Policy Significance Statement
Politicians and policy makers should be aware that data-driven technologies are inherently political and can potentially negatively impact citizens’ lives. This research shows that limited, reactive political discussion combined with fragmented treatment of technical, ethical, juridical, and political aspects can obscure decision-making and impede accountability mechanisms. This fragmentation not only weakens democratic legitimacy but also risks ethical oversight, potentially leading to dehumanization and stereotyping of citizens. To prevent these issues, policymakers should prioritize comprehensive discussions on data-driven technologies, ensuring that all relevant aspects are addressed cohesively. By fostering informed discussions, local governments can better safeguard public values and mitigate potential harms of digital welfare systems for citizens.
1. Introduction
As digital welfare systems rapidly expand within local governments worldwide, understanding their political implications is crucial for both public administrators and citizens (Alston, Reference Alston2019; Pederson, Reference Pederson, Pederson and Wilkinson2019). Data-driven technologies, including basic registration systems, dashboards, and advanced predictive analytics tools, are increasingly employed to manage social services and streamline decision-making processes in social welfare. They serve multiple purposes, such as assessing eligibility for benefits, detecting fraud, and allocating resources (Eubanks, Reference Eubanks2017; Vogl et al., Reference Vogl, Seidelin, Ganesh and Bright2020; Kersing et al., Reference Kersing, Van Zoonen, Putters and Oldenhof2022). Strikingly, data-driven technologies are often framed in technocratic terms as neutral, apolitical, administrative tools focused on efficiency, overlooking their inherent political and social dimensions (Jasanoff, Reference Jasanoff2005; Winner, Reference Winner2017).
In the Netherlands, political discussions within local governments about the digitalization of welfare remain surprisingly limited (Kruiter, Reference Kruiter2018; Goderie, Reference Goderie2022), even though previous research indicates that the use of data-driven technologies can negatively affect the work of frontline professionals (Kersing et al., Reference Kersing, Van Zoonen, Putters and Oldenhof2022) as well as citizens’ everyday lives (Oldenhof et al., Reference Oldenhof, Kersing and van Zoonen2024; Kersing et al., Reference Kersing, Oldenhof, Putters, Van Zoonen, Galis and VlassisForthcoming). Only when there is an incident, such as the Dutch childcare benefits scandal, the SyRI social welfare fraud detection algorithm that violated the European Court of Human Rights, or Rotterdam’s discriminatory benefit fraud risk-assessment algorithm, the municipal government pay attention to digitalization, but broader political debate on technology has yet to get off the ground (Algorithm Watch, 2020; Das et al., Reference Das, Faasse, Karstens and Diederen2020; Henley, Reference Henley2021; Open Rotterdam, 2023).
According to the Association of Dutch Municipalities (VNG) it is necessary for the legitimacy of data projects in the public sector that elected local politicians are more involved in discussing the development of data-driven governance, because in digitization processes, public values like transparency, legitimacy, accountability, and privacy are under pressure (Verhoeven 2019 p.7, in VNG, Reference van den Berg, Schaefer, Muis, de Graaf, Banning and Klein2021, VNG, 2022). A lack of political debate about the use of data-driven technologies can erode democratic legitimacy by obscuring how decisions are made and impeding mechanisms of accountability.
The dualized structure of local government in the Netherlands poses an obstacle to discussions on the responsible use of data-driven technologies, as it is usually viewed as an implementation issue that should be handled by aldermen and civil servants, rather than a political matter (Centre for BOLD Cities, 2023). Current literature provides limited insight into local politicians’ involvement in discussions about processes of digitalization and datafication within bureaucratic organizations, but does suggest that there is little interaction between local politicians and civil servants when it comes to data-driven technologies. Civil servants typically develop data projects but only inform or consult local politicians in a relatively late stage when such a project has already led to an actual policy, limiting political oversight (van Est et al., Reference van Est, de Bakker, van den Broek, Deuten, Diederen, van Keulen, Korthagen and Voncken2018; VNG, Reference van den Berg, Schaefer, Muis, de Graaf, Banning and Klein2021). Local politicians’ limited knowledge of data-driven technologies, along with the complexity and abstract nature of these projects, further hinders meaningful engagement (VNG, Reference van den Berg, Schaefer, Muis, de Graaf, Banning and Klein2021; Centre for BOLD Cities, 2023). Furthermore, civil servants have shown reluctance to fully inform politicians, fearing they may thwart the development of data projects (VNG, Reference van den Berg, Schaefer, Muis, de Graaf, Banning and Klein2021).
In this research, we explore the underlying mechanisms for the lack of political debate around the use of data-driven technologies by identifying and analyzing discussions in the municipal council of the city of Rotterdam in the Netherlands about the use of data-driven technologies in the domain of work and income, which is responsible for the deployment of social benefits. Despite being a frontrunner compared to smaller municipalities in experimenting with such technologies, it is relatively new in the domain of work and income, and not without problems. In recent years, the municipality has been criticized by the local audit office and investigative journalists for negative consequences for citizens of their use of algorithms in this domain, such as privacy issues and unequal treatment based on biased results (Rekenkamer Rotterdam, Reference Rotterdam2021, Open Rotterdam, 2023). Given that these recent scandals have exposed serious risks of discrimination and accountability failures, the domain of work and income is particularly relevant to focus on. Moreover, unlike domains such as infrastructure, where data-driven work is more established, work and income show a fragmented adoption of data-driven technologies, offering a unique lens on early-stage tensions between technological innovation and fundamental public values.
