaried as part of a fleet of delivery workers, self-employed and self-employed renting accounts) (Table 1). The engineers were selected based on a non-probability, purposive sampling (Green 2001).

Table 1 Description of the riders participating in the study
Full size table

The material generated in the focus groups and interviews was audio-recorded and transcribed. Screenshots of the riders' applications were also taken to document certain real-time interactions with the device. The analysis included triangulation of data and techniques. All participants signed an informed consent form outlining the study goals, the purpose of their participation, and potential research dissemination channels.

Algorithmic mediation in riders' working practices

The work performed by riders is determined by the applications of the digital platforms, which allocate orders, track deliveries and evaluate employee performance (Andersson Schwarz 2017; Griesbach et al. 2019). Each of these processes is managed by algorithmic procedures, the results of which are set out in the form of instructions displayed sequentially to workers on their device interface as they complete the chain of actions involved in each delivery (Möhlmann et al. 2021).

Typically, a rider begins work by logging into the application and completing a series of checks related to identity (e.g., facial recognition), compliance with health regulations (if the use of face masks is required) and responsibility (related to the condition of the company vehicle and devices, in the case of some salaried workers). Once the log-in protocols have been completed, the rider enters active service and is considered by the platform as a potential recipient of an order. When an order comes in, the rider receives a visual and audible message indicating the pick-up and drop-off points (usually via a Google Maps interface). Some platforms also display the estimated time to destinations and the suggested route. If the rider accepts the order (by marking the relevant option), the application links the agents involved in the various stages of the delivery and organises their courses of action. The routine of this coordination is established in accordance with the distances to the pick-up and drop-off points, and the maximum time the rider has to deliver the order. These variables are adapted to the type of vehicle used. The sequence ends when the customer and/or rider mark receipt/delivery of the order on their application. The rider then becomes operational again and available to make a new delivery.

Each of these stages involves different actions for tracking and verifying the status of the process, which the delivery worker has to complete in the application. Completing these actions enables them to advance through the sequence of actions (for example, once they have confirmed that the order has been picked up at the restaurant, they can initiate the delivery), and displays new indications about the operation. In some applications, for example, the routes and distances to the destination are displayed once the delivery worker accepts the order. Note that each of these actions requires temporal and spatial matching of the actions of all those involved in the process, in order to ensure that the parties meet and that the service takes as little time as possible. In this regard, the function of algorithmic mediation is first to ensure that the different agents' courses of action converge, ordering the sequences of interaction between customers, suppliers and delivery workers. This ensures that these interactions—e.g., pick-up at the restaurant, or final drop-off at the customer's home—are adjusted to a shared space–time geography, with each agent having access to information about the time and location of the other agents (Wu and Zheng 2020). Secondly, the function of algorithmic mediation is also to record these interaction sequences, for the purpose of logging the rider's work (for calculating wages and evaluating performance).

Consequently, by ordering and logging the successive actions to be completed by the rider, the algorithmic mediation creates a sort of framing in the operations of delivery workers, which operates much like a framework of directive action that contains and determines their work (Cañedo-Rodríguez and Allen-Perkins 2023a; Griesbach et al. 2019; Sun 2019). However, one central feature of algorithmic mediation in platform economies is that coordination tasks are not established only once, but rather need to be continuously readjusted (Cañedo-Rodríguez 2016; Thrift 2008). As in the examples given in our introduction, if algorithmic mediation is to ensure effective coordination between the parties even in unforeseen circumstances, the entanglement between the rider and the application must be capable of anticipating and introducing these contingencies into its calculation models (Finn 2018; Shapiro 2018). It must be able to operationalise these possible contingencies in such a way as to enable them to be dealt with algorithmically; in other words, they must be able to be translated into inputs that the algorithm can recognise (Finn 2018; Ziewitz 2017). In the example in which the customer did not respond to the rider's prompts, the rider was able to upload a photo to the app showing her location, in order to provide additional evidence that she had in fact been at the drop-off point. This could be used in the event of a possible complaint from the customer (as we will see, delivery workers believe that much of their work depends on positive customer reviews). Here, it is important to note that the option of uploading photos is one of the many functions that have been gradually added to the apps. Such updates illustrate the processual dimension of the algorithm and highlight a second direction in the way algorithmic mediation operates: although the algorithm induces frameworks of directive action among delivery workers, workers themselves also appropriate these frameworks, developing practices which in turn lead to continuous modifications in algorithm design (Cañedo-Rodríguez and Allen-Perkins 2023a; Griesbach et al. 2019; Ziewitz 2017).

