Friction and Promise in Data Labor

Abstract

This article develops friction as a methodological lens and mobilizes it to examine an unusual data labor arrangement in Finnish prisons. The concept of friction highlights how penal policies in a Nordic welfare state both support and intervene in tendencies to view data labor as a uniform future development. While the friction lens draws attention to infrastructural arrangements and institutional forces, it also foregrounds human involvement, imaginaries, and aspirations. Translating aspirations into institutionally rooted practices requires effort and resourcefulness, ultimately producing a “homegrown” version of data labor. The prison offers pushback against how the global data labor infrastructure attempts to reconfigure the human. We demonstrate that the friction lens provides a novel way to analyze data-based automation from a critical perspective, without collapsing differences or overlooking the potential for hopeful pathways. This turns friction into a future-oriented concept that opens multiple views on how things might evolve.

In recent years, we have faced a repeated response when presenting our work on Finnish prison inmates participating in the production of AI training data: scholars in critical data and algorithm studies start nodding knowingly. What else is prison data labor if not an effort to harness the prison-industrial complex in service of the global data-extraction machinery? The questions posed by the audience after our talks typically concern the exploitative aspects of prison labor, as if to confirm their expectations. The critical responses are fed by the current expansion of datafication and related AI initiatives that appear to spread similar data power everywhere. Data power is indeed at play in Finnish prisons, and it would be ridiculous to suggest that prisons are not characterized by highly unequal power relations and extreme forms of control (Rhodes 2001). Prison data labor intersects with the carceral environment and the asymmetries of “dataveillance” that penetrate “every fiber of the social fabric” (Van Dijck 2014, 205).

Rather than verifying the exploitative features of prison data labor, however, our aim is to pose more precise questions about data labor arrangements to query what they might tell us about the present and future processes of datafication. STS perspectives challenge us to oppose simplistic forms of politicization that hastily identify power structures and swiftly generalize from specific cases. Although, or perhaps because, similar traces of datafication are detectable in remarkably different places—including prisons, universities, schools, hospitals, and insurance companies—data power should not be understood in a manner that makes “everything everywhere the same” (Tsing 2005, 1). Despite similarities, processes of datafication and data extraction are defined by gaps, inconsistencies, and pushback that continue to matter (Milan and Treré 2019; Pink et al. 2018). By addressing the tensions and discrepancies involved in the human assistance that AI requires, we can address the contradictions and paradoxes of data power more comprehensively (Høyer 2023).

With these goals in mind, we employ friction as a methodological lens that allows multiscalar analysis and is sensitive to what is treated as universal and what is particular. As a physical phenomenon, friction is the force that resists motion between two surfaces in contact with each other. The notion of friction, however, has inspired numerous scholars to explore what happens in encounters where different entities or qualities rub against each other. The friction lens, we suggest, supports the study of developments across different scales and dimensions: friction might be geopolitical, infrastructural, interpersonal, affective, or all of these at once. Encounters with friction prompt corrective action (Edwards 2010), and friction provides the necessary traction to keep global power in motion (Tsing 2005). This means that friction can serve both as a source of resistance and a productive force. Some types of friction have negative consequences, while others appear to be beneficial (Tomalin 2023). Identifying what friction does allows the unpacking of developments that aim to reduce friction but also generate it (Bates 2018; Tsing 2005).

After introducing the friction lens, we mobilize it to revisit and build on our case study of data labor, or what is called “training AI” in Finnish prisons (Lehtiniemi and Ruckenstein 2022). Datafication processes in the prison context operate at the margins of digitalized society, but they are still intimately connected to global automation imperatives, public values promoted by means of institutional arrangements, entrepreneurial inventiveness, and future visions of society. Identifying different frictions at play in the studied case allows the examination of both universally appealing automation aims and locally embedded practices. As we discuss below, the data labor infrastructure that keeps the global AI machinery in motion is a response to one form of friction. This friction is caused by the human-powered nature of automation as, paradoxically, human input is required to keep seemingly smooth automated processes in motion (Gray and Suri 2019). Recognizing the many possible interpretations of infrastructure (Larkin 2013) and platform-mediated work (Stephany et al. 2021), we define the global data labor infrastructure as a shared framework utilized by various digital platforms and arrangements. This schematic infrastructural form serves as a reference point that allows us to highlight the important differences present in Finnish prison data labor. Mechanical Turk (MTurk), originally developed by Amazon as an internal service to remove duplicates in its catalog, is probably the best known example of the global data labor infrastructure. MTurk was established as a marketplace intermediary whereby external users could manage relations between data workers and “requesters” (Irani 2015). The explicit goal of MTurk is to reduce friction between workers and requesters. The service aids in transforming labor into a disembedded resource that can be tapped as effortlessly and “frictionlessly” as possible.

Critically for our argument, the data labor infrastructure that takes advantage of global disparities in income and welfare by transforming human labor into a disembedded resource, used with as little friction as possible, produces other forms of friction. They emerge when certain types of data labor, such as those involving speakers of minor languages like Finnish, remain underrepresented or absent due to the global data labor infrastructure. The creative endeavors and personal hopes and hesitations that address the latter friction become a focal point of our inquiry. We are interested in how data labor arrangements in Finnish prisons respond to the friction caused by this absence of data workers and attempt to resolve it by seeking alternative ways forward.