Therefore, we pose the following research question: What mechanisms explain if and how data-driven technologies in the domain of work and income are politically discussed within the municipal government of Rotterdam, and what are the consequences?
The aim of this study is to gain insight in if and how discussions about the use of data-driven technologies were politicized or not. By (de) politicization, we refer to the process of making an issue or identities (in)accessible for public deliberation or contestation (Eliasoph and Lichterman, Reference Eliasoph and Lichterman2018). (De)politicization theories offer valuable insights into political discussions around the use of data-driven technologies in the domain of work and income. They enable us to investigate three important aspects of political discussions: the where, what, and how. Firstly, where discussions are taking place refers to the physical place, such as political arenas, agencies, boards, and commissions (i.e. Wood and Flinders, Reference Wood and Flinders2014). Secondly, they indicate what is discussed. By politicization and depoliticization, actors influence what issues are up for deliberation and which ones are not (i.e. Wolf and van Dooren, Reference Wolf and van Dooren2018). Thirdly, thematic types of depoliticization used by actors give an indication of how issues are discussed (in a political, economic, technological, ethical, or legal way) (i.a. Zürn, Reference Zürn, Zürn and Ecker-Ehrhardt2013).
To answer this question, we used a sequential mixed methods design combination of automated text-analysis software ConText (1.2.0) and text-analysis software Atlas.ti (9) to analyze municipal council documents. We took an inductive, exploratory approach, analyzing all documents to make sure we would get a general view of how the municipal council discusses data-driven welfare provision both before and after scandals occur.
In the next section, we present our theoretical framework. Then we present our methods. In the fourth results section, we present our findings. In the conclusion, we summarize and critically reflect on our findings.
2. Theoretical framework
Theories on digital transformation in local government indicate that political attention and support for digitalization at the municipal level remain limited (Benfeldt et al., Reference Benfeldt, Persson and Madsen2018; Kuhlmann and Heuberger, Reference Kuhlmann and Heuberger2023). This lack of engagement is often attributed to the lack of capabilities, knowledge, and skills of political leaders (Benfeldt et al., Reference Benfeldt, Persson and Madsen2018; Gasco-Hernandez et al., Reference Gasco-Hernandez, Nasi, Cucciniello and Hiedemann2022). Although political leaders have the potential to steer digital transformation, many lack the expertise to articulate a strategic vision to guide a digital transformation or understand the value-creating potential of data (Benfeldt et al., Reference Benfeldt, Persson and Madsen2018; Gasco-Hernandez et al., Reference Gasco-Hernandez, Nasi, Cucciniello and Hiedemann2022). Consequently, there is minimal political pressure to address the governance of digital transformation at the local level (Benfeldt et al., Reference Benfeldt, Persson and Madsen2018; Gasco-Hernandez et al., Reference Gasco-Hernandez, Nasi, Cucciniello and Hiedemann2022; Kuhlmann and Heuberger, Reference Kuhlmann and Heuberger2023).
Data-driven technologies are often portrayed as neutral, objective tools used to improve efficiency, yet critical scholars emphasize their inherently political nature (Jasanoff, Reference Jasanoff2005; Winner, Reference Winner2017). Technologies embed values and hold political properties through their design and implementation (Iliadis and Russo, Reference Iliadis and Russo2016; Winner, Reference Winner2017). The ideology of dataism, rooted in neoliberal and positivist thinking, obscures this by framing digitalization as apolitical (Kitchin and McArdle, Reference Kitchin and McArdle2015; Pederson, Reference Pederson, Pederson and Wilkinson2019). This framing enables depoliticization, reducing complex moral and political issues to technical problems handled by data-driven systems (Neubauer, Reference Neubauer2011; Rodima-Taylor et al., Reference Rodima-Taylor, Campbell-Verduyn and Bernards2024). As a result, accountability is diffused and democratic oversight diminished, reinforcing inequalities through claims of technological neutrality.
The lack of political engagement in digital transformations aligns with broader concerns raised in the literature on (de)politicization. While (de)politicization literature includes cases in areas such as civil service (e.g., Peters and Pierre, Reference Peters and Pierre2005), health care (e.g., Landwehr and Böhm, Reference Landwehr and Böhm2011), and electoral politics (e.g., Majone, Reference Majone2001), empirical studies addressing (de)politicization involving data-driven technologies remain scarce. Examples include studies on how democracy is undermined by (de)politicization of certain issues through social media algorithms (e.g., Bayamlıoğlu, Reference Bayamlıoğlu2017) and analyses of the extent to which digital citizenship is a (de)politicized construct (e.g., Schou and Hjelholt, Reference Schou and Hjelholt2017). However, there is a knowledge gap regarding explanatory mechanisms behind the lack of political debate surrounding the use of data-driven technologies in the domain of work and income. This study addresses this gap by offering a unique case study that enhances our understanding of (de)politicization mechanisms in discussions about data-driven technologies within this domain. We use the notion of (de)politicization as an analytical lens to explore the underlying mechanisms that contribute to the limited political debate around the use of data-driven technologies in the domain of work and income.
