Fostering the data welfare state: A Nordic perspective on datafication

A relatively new Danish labour and unemployment law [Beskæftigelsesindsats], passed in 2019, involves an ADM system that uses data profiling to assist state social workers in their efforts to find work for unemployed citizens. Although the system is not explicitly labelled ADM or AI, but rather “a national digital tool for clarification and dialogue” (Retsinformation, 2019: §8.2), the tool nevertheless represents an attempt to digitally assess job-seekers and predict their success on the job market. The tool is based on an ADM pilot project that ran from 2016–2018 (Motzfeldt, 2019). While it is not clear whether the current tool is identical to that of the pilot project, or whether it represents a new and updated version, scholars agree that the pilot project served as the primary source for this new “ADM unemployment system” (Andersen, 2019; Motzfeldt, 2019).

The goal of the pilot project was to identify unemployed citizens at risk of long-term unemployment via a “profile clarification tool” [afklaringsværktøj]. This tool operated in two stages: In the first stage, unemployed citizens filled out a survey indicating their own evaluation of their situation – including their expectations of when they would find employment, as well as descriptions of their personal situation – in case this was relevant to their search for work. In the second stage, ADM was used to evaluate each job-seeker based on a number of predefined categories, gathered from different institutions and public platforms (relating to, e.g., education, work experience, age, ethnicity, and welfare benefit history). These categories were defined as “objective” and “observable”, while participants’ own evaluations were considered “subjective” (The Danish Parliament: 212; Mploy, 2018: 7). The combination of data across both stages generated a score, indicating whether the job-seeker was at risk of long-term unemployment. According to this score, the job-seeker was assigned services from a local job centre.

In her book Automating Inequality, Eubanks (2017) discusses the consequences of grounding social welfare policies on data prediction, on the basis of her analyses of ADM-driven social services in the US. Echoing van Dijck's (2014) conceptualisation of dataism, Eubanks criticises the assumed relationship between large population data and individual behaviour, and she describes how algorithms aimed at predicting individual needs are designed according to social group. As a result, scores used to predict individual needs are not based on individuals and their personal histories, but rather on the data mining of (pre-defined) social groups related to ethnicity, gender, civil status, neigh-bourhood, and other demographic factors. In other words, individuals are not impacted by their own actions or personal situations, but by the previous actions of “members” of their “categorical belonging”. While Danish law defines social categories (e.g., age, ethnicity, citizenship status, education, etc.) as “objective” (The Danish Parliament, 2019), Eubanks demonstrates how the history of poverty and ideas of marginalised groups as less valuable citizens play into the design of algorithm-based social policies; this is especially clear in the design of the social categories used to predict individual risks or needs, and thus the level of assistance granted by the state. She argues that, rather than facilitating new and neutral social policies, “digital tools are embedded in old systems of power and privileges” (Eubanks, 2017: 178). Such embeddedness might explain the mixed results of the Danish unemployment services pilot project: Of the sixteen municipalities that participated, nine found the ADM screening system to be effective in identifying those at risk of long-term unemployment (Mploy, 2018); the remaining seven municipalities experienced the ADM system as stigmatising, and thus undesirable for facilitating dialogue with unemployed citizens (Mploy, 2018: 32). These ambiguous results indicate a challenge in fulfilling the first pillar (justice and non-bias in processes of datafication) in the data welfare state.

All of the participating municipalities in the pilot project concluded that the ADM tool could not determine the risk of long-term unemployment independently; rather, output from the tool needed to be interpreted alongside social workers’ professional judgements. In other words, in order to maintain a transparent and sustainable data welfare state (the fourth pillar), as well as a just and non-biased data welfare state (the first pillar), the ADM tool needs to be continuously combined with professional judgements provided by “real people”. Importantly, some job centres pointed to situations in which the ADM disagreed with social workers’ evaluations (Mploy, 2018). In these cases, the social workers problematised that they could not locate the basis for the discrepancy; that is to say, they could not see the calculation behind the ADM algorithm. This points to the “black box” of ADM programming, whereby users do not understand or see the programming that feeds into a particular score or prediction. While we, as consumers, are used to ADM recommendations that do not fit our desires (e.g., suggestions from streaming services that do not capture our interest) (Motzfeldt, 2019), the same “mal matching” can have severe consequences when incorporated into social services. As consumers, we can choose to not accept ADM recommendations; but as citizens in a social system, we have fewer options to reject ADM-determined suggestions or solutions. The lack of transparency and sustainability (the fourth pillar) in the data welfare state is therefore much more critical when considering the determination of welfare provisions than it is in the provision of entertainment options.