Our analysis is based on consecutive collaborations between two startup companies and the Prison and Probation Service of Finland (RISE). A company called Vainu initially launched a pilot project with RISE in early 2019, which was abruptly terminated in June 2020 due to a shift in management. To explore the frictions that drove, sustained, and ultimately dissolved this pioneering data labor arrangement, we draw upon materials from our initial case study (Lehtiniemi and Ruckenstein 2022), which includes interviews with company representatives and prison officials, as well as documentary evidence.

In a later twist of events, a second startup, Metroc, began a new collaboration with RISE in early 2022. By 2025, the ongoing partnership had led to the introduction of data labor in several new prison wards. We secured permits to carry out research inside prisons, which authorized us to complement the initial case study with interviews and observations, and to incorporate prisoners’ perspectives into our analysis. We are aware that prisoner is a contested term in English, but it is the most direct translation of the neutral Finnish word “vanki,” which is used in prison legislation and reflects local prison practices. It is also the word that those incarcerated in Finnish prisons use to refer to themselves. The two data labor initiatives allow us to consider how friction gives rise to or interferes with data labor arrangements, while sustaining the ideas, aspirations, and collaborations that initially set them in motion.

The Friction Lens

In studies of technoscience, friction is often considered something that slows down, resists, or inhibits things. Helping to understand forces that restrict datafication, the concept of “data friction” introduced by Paul Edwards (2010) describes the challenges that arise when data moves between different people, organizations, and systems. These include the costs in time and human attention, as well as the potential for data to be misinterpreted, lost, or garbled during these movements. The term “science friction” extends the concept to the difficulties encountered when scientific disciplines collaborate and attempt to share and integrate data (Edwards et al. 2011). Even if friction as such has unhelpful or adverse effects, activities born out of the necessity to overcome friction prompt critical reflection, adaptation, and, occasionally, innovative responses that might not have emerged without friction. For instance, responses to friction in cross-disciplinary communication involve clarifying misunderstandings, refining methods, and improving data and metadata.

Anna Tsing's (2005) notion of friction, on the other hand, is coined for the purpose of probing how global connections sustain claims of universality despite becoming locally reconfigured. Much like capitalism, science, and political ideals—the universals that Tsing had in mind—datafication and related technologies spread through “aspirations to fulfill universal dreams and schemes” (Tsing 2005, 1). As Tsing observes, concepts and ideas need to travel across difference to become universally appreciated. Supported by logics such as surveillance capitalism (Zuboff 2019), data-driven value capture appears to be advancing at an accelerating pace. For Tsing, however, the forms of capitalism by which we live are not merely the result of a triumphant march of universal tendencies, but of the friction caused as universally cherished aims and local practices come together in people's lives, organizations, and societies. Datafication-related developments are exemplary in this regard, as they mobilize people and organizations to prepare themselves for impending futures with abstract and universally appealing concepts like big data or AI (Ruckenstein 2023).

For technology professionals, friction is typically something that needs to be reduced. The aim might be to ease the flow of data between organizations or to promote service design that ties humans and machines into seamless loops (Tomalin 2023). Natasha Dow Schüll (2018) describes how visions of “frictionless living” guide technology designers in their aims “to gratify us before we know our desires.” Tsing (2005), however, regards friction as a much more societally attuned and resilient notion. Friction gets in the way of the smooth operation of what appears to be “global data machinery,” but it is also necessary for keeping things in motion and letting them grow.

Thus, viewing datafication through the lens of friction suggests that data-related powers should not be treated as isolated from attempts to actively define and redefine them. Stefania Milan and Emiliano Treré (2019, 321) call for conceptual and methodological approaches to processes of datafication that can “grasp the obscure developments, the cultural richness, and the vibrant creativity emerging at the margins of the “empire.” We follow their suggestion by using friction as a methodological lens to examine infrastructural efforts to promote datafication and attempts to strengthen local aims in handling data-related developments. Our goal is not to define or typify frictions (which are found in many variations and in all sorts of relationalities), but to situate friction to study the practices and aims that it and its reduction set in motion. We promote an approach that supports studying parallel and consecutive developments across different scales and dimensions, and is sensitive to how friction is responded to and dealt with. In the case of data labor, friction draws attention to human involvements, imaginaries, and aspirations that promote different kinds of data-related futures. Since friction calls for creative endeavors, it promotes hope and anticipation, and makes people act. This approach aids in detecting ambivalences, contradictions, and paradoxes that characterize developments that are either promoted as globally disembedded or are deeply embedded in locally institutionalized aims.

We begin our exploration by examining how data labor arrangements in Finnish prisons respond to friction, attempting to resolve it by highlighting how penal policies in a Nordic welfare state support and intervene with tendencies to think of data labor as a uniform future development. We then show how datafication developments, in this case related to the human data work necessary for automation, can take distinct forms. For example, practices in Finnish prisons appear to promote the global data labor infrastructure, but still appear globally anomalous. Translating aspirations into institutionally rooted practices requires effort and resourcefulness that ultimately produces “homegrown” versions of data labor arrangements. In the homegrown variation, Finnish prisons do not simply turn into an extension of the global data labor infrastructure. Rather, institutional forces offer noteworthy pushback to how such data labor arrangements attempt to reconfigure the human. Different and potentially incompatible aspirations and aims meet in the prison, suggesting that unlikely collaborations might be needed when infrastructural support is not readily available for the automation on which companies typically rely.