2.1. (De)politicization as a lens for analyzing political discussions
Since we aim to explore if and how data-driven technologies are at all politically discussed by the municipal government, we use Eliasoph and Lichterman’s (Reference Eliasoph and Lichterman2018) interaction-centered approach to cultures of politics, which focuses on interactions between actors in institutions, formal organizations, or informal settings, as a starting point to assess if and how issues are made political. Rather than assuming that some issues are inherently more political than others, they argue that people, groups, or institutions acting in concert can politicize and depoliticize issues or people. They argue that cultures of politics are shared methods of politicizing or depoliticizing. Politicizing meaning “action, collective or individual, that makes issues or identities into topics of public deliberation or contestation.” This is not a one-way street. Actors can also depoliticize issues by “making once-salient issues or identities inaccessible to deliberation or contestation” (Eliasoph and Lichterman, Reference Eliasoph and Lichterman2018). It is therefore important to remain attentive to the dynamic movement of issues in and out of public debate over time, assuming that repoliticization, understood as countering depoliticization by bringing an issue back into deliberation, may also occur (Skoog and Svensson, Reference Skoog and Svensson2023).
(De)politicization types are particularly useful for our analysis because they indicate three important aspects of political discussions: the where, what, and how of discussions on the political level. Firstly, where discussions are taking place, here we mean the physical place, such as political arenas, agencies, boards, and commissions. Secondly, they indicate what is discussed. Through (de)politicization mechanisms, actors influence what issues are up for deliberation and which ones are not. Thirdly, thematic types of depoliticization give an indication of how issues are discussed. How actors thematically frame issues indicates whether issues are discussed in a political, economic, technological, ethical, or legal way.
While (de)politicization is often viewed as a deliberate strategy to make certain issues (in)accessible to deliberation, we conceptualize it as a phenomenon that emerges from interactions between actors and institutions within specific contexts, either intentionally or unintentionally. We acknowledge that (de)politicization may also result from limitations within the political context—such as governmental norms, habits, structures, and limited resources—that shape local politicians’ actions. Therefore, we refer to these phenomena as (de)politicization types or mechanisms rather than strategies, as they are observable phenomena rather than necessarily intentional actions. Since we ‘observe’ (de)politicization types that the documents showed us, we refrain from attributing intentionality to them.
With these adaptations in mind, our initial analysis was guided by two types of (de)politicization: (1) governmental depoliticization and (2) discursive depoliticization (Wood and Flinders, Reference Wood and Flinders2014).
2.1.1. Governmental (de)politicization
Governmental depoliticization refers to the delegation of issues from the political arena towards the administrative arena (Wood and Flinders, Reference Wood and Flinders2014). One approach, known as “the politics of ABC,” involves shifting decision-making power to at arm’s-length entities such as agencies, boards, and commissions (Flinders and Buller, Reference Flinders and Buller2006; Wood and Flinders, Reference Wood and Flinders2014; Etherington and Jones, Reference Etherington and Jones2018). Another form, known as ‘rule-based’ depoliticization, refers to delegating authority to judicial structures or technocratic rule-based systems. The introduction of new rules and regulations ‘binds the hands’ of politicians because it limits their discretion (Flinders and Buller, Reference Flinders and Buller2006; Wolf and van Dooren, Reference Wolf and van Dooren2018). A third type involves distancing personal responsibility and blurring accountability by diffusing the responsibility across a range of interdependent actors (Wood and Flinders, Reference Wood and Flinders2014). It implies that politicians use ‘the problem of many hands’ as described by Thompson (Reference Thompson1980) on purpose to blur accountability structures (Wood and Flinders, Reference Wood and Flinders2014; Etherington and Jones, Reference Etherington and Jones2018). In practice, however, (de)politicization is not a one-way street. A study of 11 municipalities in Sweden shows that, while politicians often depoliticize issues, public administrators also actively repoliticize issues (Skoog and Svensson, Reference Skoog and Svensson2023).
2.1.2. Discursive (de)politicization
In discursive depoliticization, ideas and language play an important role because they rely on language and framing to shift issues from and to different domains, thus altering their content (Wood and Flinders, Reference Wood and Flinders2014). Discursive depoliticization can restrict deliberation on certain topics, as the way language is used and issues are framed can diminish opposition by making alternative views appear “irrational” (Wood and Flinders, Reference Wood and Flinders2014; Wolf and van Dooren, Reference Wolf and van Dooren2018). Through strategically manipulating the framing, the political significance of an issue may be downplayed. Emphasizing certain aspects while downplaying others and framing an issue in a neutral or technical language can reduce its political relevance.