Another reason why the ADM prediction differed from social workers’ evaluations in some instances might be that, in line with the underlying assumptions of dataism, the ADM tool equated correlations with causality (Cheney-Lippold, 2017). In the pilot project, knowledge about unemployed citizens was reduced to correlations, whereas knowledge held by social workers was based on an observed understanding of causal relationships between various social categories (Antczak & Birkholm, 2019). Accordingly, we argue that the term “artificial intelligence” should be replaced with that of “automated decision-making”. While AI and machine learning have become the assumed hallmarks for a prosperous future, we question the assumed intelligence of datafication and stress that AI is best characterised as a future imaginary.

Motzfeldt (2019) warns that the use of ADM tools in the welfare state will most likely be expensive. While such systems are often initiated to save time and replace (salary demanding) human staff, the pilot project described above underlines that ADM cannot be a complete substitute for human staff, and it is therefore likely to add cost on top of existing salaries. Importantly, experiences with ADM-driven social services point to how professional staff tend to prioritise the ADM-score instead of their own judgement (Antczak & Birkholm, 2019; Eubanks, 2017; see also Oak, 2015); this risk seems heightened in contexts where resources are few, and there is little time to determine the reason behind the discrepancy between an ADM-score and one's own evaluation.

Public service media (PSM) represent an important component of the welfare state. PSM are largely financed through public funding, with the remit to serve the public interest. They follow and express the ideals of the welfare state – namely equality and universality – in terms of access to information and representation. Changes in media logics related to datafication are aptly captured by Andrejevic (2020), who describes a shift from mass media to automated media. Automated media operate through ADM, platforms, and utilisation of data, with a variety of consequences for mutual recognition, collective deliberation, and judgment. At their core, automated media infrastructures based on datafication propel and shape media logics and content.

The datafication of media can be understood as a “political economic regime”, whereby the accumulation and analysis of data determines new ways of doing business and governance (Sadowski, 2019). This has several consequences for media operation, including the “platformisation” (Gillespie, 2010; Helmond, 2015) of the media environment, dominated by commercial platforms such as Google, YouTube, Twitter, Facebook, and Instagram. According to van Dijck and colleagues (2018), a platform is fuelled by data, automated and organised by algorithms and interfaces, formalised through ownership, and governed through user agreements. Platformisation is but one example of how communication and sociality have become formatted to enable datafication. As Andrejevic (2020) argues, automated systems promise to augment or displace the human role in communication, information processing, and decision-making. The implication of this is that mental labour, thought processes, evaluation, and judgement must be standardised and formatted, similar to how physical labour in factories was replaced by machines.

As a result of their datafied infrastructure, media now operate in drastically different ways, using detailed audience data to customise services. While data-driven media promise more accurate services and efficient operations, they are also highly problematic in terms of the normalisation and capitalisation of surveillance and the de-skilling of comprehension (Turow, 2011; Zuboff, 2019). Additionally, the datafication of media endangers a fundamental role of media in society: to foster solidarity and civic mindedness (Andrejevic, 2020; Couldry, 2012; Nikunen, 2019). In what follows, we discuss how the datafication of the media environment affects the principles and pillars of the Nordic data welfare state.