Balancing Interests in a Nordic Welfare State

The research that led to the discovery of multiple frictions in local processes of AI training took place in Finland, where digital technologies feature in future strategies and publicly funded projects to anticipate how society and citizen skills need to be updated to thrive in the digital age. The Finnish self-image has been techno-oriented at least since the end of the 1990s, when Nokia's mobile phones became integral to the project of being at the forefront of the global scene. Digital technologies are understood to be forward-looking and modern, and figure centrally in public discussions and representations that frequently conflate science and progress with the latest technologies (Berglund 2007; Tarkkala and Snell 2022). Eeva Berglund (2007, 77) concludes that “Finland is a curiously technophile country.” In more practical terms, this means that digital developments continue to be a quest for new Finnish success stories. The authorities, driven by pressures to do more with less funds, are increasingly turning to digitalization and automated procedures in search of efficient and timely service delivery (Räisänen 2024). However, besides being a feature of austerity measures (Dencik and Kaun 2020), the active promotion of digital technologies comes with some distinct twists. We uncovered attempts by civil servants to protect welfare ideals and support existing penal policies through the implementation of new data labor arrangements (Lehtiniemi and Ruckenstein 2022). Within the prison context, data labor is employed to serve rehabilitative goals and protect prisoners’ rights, including the right to participate in digital developments. Policies that outline incarceration practices thus explain how the aspirational qualities of prison labor, aimed at training AI, can tie data labor arrangements to rehabilitative objectives.

Finland, like the other Nordic countries, retains a “rational and humane criminal policy” with roots in the welfare state's consensual political culture and public trust (Lappi-Seppälä 2007, 219). The policy's effects materialize in penal practices. When punishment is assigned, alternatives to imprisonment such as suspended sentences and community sanctions are predominantly used. Consequently, Finnish incarceration rates are low on an international scale, especially compared to systems of mass incarceration. In 2022 Finnish prisons held a total of 2,776 people, an incarceration rate of 50 per 100,000 inhabitants. The European median is more than twice as much (Aebi, Cocco, and Molnar 2023), while the US rate is around ten times higher (World Prison Brief n.d.).

Finnish prisons are divided into “open” and “closed” facilities. Open prisons, which provide roughly a third of all prison capacity in Finland, resemble minimum-security facilities in other countries, offering inmates the possibility to work or study outside of prison grounds. Closed facilities are maximum-security prisons, fortress-like arrangements of sturdy buildings containing wards, cell blocks, and walled exercise yards. People incarcerated in a closed prison—the setting of AI training—are usually serving longer sentences, or prison officials consider a less controlled environment is unsuitable for them. Closed facilities hold, for example, repeat offenders or those that have committed serious offenses, such as violent or drug-related crimes.

Broadly speaking, the stated aim of imprisonment in Finland is to increase the ability to readjust to society and lead a crime-free life. The Finnish policy maintains that prisoners are not “slaves of the state,” and their rights are protected similarly to other citizens (Lappi-Seppälä and Nuotio 2019). To minimize detrimental outcomes, the Finnish penal code (Ministry of Justice 2005, Section 3) states the intent to maintain the health and functional abilities of the incarcerated. Loss of liberty is regarded as sufficient punishment, and the imposition of additional hardship—during imprisonment or after—is neither required nor permitted. The “normality principle” expressed in the penal code asserts that conditions in prison should correspond as closely as possible to living conditions outside the walls (Lappi-Seppälä and Nuotio 2019). Normality in this sense is considered to involve the presence of stable structures of everyday life, including work and leisure. Yet normality also implies that as society undergoes changes, so should life in prison. Our civil servant interlocutors in the prison stress that today's normality involves keeping up with the digitalizing society, including maintaining or developing prisoners’ skills with digital tools and services.

Our goal is not to paint a rosy or idealistic picture of Finland or Finnish prisons. Prisons are universally harsh and hopeless places, and the primitive cell conditions in the oldest Finnish prison, built in the 1880s, have raised repeated criticism from the Council of Europe Anti-Torture Committee (CPT 2021). We stress the welfare state context to emphasize that processes of datafication do not simply land in un-normed territory and shape it to their liking (Pink et al. 2022). Existing ideologies and policies sensitize us to seeing how local configurations of universals, in this case AI training, can offer thought-provoking contrasts with developments elsewhere. Institutional and regulatory forces push back to bend any so-called universal logic. AI training is reconfigured to promote key features of the Finnish penal ideology: the minimization of harm, the normality principle, and the aim of adjusting prisoners to society.

Yet at the same time, data labor arrangements need to respond to the financial aims and aspirations of startup companies. In general, public sector collaboration with private companies raises questions about who is guiding whom, as the careful management of such partnerships to preserve public institutions and the delivery of public services is critical in terms of maintaining a solid relationship between the citizen and the state (Dencik and Kaun 2020). In this case, we would argue, the public–private partnership has managed to align interests in a manner that feels societally “right.” As we followed the project, all the parties spoke positively about it—unusual in our experience as AI initiatives in the Finnish public sector are often marked by tensions, with at least some participants evaluating ongoing projects critically (Lehtiniemi 2024). Here, however, critique and related cynicism were absent, making us think that, perhaps, when data labor enters prisons, it is possible to strike a satisfactory balance between economic and societal interests.