Discursive depoliticization of issues can also occur through thematic types of framing, where an issue is moved from or to functionally different spheres, such as the political, economic, religious, legal, educational, or scientific sphere (Zürn, Reference Zürn, Zürn and Ecker-Ehrhardt2013; Jessop, Reference Jessop2014; Kreuter, Reference Kreuter2020). Issues may be depoliticized through the processes of technocratization or economization. Conversely, politicization is defined as moving an issue into the political sphere from another sphere (Zürn, Reference Zürn, Zürn and Ecker-Ehrhardt2013, p. 21, Kreuter, Reference Kreuter2020).
2.1.3. Combining (de)politicization types
In practice, a combination of governmental and discursive depoliticization types can be used by actors. For example, first, a complex political issue is framed as a technical or administrative problem that requires expert knowledge and solutions (discursive depoliticization). Then, secondly, political debate is sidelined by delegating the issue to agencies, boards, and commissions (governmental depoliticization) (Flinders and Buller, Reference Flinders and Buller2006; Wood and Flinders, Reference Wood and Flinders2014; Skoog and Svensson, Reference Skoog and Svensson2023). This discursive framing typically emphasizes reliance on expert knowledge, enabling politicians to depoliticize contentious issues by deferring issues to non-elected technocrats or specialized consultants, thereby distancing themselves from personal responsibility (Barbi, Reference Barbi2018).
3. Methods
3.1. Research design and data collection
We used a sequential mixed methods design using automated text-analysis software Context (1.2.0) and text-analysis software Atlas.ti (9) to analyze documents and video recordings of municipal council and committee meetings (Figure 1).
Sequential mixed methods design.
We took an inductive, exploratory approach to make sure we would get a general view of how the municipal government discusses the use of data-driven technologies in the domain of work and income, both before and after scandals occur over a period of 8 years (2016–2023). This approach avoids bias towards extreme and exceptional cases by capturing routine discussions, while still recognizing the analytical value of scandals in revealing (clashing) values, political struggles, assumptions, metaphors, and hidden bureaucratic practices (Pinch and Leuenberger, Reference Pinch and Leuenbergern.d.).
The collection of data involved manually downloading all relevant documents from the municipal government information system websites (Figure 1). From the municipal council (terms 2014–2018, 2018–2022 and 2022–2026), all agendas and all meeting minutes were collected. From the three municipal council committees in the domain of work and income, the WIPV committee (2014–2018), the WIISA committee (2018–2022), and the WIOSSAN committee (2022–2026), all agendas and all motions, written questions, political commitments, and advice in the domain of work and income were collected. The council can establish committees consisting of councilors and non-council members, that prepare decisions of the municipal council on certain topics, for example, work and income, advise them, and engage with the mayor and aldermen (ProDemos, n.d.; Gemeenteraad van Rotterdam, n.d; politiekeambtsdragers.nl, n.d.). All documents and video recordings are publicly available on the municipal government information system websites.
Since the data collection resulted in a vast volume of documents, totaling thousands of pages, which was too large for manual analysis, we first employed automated text analysis software (Context 1.2.0) to make a selection, followed by a more in-depth qualitative text analysis with Atlas.ti (9) (Figure 1).
Firstly, we analyzed the material using automated text-analysis software Context (1.2.0) (Figure 1). The topic modeling and corpus statistics analysis with ConText gives insight in if data-driven technologies are discussed in the municipal council meetings and the municipal committee meetings, and what is discussed.
Secondly, we analyzed a selection of the material more in-depth using text-analysis software Atlas.ti (9) (Figure 1). This analysis gives insight into how data-driven technologies are discussed. The selection was based on preliminary desk research and critical events. We paid special attention to events that occurred between 2016 and 2023 and sparked public and/or political debate.
3.2. Data analysis
3.2.1. Automated text analysis
We used automated text-analysis software Context (1.2.0) to analyze documents of municipal council and committee meetings (Figure 1).
After manually downloading all pdfs and converting them to text files, we cleaned and pre-processed the documents (1) by applying stemming to remove plurals and endings of verbs, and (2) removing the stop words (noise words). Our analysis included Corpus Statistics and Topic Modeling.
Corpus Statistics shows for each word (1) the word frequency (how often a word occurs in the text corpus), (2) TF-IDF score (how words are distributed in documents), and (3) ratio/percentage of text a word occurs in (see, e.g., Table 1). High TF-IDF scores indicate that a word is very frequent in some documents but appears in fewer text documents overall. Low TF-IDF scores indicate noise words.
Example of corpus statistics results
We manually went through all the corpus statistics lists and identified words that could indicate a discussion about data-driven technologies in the domain of work and income. We manually cross-checked the documents to confirm if these words occurred in relevant discussions. An overview of the years of all the words that were related to data-driven technologies in the domain of work and income was visualized in charts such as Charts 1 and 2.
Topic Modelling (based on LDA) uncovers sets of words that may constitute distinct topics in a corpus. ConText displays (1) the topic number, (2) the weight, and (3) the topic members (see, e.g., Tables 3 and 4). The topic number indicates how well the topic fits the data. The topics are automatically sorted by fit. The weight indicates how well topic members fit the topic. Relatively high values represent better-fitting topics. The topic members are words that together form a theme. Each theme is represented as a vector of words, which are sorted based on their strength of association with a topic. We chose the topic numbers based on the size of the bodies of the texts.
Corpus statistics municipal council agendas work and income topics 2016–2023.