We aim to show how this adjustment in the frameworks of algorithmic action—the standardisation of courses of action and the incorporation of contingencies—is imbued with a participatory subjectivity that encourages riders to internalise the algorithm's operating patterns and respond accordingly, while at the same time, the algorithmic frameworks are re-configured based on inputs from the riders themselves. In the following section, we will describe a number of practices that illustrate the main directive framing governing riders' operations. At the same time, within these frameworks of action, we will highlight different modes of subjectivation among riders in their work, and analyse the strategies they use to ensure compatibility between the frameworks of directive action, the contingencies that arise during delivery, and their own personal agenda of tasks and expectations.

The concept of subjectivation or modes of subjectivation is grounded in extensive literature, which we will not review here. We use subjectivation to address the specific relationship between subjects, themselves, and their environment—a recursive relationship that yields flexible patterns of thought and action. These modes of subjectivation are influenced by various vectors that extend beyond the riders' work. In this text, we broadly examine the modes of subjectification concerning the diverse ways socio-technical mediations shape riders' work and life practices, as well as how they attribute meaning to them in their everyday lives.

For the sake of clarity, these strategies have been grouped into three modalities or categories involving three modulations of participatory subjectivity: strategies involving recognition by the algorithm; strategies that explicitly seek non-recognition by the algorithm (to avoid interaction), and strategies implemented by the platforms themselves to promote the involvement of the riders and the deployment of their participative subjectivity.

Participatory subjectivity in rider practice

The importance of being recognised by the algorithm

To start working, riders open the app and follow a series of instructions to verify their identity and ensure that they are compliant with the different rules and regulations governing their job. During this initial stage, the first algorithmic filter used by platforms to ensure that the person connecting to the application is the actual owner of the account is a system of facial recognition. In terms of agency with the application, the labour dependency relationship influences the way in which the initial entanglement between the rider and the algorithmic mediation is established. This is because the facial recognition check is circumvented by the shared assignment, rental and use of personal accounts, which is a common practice acknowledged by platforms, labour inspectors, trade unions and riders (Casas-Cortés et al. 2023). In the case of salaried riders, these access protocols are usually only performed once a day, when the rider logs into the application. In contrast, the operations of riders renting or sharing accounts entail a greater degree of indeterminacy, since they are impacted by the frequency and timing of facial recognition checks (Casas-Cortés et al. 2023). This indeterminacy introduces a dynamic whereby riders adapt the start of their workday to match the schedule of the actual account owners and their availability to perform the facial recognition and log into the app. Nonetheless, workers say that facial recognition is not an ‘effective’ barrier against identity fraud, because any sanctions imposed for breaching the platform's terms and conditions (e.g., temporary exclusion from the application) are generally lifted if the worker claims there has been an error in the facial recognition system (a common occurrence, even among delivery workers who do use their own accounts).

Successive updates of the applications show how the platforms themselves have ended up incorporating these practices into their algorithmic mediation processes. For example, some platforms have removed the margin of indeterminacy linked to this first interaction, requesting facial recognition only once a day, at the start of the working day. This allows account sharers and renters to be free to start work as soon as the account holder has performed the check. Moreover, the terms and conditions of platforms such as Glovo now allow accounts to be subcontracted to third parties (Glovo 2023), in apparent violation of the labour legislation in effect in some of the countries in which they operate, including Spain (Lomas 2023).