Local Data Needs

The startup company in the initial prison data labor arrangement, Vainu, takes its name from the Finnish word referring both to a hunch and the scent that an animal picks up when hunting. The name is suggestive: the company sells packaged information to customers seeking to improve business-to-business sales. Vainu retrieves data from available sources, including financial statements, articles in the business press, websites, and social media. These digital resources contain data about companies as natural language: unstructured, textual data that the company converts into structured formats, automatically and at scale. Yet human interpretative skills are needed to produce training data for algorithmic setups and AI models, and this need is ongoing. Vainu is not alone in building automated systems that rely on a continuous stream of human-annotated data. Humans doing machine-like work that machines cannot complete is vital to automation (Crawford 2021; Irani 2015; Mateescu and Elish 2019; Tubaro, Casilli, and Coville 2020). Besides annotating natural language, humans interpret sound, images, and video, and verify the outcomes of machine-learning systems (Crawford 2021; Gray and Suri 2019; Newlands 2021). This work may involve taking selfies to feed face recognition algorithms, categorizing conversation topics, or identifying parts of speech or emotions behind a statement. Human work powers AI assistants, ensures the usefulness of recommendations and search results, and removes offensive and illegal social media content (Roberts 2019). The multitude of tasks means that a developing market that caters to companies is increasingly varied (Tubaro, Casilli, and Coville 2020).

The demand for human assistance produces friction if it interferes with the seamlessly operating data machinery, and online labor platforms that match human resources with data needs have been developed to reduce it. The global data labor infrastructure decreases this friction by purposefully cutting human and social ties, as platform companies atomize and socially isolate workers from their peers (Wood et al. 2019). Companies could also hire data workers, but platforms enable them to access a workforce on demand, at a fraction of the cost of salaried staff. In other words, machine-learning calls for a flexible pool of disposable workers who can complete tasks irrespective of time and location (Tubaro, Casilli, and Coville 2020).

Platform workers mainly reside in places where even low pay is appealing, such as the US, Kenya, India, Nepal, Venezuela, and Brazil (Gray and Suri 2019; Posada 2022; Tubaro, Casilli, and Coville 2020). On MTurk, for instance, AI training is broken down into bits of work, called “human intelligence tasks” or HITs. Workers are paid per completed task. To give an example, in mid-2021 Vainu posted the following HIT on MTurk: the worker will receive 0.36 USD to “find and highlight all the organization, personal, location and job title names in a business news article.” Vainu's HIT was more complicated than some other available tasks, but it paid comparatively well.

To maintain efficiency, requesters can list new tasks through an interface that enables their automatic generation. In this scenario, workers are reconfigured as parts of a machine. The software queries human-generated data as if humans were just another piece of software, or its extension. MTurk's old tagline is suggestive: marketing itself as “artificial artificial intelligence,” the service proffered human workers as stand-ins for future pieces of software (Crawford 2021). The notion of “heteromation” (Ekbia and Nardi 2017), a neologism that stresses the human role in automation, becomes concrete in arrangements where the machine feeds the human with tasks that need to be completed. Since platform workers are an integral part of the data labor infrastructure that keeps automation up and running, they become interwoven with the global data-producing machinery.

As part of its activities, Vainu processes Finnish-language data about companies. To produce training data for this purpose, data workers need Finnish-language skills. The refining of models might also require nuanced place- and culture-specific interpretation. However, workers with the required language and cultural skills are not available on regular platforms like MTurk, at least not at the scale needed. Most Finnish speakers reside in Finland and are likely to find better compensated work in other fields. Furthermore, the welfare state provides social benefits to its residents, which means that unlike places like Venezuela (Posada 2022) where no such support exists, Finns are rarely so devoid of options that data work would be attractive. Companies could raise the pay to attract Finns to join MTurk and other labor platforms but, so far, the mechanisms of supply and demand have not worked this way, suggesting that the “global solution” to the need for dispensable human work fails in privileged settings.

The dearth of appropriate workers, then, maintains a locally felt friction that intervenes with automation aims. Since this friction discourages the production of localized AI services, it universalizes English language, as technology firms and computer scientists focus their initiatives on where training data is readily available. This threatens the survival of smaller languages in the AI world. On the other hand, and more importantly in our case, friction does not merely hinder development, it is also productive, as it spurs creative efforts that counter the universalizing trend, giving rise to experimentation with various schemes to boost local-language production of training data. For example, when the National Library of Finland digitized old newspapers in a cultural heritage project, humans were needed to interpret and correct words that optical character recognition could not handle (Chrons and Sundell 2011; Yle 2011). A game-like online interface was created for the purpose, in the hope of motivating Finnish speakers to exercise their interpretative skills by playing it. More recently, Yle, the national broadcaster, teamed up with a university to set up a system, complete with mobile apps, to which people could donate samples of everyday Finnish speech covering a range of topics. The aim was to create an openly available database for public and private actors developing speech recognition systems and voice-based interfaces (Lahjoita Puhetta n.d.). The unusual project to produce AI training data in prisons is yet another outcome of this friction. The project was motivated by Vainu's need for an efficient and cheap enough source of training data, yet by mitigating the problems caused by the absence of human resources the company was also engaging in cultural work to resist the trend of marginalizing smaller languages in the global AI scene.