Corpus statistics committee agendas, work and income topics 2016–2023.
3.2.2. Qualitative text analysis
We selected documents for the qualitative text analysis with Atlas.ti (9) based on preliminary desk research and identification of critical events that occurred from the automated text analysis.
We inductively coded the material with the (de)politicization types as sensitizing concepts. We coded the material in the first round of open coding, followed by a second round of axial coding (Neuman, Reference Neuman2014).
In the first round of open coding, we used depoliticization types as sensitizing concepts to interpret the data while keeping an open mind to new codes. While some codes aligned with the sensitizing concepts, other codes revealed depoliticization types that were not present in the literature, such as “content chopping.”
In a second round of axial coding, we organized, linked, and grouped codes into categories like “lacking knowledge,” “content chopping,” and “dehumanizing.” During this round, we identified relationships between codes and categories. For example, how categories such as “lacking time” and “lacking knowledge” are related to “distancing responsibility and diffusing accountability,” and how they, in tandem, can lead to “content chopping.”
Keeping an open mind allowed us to also identify risks by grouping codes into categories such as “dehumanizing people” and “stereotyping.”
We used several triangulation strategies to ensure the data quality.
Firstly, we used method triangulation by using different methods and software to minimize retrieval bias (Grimmer et al., Reference Grimmer, Roberts and Stewart2022). A potential risk was that discussions included terminology that differed from our expectations, which could lead to missing important discussions. To mitigate this, we cross-checked results from the corpus statistics and topic modeling done with ConText (1.2.0) with the results from the qualitative analysis with Atlas.ti (9).
Secondly, we used data triangulation by cross-checking information from different sources to avoid resource bias (Grimmer et al., Reference Grimmer, Roberts and Stewart2022). We cross-checked the information gathered through the document analysis and the video recordings. The recordings provided valuable context for discussions in the committee meetings because there were no ad verbatim meeting minutes from the committee meetings.
Thirdly, through researcher triangulation, we discussed and refined our findings. The first author collected, analyzed, and coded the data, which were then discussed in conversation with the other authors. (Mortelmans, Reference Mortelmans2020).
4. Results
4.1. Reacting to scandals and criticism “…and proceeds with the order of the day”
Our results show that while data-driven technologies are used in the domain of work and income, there are rarely discussions (answering the if question) in the municipal council on how the data-driven technologies are used (Table 2, Chart 1). Chart 1 shows the only instances in eight years of 150 municipal council agendas where words related to data-driven technologies in work and income appear in the corpus statistics. In 2016, 2017, 2018, and 2020, no words related to the use of data-driven technologies in work and income occurred in the corpus statistics and topic modeling. On the rare occasions that data-driven technologies are discussed in the municipal council, it is discussed either (1) in reaction to scandals, or (2) in reaction to criticism (Table 2, Chart 1).
Timeline 2016–2023 topics discussed in the municipal council and council committees
While municipal council members occasionally raise critical questions, formal ethical reflections on data-driven technologies are mostly described in reports from external audit organizations such as the municipal ombudsman, the Rotterdam Audit Office, and the Concern Auditing. Internal reflections focus primarily on technical issues within committees (Table 2, Chart 2).
Chart 2 displays all instances over eight years in 143 committee agendas where words related to data-driven technologies in work and income appear in the corpus statistics. These instances peak in similar periods as in the municipal council agendas, with similar topics. However, the committee agendas contain a greater number and specificity of words, indicating that discussions occur more frequently and in greaterdetail within the council committees (Table 2, Chart 2).
Over time, most discussions about scandals and critical reports shifted towards municipal council committees, employing governmental depoliticization mechanisms such as the ‘politics of ABC.’ For example, the discussion about the Benefit frauds analytics project (Box. 1) was shifted from the council to committee meetings (Charts 1, 2). Furthermore, scandals and critical reports involving externally developed data-driven technologies (e.g., SyRI, childcare benefits scandal algorithm) tend to be discussed in council meetings, while those involving internally developed technologies (e.g., the analytics benefit fraud project) are primarily addressed in committee meetings (Charts 1, 2).
Benefit fraud analytics project
The municipality of Rotterdam used a benefit fraud risk-assessment algorithm from 2018 until 2021 to select benefit recipients for re-assessment.
In 2016, the municipality hired Accenture to develop a pilot model that assigns individual benefit recipients a risk score that indicates potential welfare fraud. People with the highest scores are selected for re-examination.
In 2018, the municipality implemented an improved version. After an evaluation, Accenture indicates that detecting fraud is difficult, but that the model can predict “unlawfulness”: the likelihood of issues with benefits (Klaassen and van Dijk, Reference Klaassen and van Dijk2023; Open Rotterdam, 2023).
In 2021, the Rotterdam Audit Office warned in their report “Colored Technology’”that the algorithm can lead to biased results because attention for transparency and responsibility was lacking (Rekenkamer Rotterdam, Reference Rotterdam2021).
In 2023, investigative journalists revealed that the algorithm unfairly targeted young single mothers, increasing their chance of being re-examined for welfare fraud (Open Rotterdam, 2023).