Once they have logged in, riders can start receiving orders and are linked to the other delivery workers, customers and suppliers who make up the network mediated by the platform. Algorithmic mediation then starts to organise the agents' courses of action. In the case of riders, it orders sequences of order allocation, delivery tracking/verification, and performance evaluation. As already discussed, the routine of each of these phases generates a framework of directive action in the riders’ work, depending on the algorithmic models used by each company (Griesbach et al. 2019). It should be noted that these frameworks for action are, to a large extent, conditioned by the nature of the company's labour relationship with its riders. In general, the remuneration of salaried riders (i.e., those who have an employment contract with the platform) does not depend on the number of orders they deliver, but on the length of their working day; whereas the earnings of self-employed riders are proportional to the number of deliveries they make. These differences not only affect the way in which the framing is configured, but, as discussed earlier, also impact the delivery workers' workflow and their ability to plan ahead (Shapiro 2018; Sun 2019).

Take, for example, the time and place at which they start work. Salaried workers usually have a fixed workplace and/or a pre-established timetable set by the company. In contrast, the first thing self-employed riders have to do is find out when and where they are scheduled to start work. There is a common assumption among delivery workers that starting the day in a location with a large concentration of suppliers—such as shopping malls or areas with a large number of restaurants—favours subsequent order allocation sequences. While the work of salaried delivery workers is not dependent on the place from which they log in (which, in some cases, is mostly configured within a radius of action from their workplaces), self-employed delivery workers commonly talk about travelling to places with a high concentration of suppliers, either to start working there or (as discussed below) to try to restrict their drop-off points to these areas.

If starting work in these ‘hot spots’ fosters certain expectations about order allocation, the riders' algorithmic expertise also extends to evaluating the feasibility of the orders they are assigned (even several orders at once, as Sun 2019 and Chen and Sun 2020 note in their fieldwork with riders in China). After being assigned an order, the rider receives a notification with details of pick-up and drop-off points. They then have a window of about one minute in which to accept, reject or reassign the order. In terms of their scope for action, salaried delivery workers are under a contractual obligation to accept orders; generally speaking, they cannot turn them down—except by omission, when they fail to spot the notification and, after the acceptance period expires, the order is reassigned to another rider. In contrast, self-employed delivery workers do have the option of refusing orders. This happens mainly in the case of ‘long-distance’ orders—i.e., those that exceed what they calculate to be the limit for feasibility (around 5 km according to our informants working on bicycles); and when, for a variety of reasons, they do not think they will be able to complete the order within the deadline set by the company (e.g., in the case of suppliers who tend to take a long time to prepare orders and delivery routes traversing areas of high traffic density).

It is clear that, when assessing the feasibility of orders, riders use different calculation strategies. Whereas salaried delivery workers are somewhat limited in their range of action because they have to accept the orders assigned to them, self-employed riders can choose which orders to accept and when to do so. Such calculations are the basis for one of the categories with which rider subjectivity is usually identified in existing discourses: the ‘autonomy’ to choose their working hours and the orders they deliver (Griesbach et al. 2019). Although this category is imbued with a certain business rhetoric that tends to link riders' ability to choose with their ‘freedom’ (Lehdonvirta 2018), it is clear at this point that the need to ensure their 'autonomy as choice' requires self-employed delivery workers to incorporate a whole series of strategies into their practice that salaried delivery workers do not need to employ. We shall return to this issue later.

A closer look at the strategies deployed by delivery workers in their work reveals that discourses on ‘opacity’ are a constant in this field (Cañedo-Rodríguez and Allen-Perkins 2023b). This ‘opacity’ with regard to the design of the algorithms and users' lack of knowledge of how they work means that many of the tactics used by delivery workers are based on trial and error, following the assumptions, rumours and conjectures that are passed amongst them (Diz et al. 2023). In terms of the possible effect that accepting or rejecting orders may have on other processes, riders do not know what criteria the platforms use to allocate orders, what parameters they employ to calculate how much the rider will make on each delivery, or what effect accepting or rejecting orders will have on their evaluation metrics. They merely conjecture that long-distance orders tend to offer higher payment, and that repeated rejections will prompt the company to ‘switch off’ the rider. According to our informants, minimising rejections leads to a higher frequency of assignments. This fosters narratives of ‘self-exploitation’ and ‘self-demanding attitudes’ among delivery workers, especially the self-employed, who tend to work beyond the time they had initially set themselves (Griesbach et al. 2019). At the same time, this willingness to be driven by assumptions about how the algorithm works is reinforced by the strategies used by the platforms themselves to encourage order acceptance. For example, as well as the feature whereby orders can only be accepted within a certain time window, in some platforms, the rejection of an order triggers new waiting times and notifications that the rejected order is in the process of being reassigned—which de facto delays the entry of new orders.