Prisoners are No Ghosts

Most data labor remains publicly unacknowledged, but the Finnish arrangement has been exceptional enough to be covered by international media (Chen 2019; Meaker 2023; Peteranderl 2019). An article on the online magazine The Verge (Chen 2019) framed the data labor arrangement by juxtaposing two potential outcomes: emancipation or exploitation. Quoting Lilly Irani, an established scholar of work in global digitally mediated economies, the article pointed out how the hype that surrounds AI “can masquerade really old forms of labor exploitation as ‘reforming prisons.’” Wired magazine (Meaker 2023) raised similar concerns, noting that exploitation at the margins of society is part and parcel of how technology companies typically operate.

Data work on platforms is “organized to foster alienation” (Tsing 2019, 150). Workers, defined as contractors, perform tasks on a platform that deprives them of the stability, benefits, and social support that regular employment would provide (Tubaro and Casilli 2022). Tasks are distributed among workers by algorithmic means, which can create “digital cages” that treat workers arbitrarily, isolating and disenfranchising them (Vallas and Schor 2020). As an intermediary between companies and workers, MTurk is a case in point (Irani 2015). Task requesters—or rather, their automated software programs—send HITs to MTurk, which lists them for workers’ selection. This disconnects individual workers from requesters in a very concrete manner: they are represented by ID numbers or codes, not identifiable by name or location. Rather than specifying who does the work, a requester specifies eligibility criteria, and the internal workings of MTurk take care of the rest. In this process, workers are converted into unidentifiable and interchangeable machine extensions that need to compete for tasks against nameless peers, who only become visible on the platform through changes in task lists. They must identify the most lucrative tasks before others, while maintaining performance metrics that make them eligible for the more desirable tasks.

As suggested above, turning workers into disposable parts of the data-extracting machine is a response to the friction produced by the data needs of the global economy. Gray and Suri (2019) label the work done on or through online platforms as “ghost work,” a high-tech version of more familiar types of labor that “disappear from our observations and reckonings” (Daniels 1987, 403), such as domestic gendered work that remains outside of the purview of labor relations and markets. Ghost work, on the other hand, disappears because it is hidden by way of its technical organization.

Since technology is supposed to work effortlessly, companies are incentivized to downplay and obscure the human help needed. Gemma Newlands (2021) frames the invisibility of data work as the strategic concealment of human involvement and, indeed, leaders and employees of AI companies are typically unwilling or unable to discuss their use of human workers (Tubaro, Casilli, and Coville 2020, 4). Platforms organize work in ways that separate or disembed workers from customary labor markets (Lehdonvirta 2022; Posada 2022). This lack of ties to regular employment means that workers need to seek alternative support systems to reduce the uncertainties, risks, and hardships involved in data work. In the absence of unions and workplaces, people adjust to platform requirements by organizing collectively using social media and collaborating to get work done (Wood, Lehdonvirta, and Graham 2018). They restore the sociality that has been denied them. In Venezuela, for instance, annotating data for AI is supported socially and economically by family and household members and fellow workers in social media groups (Posada 2022).

Incarcerated people tend to be invisible in society, and the prison walls provide an even more opaque buffer than platformed labor arrangements. At first sight, when prisoners perform data work, universals are in active play, contributing to the disappearance of human input in AI. From Vainu's point of view, arrangements with the prison are like those with MTurk, as the visibility of labor arrangements is lost when tasks reach prison walls. As Tuomas Rasila, the company's co-founder, explains, “From our perspective we’re dealing with a workforce we don’t know. So we have to assume it includes unskilled workers, and we have to create processes to ascertain the quality of their output.” Companies using platforms like MTurk have developed automated quality control measures. Cross-validation, for example, can be employed to detect random clicking patterns associated with what Vainu's representatives called “unmotivated” workers. This problem, we were told, is in fact much more severe in MTurk than in prisons. Nevertheless, the same validation procedures are at work in both cases.

Vainu also needs to rely on an intermediary to distribute tasks to workers in both MTurk and the prisons. As prisoners are represented by pseudonymous IDs, the company has no control over, or knowledge of who does the work or is logged in with a particular ID. Rasila states that “they stay in their role,” meaning that they do not interfere with what goes on in prisons: “We’re a tech company collaborating with RISE, and we’ve expressed our willingness to pay for each completed task. We don’t have an opinion on how to run prison administration, and we don’t want to have one. We are assuming they’re competent in organizing prison labor.”

Despite similarities, it would be a mistake to push the analogy between ghost work and data work in Finnish prisons further. On the usual platforms, workers organize their own work, in principle autonomously yet often ending up working long and irregular hours at nights or in isolation (Gray and Suri 2019; Wood et al. 2019). Platformed arrangements for mediating tasks and workers are designed to function without human intervention or attention to particularities in the everyday situation of an individual worker. The arrangement in Finnish prisons is fundamentally different: prisoners are required to act and do things that are valued by the prison system. Within the prison walls, they cannot disappear into ghosthood, even if they wanted to. They are expected to participate in “prison activities,” which consist of work, courses, training, and various rehabilitation programs. The prison controls daily schedules and prisoners’ movements, and when it is time, they work under the watchful eye of prison officers in spaces allocated for the purposes of work, often in the ward's common area. Instead of machines or faceless market arrangements, civil servants, prison officials, and counselors participate in matching prisoners with data labor tasks. This human involvement in task allocation introduces its own kind of friction, as the prison system needs to ensure that someone represented by the pseudonym is there to carry out the task. The aim, as expressed and codified in penal policies, is not to squeeze economic value out of the prison, but to offer activities that could aid in adjusting incarcerated people to society. Typically, prisoners do not work more than a few hours at a time, and this is also the case with AI training. Unlike ghost workers whose compensation is task-based, prisoners receive a small daily allowance that does not depend on the number of tasks completed (see Lehtiniemi and Ruckenstein 2022).