Nordic PSM have traditionally been a strong and important part of the so-called media welfare state (Syvertsen et al., 2014). Nordic PSM prided themselves on their early move towards digitalisation; however, they now find themselves surrounded by commercial platforms owned by tech companies. Though sharing PSM content on social media platforms appears necessary to reach audiences – particularly young audiences (Andersen & Sundet, 2019) – Nordic PSM reluctantly do so, as a side effect of this sharing is the involuntary support of commercial platforms, which blurs the boundaries of the public and private sector and constitutes a threat to the second pillar (decommodification). For this reason, and to maintain ownership and control, some Nordic PSM have adapted to datafication by creating their own platforms, collecting data, and using customised services (Hokka, 2018; Moe, 2013).

Most Nordic PSM employ audience metrics, user profiling, or algorithms to create content, customise distribution channels, and generate recommendation systems. In line with their early adoption of digital technologies, YLE (Finland), DR (Denmark), SVT (Sweden), and NRK (Norway) considered data-driven personalisation services highly relevant to their strategy (Andersson Schwarz, 2016; Van den Bulck & Moe, 2018). But how does data-driven personalisation affect the non-bias of media content, in other words, the universal principle of PSM? PSM justify the use of data-driven customised services as a “new universality”. However, personalised recommendation systems may also erode universalism and the idea of a shared public sphere (the first pillar of the media welfare state) by categorising and profiling audiences. In particular, recommendation algorithms allow audiences to filter out content and viewpoints they are not interested in, thereby creating “filter bubbles” (Pariser, 2011). The resulting customised – and therefore fragmented – media environment erodes the experience of an imagined community (Andersson, 1983), which is central to solidarity, trust, and civic mindedness. In other words, the infrastructures and architectures of digital media emphasise individual tastes and personalised brands over collectiveness and community (the first and third pillars of the media welfare state).

To counter these polarising tendencies, PSM have integrated “public service algorithms” into their recommendation processes (Beckett, 2020; Bennett, 2018, Nikunen & Hokka, 2020). For example, YLE uses algorithms that seek to diversify, rather than polarise, media use (Nikunen & Hokka, 2020).

There are some implications that data-driven personalisation can also be used to better foster diversity of audiences. Through personalisation, PSM have been able to recognise new audiences (Hokka, 2018) instead of addressing the abstract average citizen, who is unlikely to belong to a minority; however, data-driven personalisation may create new forms of marginalisation and groups of discrimination (Mann & Matzner, 2019). Commercial platforms such as Netflix have found a global niche in audiences who are interested in content produced by or representing ethnic and racial minority groups as well as gendered and sexual minorities. Nationally, these audiences may be small, when considered globally, they are substantial. While catering to these audiences is important, the fact that it is driven by commercial interests, rather than by public values, tends to put emphasis on sensational and trendy elements that further marginalise those already in the margins (Saha, 2018). These challenges exemplify the entangled nature of the first and the third pillars of the data welfare state in the context of media, where the question remains of how to ensure non-bias in data-driven systems and take into account the different needs and aspects of diverse society.

Andrejevic (2020: 49) suggests that, instead of focusing on how algorithms affect content discovery on media platforms, we should attend to the ways in which the “combination of platform logics and communicative practices with broader social policies undermines the conditions for democratic deliberation” (see also Baum & Groeling, 2008; Campbell, 2018; Pariser, 2011). For many, such a consideration culminates in a vision of the public sphere guided by commercial profit rather than public values. Indeed, datafication has intensified the entanglement of public and private media infrastructures, which takes us to the data welfare state's second pillar of decommodification: The domination of global platforms has given rise to a media ecosystem in which virtually all media platforms are dependent on the infrastructural services of global tech giants (van Dijck et al., 2018). As this applies to PSM, the entanglement with data-driven media platforms risks undermining the public's trust. They also constitute a risk to sustainability (the fourth pillar) – as management and responsibility for the platform are outside the control of individual countries – as well as data diversity (the third pillar), as the data (and control with data) is gathered on very few platforms. While many European PSM allow for some private funding and advertising on their websites, Nordic PSM have tried to maintain independence (Sørensen & Van den Bulck, 2020). As discussed, some have developed their own platforms (YLE Areena in Finland, SVT Play in Sweden, NRK TV in Norway, and DR TV in Denmark); however, to be discovered by potential users, they must also appear on set-top box interfaces and third-party software (e.g., Apple TV, PlayStation, ElisaViihde [Finland], Strim [Norway], Tv Hub [Sweden], and YouSee [Denmark]). Whenever PSM content and platforms become embedded in these interfaces, they must submit to their hosts’ datafied logics, marketing, and commercialisation; this challenges the second pillar (decommodification) of the data welfare state.