Conjecture about how the algorithm allocates orders also leads to other strategies in which riders reorient their actions in ways they think will match the patterns used by the algorithm in its calculations. One of the most common is to take up position at specific points that confuse the platforms' geolocation systems. For example, in cities like Madrid, it is very common to see riders waiting on bridges or above one of the many underpasses that criss-cross the city centre. This is due to an assumption among the workers about the procedure used by the algorithm to calculate the expected earnings of a delivery, which, among other inputs, supposedly takes into account the distance between riders and the pick-up and drop-off points. According to these shared beliefs, the algorithm establishes this amount using at least two spatial parameters: the linear distance between the rider and the pick-up and drop-off points; and the actual distance they have to travel to reach those locations. The reasons the riders stop on bridges or over underpasses is that it enables them to minimise the linear distance to the pick-up points (thus augmenting their proximity to a ‘hot spot’), while at the same time increasing the distance actually travelled during delivery (thus increasing their anticipated earnings).

When we analyse the strategies employed with regard to facial recognition and the tactics designed to boost order allocation, we observe how this search for practices aimed at using the algorithm in favour of the rider questions the panoptic nature with which some approaches conceptualise algorithmic control within the platform economy (Woodcock 2020). Also, we begin to understand that riders do not only base their working practices on fulfilment and verification of the chain of actions displayed by the application; rather, in effectively completing these tasks, they use practical knowledge to align them with their assumptions on the workings of the algorithm (Shapiro 2018). This ‘know-how’ is particularly important during the delivery phase, when riders are more likely to have to respond to contingencies that have not been envisaged in the algorithmic models. There is one particular period of delivery around which a whole series of strategies have been created, namely ‘drop-off’ time. Drop-off time is the length of time riders have to deliver the order. It is measured from the moment the app detects that they have arrived at their destination to the moment the delivery is received by the customer. The application uses at least two measurements to verify that they comply with this time: the moment when it considers that the delivery worker has arrived at the delivery point, based on their geolocation with respect to the final location (usually a linear distance of 100 m to the drop-off point); and the moment when the customer and rider notify the system that the order has been delivered, using the application.

Given that the start of the drop-off time is based on a spatial estimation that uses linear distances, within the space of five minutes riders need to factor in all the contingencies that may arise between the moment when the application notifies them that they have arrived and the moment they actually hand over the order. Our informants frequently mentioned that during drop-off, they have to factor in different inputs impacting the delivery, which are not generally considered by the algorithm. These include access controls to the locations; availability of parking places for their vehicles; the architectural design of residential areas; building height; the presence of lifts (and whether or not they are working); the weight of the order; apps 'crashing' during drop-off; and non-responding customers (Sun 2019; Wu and Zheng 2020).

The issue of drop-off time reveals the performative nature of the entanglement between rider and application, given the formers’ ability to translate contingencies into algorithmically computable inputs (Cañedo-Rodríguez and Allen-Perkins 2023a; Ziewitz 2017). This can be seen in the strategies used by delivery workers to try to minimise contingencies and, at times, use the workings of the platform to their advantage. One strategy commonly reported among our informants was to try to extend the drop-off time to as close as possible to the 5-min limit that the application provides for. Less frequent tactics used to achieve this ‘break’ include motorcycle delivery in pairs (a method used by one couple of self-employed Venezuelan riders in Madrid), where one drives the vehicle and the other hands over the order; or the inclusion of practices during drop-off that make up for the absence of certain functions on the application. One such approach is to make sure the customer does actually check the ‘order received’ option. As our informants noted, some customers (mostly university students) deliberately fail to check in the order, resulting in the rider having to bear the cost. Although some platforms have intermittently introduced a variety of systems to ensure receipt of the order (such as using a PIN code to verify hand-over), one recurring practice is to take photos of the customers upon drop-off, in order to generate additional documentation that can be used as evidence if necessary. As noted, the possibility of uploading photos to the apps is a function that has been introduced as the result of the practices of the riders themselves. However, in one focus group, several informants said that the reduction in drop-off time (from the previous 10 min to the current 5 min) was the result of the algorithmic frameworks being reformulated to adapt to the common practice among riders of extending their break time between orders.