The topic modeling analysis supports the corpus statistics results, showing that discussions about data-driven technologies in work and income are rare. From 193 agendas and meeting minutes spanning 150 municipal meetings over eight years, only 2 relevant topics emerged. The relatively low topic numbers (13 and 14 out of 20) indicate that these topics received significantly less attention compared to other topics (Figure 2).
Visualization of 2 relevant topics in 193 documents period 2016–2023.
The first topic appears in council meeting agendas in 2022, where the words “governance” and “algorithms” refer to a new governance approach for governance of sensors, data applications, and AI (Table 3).
Topic modeling agendas 2016–2022
The second topic occurs in the agendas of committee meetings in 2023 (Table 4) and refers to two committee meeting agendas listing a letter from the Dutch Data Protection Authority (hereafter DDPA). In this letter, the DDPA raises concerns about the risk-assessment algorithm used in the Benefit Fraud Analytics Project (Box. 1), requiring the municipality to submit a report on the impact of the algorithm.
Topic modeling agendas 2016–2023
4.2. Barriers to discussion
Our inductive analysis revealed two discursive factors that are used to justify limited and shifting political discussion towards council committees: (1) claims of lacking time and knowledge, and (2) distancing responsibility and diffusing accountability.
4.2.1. Lack of time and knowledge
Both council members and aldermen claim to have limited time and knowledge to do their job. Earlier research indicates that even though aldermen have a full-time paid position, whereas council members do not, both experience high work pressure and insufficient knowledge and skills to do their job properly (ProDemos, n.d.; Boonstra, Reference Boonstra2024; Raad voor het openbaar bestuur, 2018, 2020).
During discussions, both aldermen and council members indicate that most of them lack sufficient knowledge to critically question data-driven technologies, often resorting to various depoliticization mechanisms to deal with this knowledge gap. Although they are aware of these limitations, they do not indicate whether or how they plan to allocate more time or address the knowledge gap, potentially using these constraints as a convenient justification to depoliticize the topic of data-driven technologies. We illustrate this de-politicization mechanism with the Benefit Fraud Analytics Project (Box 1).
In a response letter to Concern Auditing’s recommendations on improving the transparency around the risk-assessment model in the benefit fraud analytics project, the alderman writes:
The note aims to provide you with the necessary clarity, albeit concise, about the (further) development and use of the risk assessment model. Because this involves complex mathematical matters, it remains difficult to fully understand how the model works. I find it crucial that we work with scientifically substantiated programming and internationally validated software, but above all: expert employees. (Letter alderman, August 25th, 2021).
By describing it as a “complex mathematical matter” the alderman shifts the issue from the ethical sphere (lack of transparency), towards the mathematical sphere.
By acknowledging that “it is difficult to fully understand how the model works” the alderman employs the depoliticization mechanism of distancing responsibility, by diffusing the responsibility among various actors such as programmers, software makers, and expert employees.
The alderman shifts the issue towards the scientific sphere by referencing “scientifically substantiated programming,” and an authoritative international community that can “validate software,” while also moving the issue from the political to the bureaucratic arena by emphasizing working with ‘expert employees.’
4.2.2. Distancing responsibility and diffusing accountability
Political discussion is further limited by the formal separation of responsibility for content and process of data-driven technologies: the ‘organization’ alderman is responsible for ICT systems, while the ‘sector-specific’ alderman is responsible for ensuring the system’s use and cultivating a culture where necessary adjustments or new releases are requested. This division separates the responsibility for the design and use of ICT systems. The organization’s alderman stated:
I am not responsible for whether a benefit is paid correctly, although I am responsible for the system. The sector-specific alderman is responsible for and involved in the functioning of a cluster. (Aldermen in municipal council meeting in 2017).
In this discussion, it was also mentioned that sector-specific aldermen typically lack sufficient ICT knowledge, making it a logical choice to further delegate responsibility to the organization alderman or to technical sessions in committee meetings. In this context, the delegation of responsibility results from the division of tasks within the organization.
Concerns were raised about the diffusion of accountability. A council member remarked that “it’s always useful to outline responsibilities on paper, but the problem is that it’s just paper.” It was concluded that shared responsibility repeatedly leads to issues. As one council member aptly summarized the problem of many hands (Thompson, Reference Thompson1980): “When responsibility is shared among many, it effectively falls on no one.”
4.3. Content chopping
The claims of lacking time and knowledge, combined with a lack of responsibility, lead to a depoliticization mechanism that we call “content chopping.” We define this as chopping an issue into small content pieces, for example, technical, ethical, political, or executive aspects, and spreading them into separate documents and discussion arenas. Thereby, it obscures the overall coherence of an issue, which diffuses critical concerns. The mechanism of content chopping has clear parallels in science and technology studies, where scholars have shown how complex issues are often fragmented in ways that undermine holistic political discussions (Thoreau and Delvenne, Reference Thoreau and Delvenne2012; Chambers, Reference Chambers2023). We will illustrate the mechanism of content chopping with the Benefit Fraud Analytics Project (Box 1).