As very little is known about the data traffic in these contexts, transparency (the fourth pillar) is seriously challenged. Moreover, some Finnish, Swedish, and Danish PSMs are connected to third-party servers that collect user data for advertising or data management purposes (Sørensen & Van den Bulck, 2020). While there is no evidence that audience data is gathered by these servers, it is nonetheless clear that PSM are significantly integrated into digital business networks, and it is highly problematic for public trust if PSM user data is tracked and sold to third parties. In addition, data gathering on PSM remains modest compared with that of the tech giants; thus, PSM must seek new collaborations to access more data. These collaborations with commercial data-driven platforms further blur the line between public and private media, directly affecting the second pillar (decommodification) of the data welfare state.

In Sweden, public service radio has integrated the commercial music streaming service Spotify – a phenomenon referred to as the “Spotification” of public service media (Burkart & Leijonhufvud, 2019). Furthermore, some PSMs (e.g., NRK) co-produce content with Netflix (Sundet, 2017). While datafication is embraced in PSM news services as an efficient way to reach and serve audiences, it is simultaneously seen as highly problematic for the PSM principles of universalism and equality (the first pillar). Growing dependence and entanglement with commercial digital platforms generates uncertainty over data practices and trust. In recent years, Nordic PSM have paid more attention to the transparency of their data practices and the independence of their platforms. This reflects their strong investment in data-driven technologies and the challenges that these systems entail for PSM. As illustrated by the ADM unemployment services case, algorithmic operations represent trade secrets, and data collection practices are often hidden from the public. To serve public values, PSM must maintain transparency (the fourth pillar) in their operations and ethics in regards to their data practices.

It seems PSM experiences various challenges in upholding the four pillars of the data welfare state. Particularly the second pillar of decommodification appears to be difficult to foster in the current platformed media ecosystem. The effects of blurring the boundaries between the public and private sector are illustrated in the ways in which PSM have adopted the logics of commercial platforms with practices that challenge non-bias and diversity principles (the first and third pillars), as well as undermine sustainability and transparency (the fourth pillar) with increasing dependency on powerful platforms.

In the Nordic setting, corrections are part of the public sector – aiming to rehabilitate and resocialise individuals – and are often viewed as Scandinavian exceptionalism (Pratt, 2008; Pratt & Eriksson, 2012). Similar to other areas of the welfare state, the corrections sector is increasingly implementing digital technology in order to become more efficient, particularly with respect to decision-making. In the Swedish context, the aim to “smartify” corrections – in part, through datafication – has materialised in the Krim:Tech initiative, which the Swedish Prison and Probation Service [Kriminalvården] launched in 2018. The main aim of Krim:Tech is to gather and recruit technology developers to renew and digitalise work with incarcerated individuals (Kaun & Stiernstedt, 2020). The Swedish Prison and Probation Service describes this in the following terms:

Krim:Tech is the new digitalisation initiative by The Swedish Prison and Probation Service. With the help of the latest technology and research, the initiative will support the development of new and improved digital solutions within the authority. Krim:Tech is an inventor's workshop and test bed for digital technology. Does an ankle monitor actually have to be an ankle monitor or could it be something else instead? How can we use IT to keep our security class 1 facilities calm? How can we prevent children and families who are visiting their father or mother in the prison from becoming afraid? Can we do security scans with a toy instead of metal detectors and full body scanners?