All this suggests that while the demand for frictionless human input in automation tends to reduce humans to nameless and disembedded resources—essentially turning them into ghosts—Finnish prisoners involved in data work do not undergo a similar transformation. As data workers, they are structurally better positioned to receive individual and attentive treatment than workers in the Global South. Thus, the lens of friction allows us to identify developments that suggest different kinds of data-related outcomes. Whereas the friction-reducing infrastructural form of data labor obscures, anonymizes, and renders uniform the human inputs to automation, the friction caused by human involvement in automation can also create locally visible, non-anonymized alternatives.

Aspirational Data Labor

Novel technologies are entangled with future aspirations as they “move forward with hope, as part of everyday anticipation” (Pink 2022, 44). Hope, anticipation, and aspiration were also in the realization of the prison pilot. The project was set in motion by Rasila, one of Vainu's founders, who understood the need to develop new arrangements to do AI training in Finnish. In line with the friction approach, appealing invitations to become part of global developments, saturated with hope and anticipation, make people act. Rasila, exemplifying personal aspirations to innovate and strengthen processes of datafication, saw unusual potential in prisons and thought that if data work were done in prisons, opposites would meet to produce a highly thought-provoking dynamic. He imagined the prison's “unwanted, low-end workforce,” usually associated with menial work, performing high-tech tasks in an environment defined by the most severe digital deprivation imaginable in a Nordic country. Doing data work, he believed, would lead to individually felt empowerment. By training a machine to perform better, those at the margins of society could participate in permanently improving something, in making technology move forward. As Ghassan Hage (2009) suggests, hope is nourished by an “imaginary mobility,” a sense that things are “going somewhere.”

For a startup entrepreneur, the prison pilot is also a proof of concept that probes the frontier of technological scalability (Tsing 2019). It examines whether it is possible to build a pipeline ending in a public institution, engaging people in data work and automatically auditing the quality. Rasila and colleagues semi-seriously entertained the possibility of turning the piloted arrangement into a general-purpose platform that would mediate data work. With such a platform, Finns could do data work when sitting on a bus or standing in a queue, or when they are ill or retired—basically, whenever they had extra time in their hands. Vainu, however, was not in the business of platform creation, and Rasila made it clear that the company positioned itself as a platform user. Nevertheless, the pilot proved the point. If data work could be done in prisons, it could be done anywhere. The prison could operate as an accelerator of automation processes, a test bed for technological development that could later be transitioned from prisons to other domains of application (Kaun and Stiernstedt 2021).

Whereas the tech entrepreneur's aspirations are inspired by universally appealing AI futures, advocates of the pilot project within the prison system entertain more qualified hopes for data work—hopes related to rehabilitation and adjustment to society with aims to update citizen skills for the digital age. Importantly, however, there was no tension between the tech entrepreneur's and the prison system's aspirations, suggesting a fertile ground for collaboration. Despite attempts to build “smart” facilities, prisons have generally remained digitally deprived environments, further framing incarcerated individuals as a disadvantaged group (Reisdorf and Jewkes 2016). Although some Finnish prisons have cell terminals and permits can be granted for limited whitelist-based internet access upon request, computers are not a part of daily prison life. For civil servants, the normality principle in the Finnish penal code states that conditions in prison should be as close as possible to those in society more broadly. For this reason, prisons cannot remain analog islands within an increasingly digitalized world. Data work presents an opportunity to place prisoners in front of computers, reconfiguring them as data producers and aspiring participants of a future AI-powered society.

The temporal element is central to aspirations, in that the value and rewards are expected to materialize in the future. Kuehn and Corrigan (2013) coined the term “hope labor” to describe voluntary, online social production practices undertaken in the hope that the experience and exposure they provide will lead to future employment opportunities. Similarly, Brooke Erin Duffy (2017) suggests that “aspirational labor” defines activities that have the goal of bringing future rewards, including social and professional prospects. Prison officials similarly envision future value for data work; it is imagined to be beneficial because it supports the eventual resettlement of the incarcerated.

RISE officials’ positive attitude to data labor is further reinforced by the increasing difficulty of finding suitable work for prisons. The trend is international, as small-scale industrial production has been largely eradicated in the context of global trends of automation and outsourcing (Kaun and Stiernstedt 2023, 33). Emerging AI technologies are seen to create conditions for feasible work within prison settings. The human work that keeps AI systems running is often considered repetitive, to the extent of being rote and menial. Yet, as we were repeatedly told, success in a closed prison environment comes in tiny steps. Any promise of progress or improvement in the lives of prisoners and the overall prison system is welcomed. Prison counselors view data labor as more cognitively rewarding than current forms of prison work—packaging electrical components or small screws in boxes, or folding laundered towels and socks—which, as one of the counselors expressed, allows those doing the work to “check their brain at the door.” Although data work might be routine, it still introduces knowledge-oriented tasks, such as reading and annotating texts, into the prison setting.