One final process that shows the need for riders to be algorithmically recognisable is the rider evaluation system, in which customer ratings are assumed to be a fundamental variable. In general, delivery workers associate a higher score in performance metrics with a higher frequency of order allocation or, as discussed below, better options when it comes to choosing work shifts. This perceived relationship between higher ratings and increased frequency of order allocations prompts delivery workers to ask customers to give positive scores when they complete the delivery. This is especially significant among riders who rent or share an account. Indeed, renting or sharing accounts—a common practice among the Venezuelan riders we interviewed—is viewed as a ‘help’, because it allows riders to protect their administrative status in Spain, while at the same time enabling them to access a ‘customer portfolio’ without having to build up their reputation from scratch.

The importance of being ignored by the algorithm

In contrast to the aforementioned strategies used by riders to render themselves more readily recognisable to the algorithm, there are also situations in which their aim is to achieve the exact opposite, namely to not be recognised or to be ‘ignored’ or passed-over. These ‘disconnection’ strategies are used to adjust (or disrupt) order allocation sequences. For example, as discussed earlier, there is an assumption among riders that being close to points with a high concentration of suppliers tends to boost order allocation. One tactic used by self-employed riders is to repeatedly switch the application on and off, in order to give themselves time to gradually return to areas of higher demand. This tactic, which is not available to salaried workers, is also commonly used by self-employed riders when they want to return to familiar locations, interrupting the allocation of orders in order to reduce the distance to those points.

In the case of salaried delivery workers, possible disconnection strategies are limited to the moments immediately before the end of their working day, when the assignment of a new order might mean that they have to work beyond their scheduled finishing time. The tactics used in this case vary. Some try to ‘extend the drop-off time’ or—when the app functions allow—activate ‘busy’ mode, which interrupts the order allocation sequence. We also identified cases in which workers get help from colleagues, for example by exchanging mobile devices in the workplace.

A third strategy appears to replicate the previously-mentioned practise among some customers of not registering their order, and involves the reallocation of incoming orders. Indeed, order reallocation is seen not only as a way of ‘disconnecting’ in order to interrupt assignment sequences, but also sometimes as a means of exploiting algorithmic knowledge ‘in the rider's favour’. Examples include delivery workers who ‘eat for free’ by picking up an order without marking the ‘order received’ option, and then reassign the order to a third party.

At the same time, as well as the tactic of taking up position at particular points that play around with the linear and real distances to the pick-up and drop-off points, other strategies involve circumventing rider geolocation so as not to be recognised by the algorithm (Heiland 2021). For example, we came across practices intended to identify ‘dead spots’ where low signal coverage makes it difficult to allocate orders, including underground locations and points situated at a distance from residential areas (Diz et al. 2023). Another tactic is the use of applications that generate false locations, although the riders we interviewed reported that this is not frequent practice. Apps such as Fake GPS Location and Spoofr tend to be more common among drivers on platforms such as Uber, because they enable them to increase distances to pick-up points and thus hike fares (Abuya 2017; Arubayi 2022). In contrast, among food delivery riders, the few examples of these applications being used date back to the time when the platforms were just starting up, before the code in the applications was updated to prevent riders from tampering with the geolocation system.