In 2021, the Concern Auditing provided advice requested by the municipal council about the risk assessment model. This advice included the recommendation to improve the lack of transparency regarding the use of the risk-assessment model. In a response letter, the alderman explains which steps have been taken in the past period and which actions are planned in the coming period. In this letter, content chopping occurs, with the different aspects of the issue spread across multiple documents.
The first document mentioned in his letter is a note named “Explanation of the operation of the risk assessment model” that has the aim of explaining how the model works. The technical aspects of the model are isolated from other aspects in this note.
The second document, named “Re-examination Work and Income (background, working methods, re-examination methods)” explains how re-examinations are carried out. The alderman writes that because this document already explains how re-examinations are carried out, this will not be discussed in the note about how the risk assessment model works. Hereby, the policy aspects of this model are isolated from the technical aspects in this document.
A third document is a letter that discusses “increasing the human dimension when carrying out re-examinations.” In this letter, the ethical aspects of the model are isolated from technical and policy aspects. However, this letter is nowhere to be found in the council information system or outside of it.
4.4. Risks of content chopping
Content chopping can especially be harmful when different aspects are separated from each other, because when these aspects are not coherently discussed, one aspect can become more prominent and take precedence over other aspects. The main risks of content chopping are that it can lead to (1) dehumanizing people and (2) stereotyping.
4.4.1. Dehumanizing
We use the benefit fraud analytics project (Box 1) to illustrate how content chopping can lead to dehumanizing people, which we understand as abstracting issues to the point where human elements are lost. In this case, different aspects were dispersed across separate discussion arenas and documents, with a strong focus on the technical aspects compared to other aspects (see 5.3).
Another follow-up to the Concern Auditing’s recommendations about transparency in the benefit fraud analytics project were presentations in two technical sessions of the committee by two data analysts. The first session focused on explaining how re-examinations are carried out, while the second session aimed to explain the workings of the fraud risk prediction algorithm.
In the second technical session, characteristics of benefit recipients were replaced by characteristics of apples in an example to explain how the algorithm works. In an attempt to simplify the complex topic for council members, the two data analysts used an unfortunate comparison likening potential fraudulent benefit recipients to “apples used to make applesauce,” implying that certain characteristics lead to its unsuitability for sale and its processing into apple sauce. They explain how they, based on looking at characteristics of 12 apples in the table, created a decision tree to figure out which apples are likely to end up in the applesauce and which ones will get stickered and sold separately (Figure 3). The data analyst, in a reflection on their attempt, said:
What stands out about our attempt? That with a few rules, we can still fairly well categorize apples. (…) But also, that none of the rules we have come up with are flawless (…) And what also stands out, at least for us, is that this was somewhat, well, randomly looking at that table, figuring out what could be selected. So, we just searched randomly. (Data analyst, technical session in 2021).
PowerPoint slide technical session.
He does not express any worries about the potential consequences of apples being miscategorized, nor about benefit recipients that are unfairly flagged as potentially fraudulent and invited for a re-examination conversation based on an algorithm that does not function properly. Isolating the technical aspects within a technical session leaves little room for reflection on ethical aspects, such as the social impact on benefit recipients’ lives, and ensuring equal treatment and fairness. The metaphor, referring to the “one bad apple can spoil the barrel” saying, illustrates the dehumanization of benefit recipients. While crushed apples taste quite good, crushed citizens certainly do not.
4.4.2. Stereotyping
Another risk of content chopping is that it can reinforce stereotypes, such as the assumption that individuals from certain ethnic or racial groups are more likely to commit benefits fraud. This was particularly evident in 2018 during a municipal council discussion on whether Turkish Dutch welfare recipients were more likely to conceal foreign assets. Content chopping can lead to the embedding of human biases in data-driven technologies through stereotyping, particularly when a dominant political perspective overshadows ethical and scientific aspects. We illustrate this with the risk profile foreign assets case (Box 2) in which negative attitudes toward Turkish Dutch recipients became embedded in the design of a risk profile.
Risk profile foreign assets
In 2019, the alderman of Work and Income wants to set up an experiment in collaboration with the International Bureau Fraud Information (IBF), which is part of the Institute for Employee Benefit Schemes (UWV), and the Social Insurance Bank (SVB). The aim is to investigate how the detection approach to find concealed foreign assets can become more effective. In this experiment, a risk profile is used to select people for re-examination. The risk profile is based on characteristics such as non-Dutch country of birth, repeated residence abroad, and spending over 28 days abroad annually.
In 2018, the motion “Decisive action against undisclosed foreign assets” was discussed in the municipal council. This motion claims that “an estimated 20% to 30% of Turkish Dutch welfare recipients conceal foreign assets.” During the discussion, council members criticized these statements, as they were based on research conducted by a foreign commercial agency, whose reliability was questionable. Moreover, these figures did not align with trustworthy studies conducted in the Netherlands. The alderman warned:
The only thing I see is that people are merely copying headlines [from the study], thereby creating their own version of reality. (…) And based on that self-created reality, we are being asked to completely overhaul our policy.