A repeated feature of discussions with RISE officials about the prison pilot was a slippage from AI training, to the broader context of prisons in the digitalizing society. AI training appeared to belong to an imaginary universe of technological advancements along with smarter prisons, VR headsets and chatbots experimentally used in rehabilitation, cell terminals hooked to the internet, and even the prison's new information system. In civil servant talk, these high-tech advancements contrast sharply with the prevailing image of a prison system unable to react to societal change. Data work mobilizes the hope of movement and progress in the digitally deprived prison system. In light of optimistic stories associated with AI training, it is not only the most marginal prisoners who appear a little less marginal, but the whole prison system. The hope is that Finnish prisons will finally become “modern”—a term used by civil servants—and that prisoners, with little to expect from life in terms of work, will have slightly better opportunities to become functioning participants in society.

Friction in AI Training

The imagined future returns of work are not necessarily the prisoners’ own aspirations. Counselors suggest data work to those they consider could benefit from it. Participation is voluntary and, as many prisoners think they are unsuited for AI training, they may decline upon hearing the details of the work. In this instance, the friction in data labor is between the prisoner and the AI training. People who end up incarcerated in closed prisons often suffer from difficulties with learning and concentration, and they simply cannot do the job. The pool of data laborers in Finnish prisons is therefore limited. Both prisoners and counselors told us about those who had tried the work but were distracting others or simply clicking away randomly. For a group of youngsters who had been considered a good fit based on computer skills, concentrating on the work turned out to be impossible. The work was then tried with a group of seniors, who fared better even though some had almost no prior experience with computers.

A prison counselor who oversaw female prisoners’ data work described observing them sitting silently with laptops, concentrating on completing tasks. At times, they would discuss how to solve a problem. What caught the counselor's attention was that prisoners were collaboratively engaged. The same observation was later confirmed by our own fieldwork. While data work is in a sense solitary, performed individually on a computer, undertaking problem-solving tasks in a shared space involves social interaction. For the counselor this is a highly positive outcome, as collaboration cannot be taken for granted in the prison environment. Here, data labor appears to reduce friction, as it helps smooth out prison relations in ways that are beneficial for rehabilitation. We also repeatedly heard an anecdote about an elderly female with no experience with computers. After being encouraged to try data work, the woman said she enjoyed reading the business news items it involved, as she had never encountered such texts before. She also liked being able to say that she had trained AI while serving time. The retelling of this anecdote accentuates the sense of possibility associated with data work. In interviews, prisoners confirmed being exposed to new kinds of textual material and finding this beneficial or somehow intriguing. Technologies in the prison are typically apparatuses of control and surveillance, but AI training is associated with a curious form of techno-hope. With new digital tools and computer skills, and a sense of advancement, prisoners seem to be a little less stuck in the closed environment.

Still, most prisoners consider data work to be simply work. Rehabilitative aims and prospective AI futures feel distant, but those with whom we spoke appreciate the opportunity to break the tedium of cell-centered life. Generally, the aspirations of prisoners are constrained and focused on surviving day-to-day. Nevertheless, many found data work meaningful and believed that more willing workers could be found in other wards. While the work is repetitive, it is, as one prisoner put it, “good to use the brain under these circumstances.”

While some were quite knowledgeable about what it meant to produce data for training AI, for most, the exact purpose of their work remained a mystery. Nevertheless, everyone appreciated the fact that their work was related to AI. Even those who lacked comprehension of the broader context and the use of the data they were producing reported performing their tasks to the best of their abilities, as random clicking would obviously be useless and render their effort meaningless. Some expressed that they derived purpose from data work because it is outsourced to the prison by a company that relies on prisoners’ contributions. They shared a sense of responsibility to do the work properly, as they valued its existence, and did not want to discourage the company from collaborating with the prison. Thus, the prisoners appeared to be aware of their role in frictionless data production.

Growing and Cutting Weeds

The imaginary mobility associated with the prison data labor pilot, and the sense of people moving forward alongside technologies, solidified the project's success. Vainu's representatives were pleased with the quality of the work, which was at least on a par with, or even better than the results expected from MTurk. Halfway through 2020, however, Vainu's new management team, with no first-hand experience of the pilot, abruptly ended the project. RISE tried to convince Vainu to continue the collaboration, but without success. In the end, data labor in prisons was a “rönsy’—a stolon or runner—a Finnish way of saying that the project was an outgrowth of the core business that needed trimming away. In Tsing's (2019) vocabulary, the prison pilot might be considered “a weed” that appears superfluous, something that results from a disturbance and, in turn, disturbs the overall growth of well-organized operations. A streamlined startup company has no time to cultivate weeds.

Our informants were disappointed that a pilot that appeared to be a win-win scheme, supporting both the prison system's rehabilitative aims and the company's need for data, was terminated. The desire to hold on to the project suggested to us that the homegrown pilot had successfully aligned various interests. The friction lens helped us to disentangle institutional practices and aspirational orientations, revealing hidden and less obvious dynamics. The alignment of parties involved in the pilot had required the initiative, effort, and resourcefulness of the tech entrepreneur and his colleagues, who practically set up the AI training pipeline. The pilot further benefited from the receptive and supportive attitude of civil servants who secured the collaboration, and the prison counselors who oversaw the new type of work. Prisoners were also onboard, with enough of them volunteering to engage in data labor to make the initiative a viable concern.