This description of Krim:Tech reinforces the idea of renewing the entire organisation – including incarcerated individuals – with the help of smart data-based technology, while simultaneously considering the prison context as a test bed for new technologies (Kaun & Stiernstedt, 2021). This idea of renewal and even reinvention is emphasised in an unpublished policy document, shared with the authors, which carves out a digital agenda for The Swedish Prison and Probation Service. According to this agenda, there are five ambitions for smart technology within the sector: leveraging digital resources to overcome social isolation, preparing prisoners for life in the digital society, increasing efficiency through digital resources, preempting recidivism through digital resources, and striking an appropriate balance between security and the use of digital resources.

Besides the larger visions expressed by Krim:Tech and the digital agenda, there are also specific projects and implementations of digital technology in The Swedish Prison and Probation Service that illustrate particular aspects of datafication. One example, which is used to support probation services, is an application called Utsikt [View, Prospect, or Outlook] (Kaun & Stiernstedt, 2020). This tool, representing the first of its kind, was developed by Krim:Tech in 2015, with support from the state agency for innovation (Vinnova) (The Swedish Prison and Probation Services, 2015). During a trial period in early 2017, a group of 19 individuals on probation tested the application and provided feedback for improved functionality (The Swedish Prison and Probation Services, 2017). According to the project leader, Lena Lundholm, the application was designed to improve attendance at probation meetings and provide clients with preemptive exercises derived from cognitive training in challenging situations. Accordingly, the application includes breathing exercises for stressful encounters and links to hotlines for support during critical episodes. Furthermore, the tool enables users to track their moods and provides them with different scenarios for problem-solving.

Descriptions of the application emphasise that it is merely meant to prevent recidivism and not to control or supervise clients. Use of the application, which is available as a free download, is voluntary, but it is thought to complement the work of probation officers. The application requires an iPhone 5 or Android version 5 (or later operating system), as well as an App Store or Google Play account. These preconditions are potentially challenging for some clients, especially those who have recently been released from long prison sentences (Jewkes & Reisdorf, 2016); this lack of access potentially challenges the first pillar of justice and non-bias in the processes of datafication.

Together with a probation officer, users input data on their rehabilitation process and support, as well as control measures, into the application. All data stored in the application are inaccessible by Prison and Probation Service officers. In fact, the only link between The Swedish Prison and Probation Service and the application is the automatic synchronisation of meetings via the calendar function. Beyond this, no data are saved, and it is the responsibility of the users themselves to back up their personal information. However, it is conceivable that, in the future, data stored in the application could be used to predict users’ recidivism or critical moments. This can threaten the transparency and sustainability (the fourth pillar), if users do not know how the information they insert into the application is used to predict future risks of recidivism.

Concern for the secure treatment of data is made very explicit in the application's promotional material. However, datafication concerns not only the treatment and usage of data, but also the transformation of complex processes (i.e., rehabilitation) through datafication; in the case of this application, this transformation involves the differentiation of rehabilitation into distinct periods, tasks, and risks to be mitigated, leaving little room for the user or probation officer to navigate. This highlights a similarity with automated decision-making in the employment services and the use of data-driven methods in PSM discussed above, in which human-centred approaches were also sidelined.

Following the principles of a data welfare state, datafication in the corrections sector needs to follow the four pillars of non-bias, decommodification, diversity, and transparency. More specifically, if data-based prison technologies are to avoid comprising the principles of the data welfare state, they would need to strive for nondiscrimination by avoiding specific biases (the first pillar), for example, by not predicting recidivism based on risk-scoring and historical data, instead allowing for rehabilitation and social mobility of incarcerated individuals. Additionally, datafied corrections within the welfare framework should strive for decommodification (the second pillar) by reducing the presence and influence of commercial actors in the corrections sector, including going beyond the problematic discourse of technological backwardness currently dominating and justifying datafication projects in public–private partnerships. In order to uphold the third pillar (data diversity), the corrections sector should strive to acknowledge diversity and work against algorithmic standardisation that forestalls the importance of individual needs in, for example, rehabilitation. Lastly, the corrections sector should strive for transparency (the fourth pillar) in terms of meaningful information for incarcerated individuals on how decisions impacting their everyday life – for example, about their placement, work assignments, and programme activities – are made, and in which ways algorithmic systems influence these decisions.