Again, on/off strategies show that margins of action vary greatly between riders. Although some riders can choose their working hours and the orders they deliver, they also have to use a large number of strategies to stick within areas of highest demand and to coordinate their personal and working schedules. In contrast, other delivery workers are more limited in the extent to which they can disconnect, at least in terms of when they can implement such options. However, they themselves say that this limitation on the strategies they can employ does not necessarily mean that they are less autonomous in their work. On the contrary, because they have no need to employ strategies involving moving nearer and further away (they are assigned allotted delivery areas and times), salaried riders associate their ‘autonomy’ not so much with their ability to choose, but with the ‘relaxed’ pace of their work routines, the ‘freedom’ of having paid holiday time, and the ‘security’ of being able to access health coverage in the event of an accident at work. This demonstrates that the autonomy of delivery workers cannot be viewed in the abstract, but is rather configured in each specific case by the particular relationship established between them and the application.

Gamification strategies and competition

As well as the strategies used by riders to render themselves more recognisable (as described in "The importance of being recognised by the algorithm" section) or less recognisable (see "The importance of being ignored by the algorithm" section) to the algorithm, the platforms themselves sometimes add functions designed to encourage riders to internalise algorithmic frameworks of action. Some of the most common are gamification strategies, which allow delivery workers to enter inputs into the algorithmic models with the aim of giving them a sense of ‘individual choice’ (Gandini 2019; Sun 2019). These models influence the range of actions the workers perceive as being possible, thus generating different affordances in their work (Bucher and Helmond 2018).

In our ethnographic study, we identified two main models of gamification. First, there are what are known as ‘challenges’, in which delivery workers receive a bonus for completing certain tasks in a limited timeframe and within specific time slots. These include cumulative bonuses (€25 for completing 18 orders, €30 for 21 orders, etc.) and payments for performing tasks in very specific time windows, such as completing 8 orders in two off-peak hours. According to some of our interviewees, these challenges reinforce the playful dimension of their work, as they allow them to enjoy cycling, one of the main attractions they find in their occupation. However, they also point out how these challenges, while offering an incentive to the delivery workers, clearly demonstrate that ‘the algorithm is rigged’, because on occasions, the order allocation chain is interrupted just before they complete the assigned challenge.

Second, platforms such as Glovo allow self-employed delivery workers, once a day, to adjust the value of the so-called ‘multiplier’—a numerical coefficient usually between 1.0 and 1.6 by which delivery workers indicate how much they are willing to charge per order, in relation to the baseline reference of 1.0 used by the algorithm in its calculations. Given that this value can only be adjusted once a day, riders have to project their calculations to take into account expected demand conditions and incorporate multiple variables in addition to those mentioned in the previous strategies, such as the weather, the day of the week, the work period and any possible high-attendance events (Cañedo-Rodríguez and Allen-Perkins 2023a). To calculate the most suitable multiplier value, they need to be aware that increasing it before peak demand periods generally leads to fewer orders, but they also need to anticipate these variable conditions and to recognise when they are having a ‘good day’ (more orders than expected), so they can raise the multiplier value before moments of peak demand. In turn, the multiplier links the riders' expectations to those of others. As Diz et al. (2023) note, it is not uncommon for riders who frequently cover the same delivery points to agree to set the same multiplier value, in order to ensure that they are all allocated orders. This strategy is a way of avoiding working below the value estimated by the algorithm (which used to happen before Glovo raised the minimum value of its multiplier to 1.0 from the initial rate of 0.7) (Alcalde Lucas 2021). This setting also shows how affordances not only shape choice options among users, but are interpreted in terms of riders’ intentions or what they want to achieve (Bucher and Helmond 2018).

As well as employing gamification strategies, the platforms also promote participatory subjectivity among riders with features designed to foster competition (Griesbach et al. 2019; Rani and Furrer 2021). One such practice was the shift selection system that some platforms used in previous models. The system opened twice a week at a given time; riders then had fifteen minutes to choose their work shift from the slots shown in the application. The hours were then allocated in order, using rider evaluation metrics. Riders with the highest scores (on a 100-point scale, based on customer ratings and other undisclosed variables) were given first choice in selecting their hours. Workers with lower scores could only access the slots that had not been picked up in the previous stages. Once the selection process was complete, riders could ‘drop’ any hours they were unable or unwilling to work. These hours appeared outside the shift selection opening windows and could be selected by any riders, regardless of rank. In most cases, this system led to intermittent shifts that varied from week to week, as well as shifts outside peak demand periods. It also created situations in which riders were ‘fishing for hours around the clock, because they could open up at any time’. Shift selection contributed to this situation by fostering a particular routine that ended up ‘trapping’ riders (Ziewitz 2017), generating ‘self-exploitative’ and ‘self-demanding’ attitudes such as those discussed above, because workers tended to try to ‘hog’ more hours than they had initially been willing to work (Griesbach et al. 2019). Finally, this system encouraged riders to incorporate the companies' metric evaluation procedures into their work practice. As one worker said: ‘because you're afraid you'll get a lower score […] you ask customers: “Please give me a smiley face, please”’.