In terms of local automation, projects that are regarded as weeds that “shout challenges to stability” by growing human ties rather than cutting them (Tsing 2019, 34), the prison pilot is suggestive. The failure of Vainu's management to appreciate the collaboration implied that they did not recognize how the project could be more than just a mere outgrowth. Companies that need data labor tend to appreciate a streamlined form of data labor that actively reduces friction and ties humans and machines into seamless loops. MTurk and other companies operating on a similar logic aim to turn human workers into isolated, machine-like parts of automation processes, with the workforce unable to interfere with the logic of the platform. From the platform's perspective, human diversity has no value; it does not, and it cannot, matter which ID code, and where, sits in front of the computer. In fact, diversity can be seen as so detrimental to platform operations that social ties need to be actively, even aggressively, cut.

The desire to keep their automation operations frictionless means that companies have little interest in more idiosyncratic projects that foster local or personal aspirations. The aim is to erase human and social involvements that might bring messiness and uncertainty into the picture. Instead of diversity, the tech world tends to nurture “a one size fits all” imaginary. This, however, is not a product of technology as such, but of the way technology is envisioned and promoted (Seaver 2021). The desire for frictionless automated solutions with as little human intervention as possible, blocks potential AI futures. Technology companies follow the mainstream notion of what automation does; they aim to operate in disembedded rather than locally embedded ways, while simultaneously trying to replicate and mimic what others have done before in an isomorphic way (Caplan and boyd 2018).

The technology industry claims to nurture innovation and promises disruption, but our case serves as a reminder that they often follow narrowly defined paths. Considering the association of technology with an open future, this kind of imaginary “stuckedness” (Hage 2009) cuts down novel developments. With such predefined aims, it is impossible to commit to weeds—like the prison pilot—even though these projects have potential, in that responding to a locally felt friction could become the source of a new kind of value.

Ways Forward

The methodological lens that we have developed to identify responses to friction and what they promote, challenges data labor as a uniform future development. It reveals how personal aspirations, institutional context, and power dynamics operate in complex and even contradictory directions. The infrastructural form that supports global data labor arrangements reduces friction associated with operations that need masses of human-generated data. Nevertheless, diversity in data labor is possible within particular social, political, cultural, and economic contexts, exposing other kinds of aspirations and tensions. AI training conducted in Finnish prisons is deeply entrenched in promissory data-driven futures and in existing inequalities. The whole project was launched because non-incarcerated Finns can sustain themselves without resorting to this kind of work. At the same time, however, the project is an outcome of aspirations to alleviate inequalities by providing incarcerated individuals with better prison conditions and more hopeful future horizons.

The case of prisoners training AI demonstrates how data power critiques can distort the perspective to the degree that we lose sight of important details. Finnish developments become lumped together with other data labor arrangements without attention being paid to prevailing penal philosophies and practices, or to the values, goals, and aspirations that move the people involved. The friction lens enables separating specific forms of data labor and the various human entanglements in play. Data labor arrangements can treat humans as mere extensions of machines and transform them into ghosts, yet the friction approach can also identify other kinds of human involvements and aspirations. This turns friction into a future-oriented concept. At “the edge of the future” (Pink 2022), the friction lens can open multiple views onto how things might go forward. The emphasis on the aspirational does not merely add nuance to the current conversation but serves as a broader argument for the role of datafication-related technologies in society.

Current appeals for AI ethics and regulation reflect a desire for the path to an AI future to be properly managed, with warnings of risk and bias along the way. Nevertheless, it takes considerable effort to stabilize AI as an object of regulation (Ruckenstein 2024). The friction approach does not shy away from instability and uncertainty. Rather, it aids in exploring the ambivalences and contradictions that reveal the shakiness of current ethical and regulatory approaches. Mainstream approaches to AI futures tend to overlook that the generation of datasets promotes disembedded forms of labor. While technology enthusiasts might be convinced that ultimately data work will be fully automated, the growing and shifting demands for data will keep the need for human data work high in the foreseeable future. Labor platforms attract a human workforce with few other options, resulting in a dispersed global underclass of ghost workers operating in the invisible backstage of AI-assisted society. Most low-paid workers producing the datasets used to train AI models reside in the Global South. This generates profits in the wealthier regions and actively promotes the transfer of value from South to North (Casilli 2017). Such a dynamic suggests a distinctive form of data colonialism that normalizes data-related extractivism. However, this aspect of data work has not been adequately addressed in AI ethics debates.

With its institutional and human arrangements, the Finnish prison system is far from a streamlined data production operation, and given the prevailing penal philosophies and regulations, it is unlikely to become one. Ultimately, this is probably the reason why the collaboration between Vainu and RISE ended. Rather than harnessing “neutral” technology to outsource work to prisons, extending the collaboration would have required ongoing human dealings with the prison system. While for Vainu the data labor arrangement was ultimately an outgrowth, the friction produced by the dominant data labor infrastructure has not disappeared. The collaboration between Metroc and RISE shows that there is still a shortage of Finnish-language data workers. This friction remains productive and continues to sustain the new data labor arrangement.

In our continued engagement with locally sourced data labor, the friction lens aids us in tracing how interests are balanced, and how diverse aspirations and social relations grow or wither away. The Finnish arrangements serve as a testament to the possibilities of building AI initiatives that do not treat data workers as mere extensions of machines. Of course, this does not mean that data labor, in general, is without downsides. AI developments continue to be ethically and politically problematic. However, the focus on friction helps distinguish these developments by drawing attention to the human involvements, imaginaries, and aspirations that promote a range of data-related futures.