Although the shift selection system is no longer included among the app functions, the platforms now employ other systems to encourage competition between riders. One of the most common, especially in what are known as ‘dark stores’ (distribution centres not open to the public, from which riders deliver products bought by customers over their mobile apps) is the use of devices listing the fastest delivery workers (e.g., white boards in the work centres) or ‘employee of the week’ lists circulated on the messaging channels shared by the workers in each centre. The rationale behind these devices is that they supposedly ‘motivate’ delivery workers. However, our informants believe that rather than fulfilling this function, they tend to be viewed as evaluation devices, similar to those provided by other metrics, because when incentives are offered to the fastest riders, they consist simply of batches of products from the establishment itself that are close to their expiry date.

Conclusions: weaving the algorithm

The practices of delivery riders are structured around an algorithmically-organised environment, in which the interaction sequences of the agents involved are woven and unwoven around the links made possible by the algorithm on a daily basis. In this paper, we have explored the way in which algorithmic mediation fosters a participatory subjectivity among the agents involved, taking into account the mutual interaction between the algorithmic calculability that permeates the reasoning of users and the way in which such behaviours reconfigure the design of the algorithms themselves.

We have posited three categories that demonstrate and expand on the concept of participatory subjectivity. First, we consider strategies that involve algorithmic recognition of users, such as ID checks, order allocation, maximisation of number of deliveries, and rider evaluation. These strategies show how riders incorporate assumptions about how the algorithm works—proximity to ‘hot spots’, avoiding order rejection, maximising customer ratings—and also reflect how the algorithmic frameworks are updated on the basis of input from riders. Secondly, we discuss ways in which riders implement strategies that enable them to be overlooked by algorithmic mediation. These tactics, which include logging in and out of the platform, reassigning orders, or identifying geolocation ‘dead spots’, highlight variations in the margins of action enjoyed by different workers. These operational limits show how some of the categories commonly used in the field (such as the ‘autonomy’ of the delivery workers) cannot be understood in the abstract, but rather in the way they are configured in the practical context of action. The third category, in which the algorithm itself is designed to encourage user participation (e.g., gamification and competition strategies), reflects the way in which algorithmic mediation orients delivery workers' affordances and how those workers interpret these options in relation to their expectations.

The characterisation of these three modes of subjectivation shows how the algorithmic frames that predefine riders' operations are conditioned not only by unforeseen variables that arise during delivery, but also by other more fixed variables, such as the riders' type of contract. As shown by the examples we have given of riders' participatory subjectivity, the labour relationship (i.e., salaried or self-employed, in its various forms) influences the different affordances within the algorithmic framing. This requires us to view the effects of algorithmic mediation from the situated and particular context in which the relationship takes place.

The three forms of participatory subjectivity outlined here show that, in the field of home food delivery, the subjectivity of delivery workers is linked to vectors that constantly refer back to values of adaptation, learning, response and mutual interaction. This openness calls into question the view of riders as passive or resistant subjects and, from a more complex perspective, points to the performative condition of their relationship with the application and a certain degree of algorithmic calculability in their work.

Although in this paper we have focused our research on the way in which participatory subjectivity is configured around the rider/application relationship, we believe it would be useful to analyse how these modes of subjectivation operate with regard to other agents in the network. We also believe that extending this study to the dynamics that structure the practices of restaurants and customers would provide new perspectives from which to observe the reconfiguration of the relationships between agents, thus complementing our analysis of the forms of subjectivity that shape algorithmic mediations and the way in which algorithms themselves are employed as social-technical fabrics.