These discussions reveal an ongoing political struggle over the stereotyping of welfare recipients who are likely to commit fraud. In this case, the stereotype concerns fraudulent Turkish Dutch recipients. Throughout this discussion, the dominant political perspective threatens to overshadow scientific and ethical aspects, such as basing policy on trustworthy scientific research. This is also reflected in the design of data-driven technologies aimed at detecting fraud because eventually characteristics such as “country of birth is not the Netherlands,” “residence abroad has been established on several occasions,” and ‘priority for job seekers who spend more than 28 days abroad per year’ were used to determine the risk profile.
In 2021, the Socialist Party challenged this stereotype by publishing a “Black book on welfare abuses” in which they write:
On the one hand we see a municipality that spends millions on combating fraud, on the other hand we see millions being cut back on job guidance. For example, the council invests in the discriminatory experiment to detect foreign assets. However, there was no question of foreign assets. Yet benefits have been stopped because, among other things, Rotterdam residents did not respond to the invitation (SP Rotterdam., Reference Rotterdam2021, p. 5).
They counter the stereotype of Turkish Dutch recipients by replacing it with a stereotype of vulnerable individuals who do not get the support they need in their job search. They counter the dominant political perspective by bringing ethical aspects, such as equal treatment and non-discrimination, and executive aspects, such as cutbacks on job guidance and Rotterdammers not responding to invitations, back into the discussion. In a response letter, the alderman denies any discrimination but ultimately yields to the criticism, leading to the decision to stop using distinctions based on country of birth.
This case illustrates how human biases, such as negative attitudes toward Turkish Dutch benefit recipients, can become embedded in the design of data-driven technologies like this risk profile when a dominant political perspective overshadows ethical and scientific considerations. The stereotype of fraudulent Turkish Dutch recipients frames the discussion in a way that makes a more neutral, individualized perspective- one that treats welfare recipients as human beings—less accessible. Adhering to a stereotypical conception of welfare recipients obscures a nuanced understanding of their identities. However, in this case, the negative stereotyping of Turkish Dutch recipients eventually failed because reintroduction of scientific and ethical aspects into the discussion led to the exclusion of distinctions based on country of birth.
4.5. Governance of algorithms: efforts to prevent applesauce
In 2022, a new governance approach for sensors, data applications, and AI was introduced (Table 3) in response to a report by the Rotterdam Audit Office about transparency and responsibility issues in algorithm use (Rekenkamer Rotterdam, Reference Rotterdam2021, Box. 1). This governance approach includes instruments to ensure transparent and responsible algorithm use. One of the instruments that was introduced was an algorithm registry with the aim of ensuring transparent communication with citizens about algorithm use. Additionally, an algorithm expert has been appointed to provide technical, ethical, legal, and informational perspectives on algorithms. This algorithm expert also advises the newly created Algorithm Advisory Board, an external committee reporting to the executive council. Furthermore, a Human Rights and Algorithm Impact Assessment (IAMA) instrument facilitates balanced discussions among relevant actors in the development phase of algorithmic applications. Lastly, awareness initiatives for individual employees include presentations on what algorithms are, how and where they are used, and which risks and safeguards there are, along with separate roundtable discussions on ethical dilemmas in practice.
While Rotterdam’s recent efforts represent progress toward responsible data-driven governance, they largely depend on expert input and procedural safeguards. It remains unclear how these measures foster democratic oversight and stimulate meaningful political engagement. This raises important questions about what re-politicization could look like in local democracies. Beyond resisting depoliticization mechanisms like content chopping, it involves addressing the normative and societal dimensions of data-driven governance. Re-politicization requires creating space for political struggle and collective deliberation, rather than reducing data-driven governance to technical or administrative matters. This involves not only improving council members’ technical knowledge but also strengthening their willingness and capacity to engage with the broader societal implications.
5. Conclusion and discussion
The goal of this research was to gain insight into what mechanisms explain if and how the use of data-driven technologies in the domain of work and income are politically discussed in the municipal government of Rotterdam, and what the consequences are. Our findings confirm earlier research suggesting limited political attention to digitalization (Benfeldt et al., Reference Benfeldt, Persson and Madsen2018; Kuhlmann and Heuberger, Reference Kuhlmann and Heuberger2023), showing that discussions in the municipal council are rare and occur primarily in response to scandals or in reaction to criticism from audit offices.
In line with previous studies, our analysis shows how (de)politicization mechanisms, such as distancing responsibility and content chopping, explain the limited political debates around data-driven technologies in work and income (Bayamlıoğlu, Reference Bayamlıoğlu2017; Schou and Hjelholt, Reference Schou and Hjelholt2017; Wood and Flinders, Reference Wood and Flinders2014). This study also contributes to new insights by broadening the conceptualization of (de)politicization as an emergent phenomenon shaped by interactions between actors and institutions within specific contexts. By emphasizing observable mechanisms rather than intentional strategies of actors and considering the influence of structural and resource constraints on political actions, we expand the application of (de)politicization mechanisms beyond the deliberate strategies to include unintentional dynamics within political processes.
Political discussions about these data-driven technologies should not be limited to responses to scandals and criticism, especially given the risk of embedding human biases, such as negative stereotypes, into these systems throughout all phases of their existence, as seen in the Risk profile foreign assets case (Box 2). When the different aspects of an issue, such as political, technical, scientific, executive, and ethical aspects, are not discussed coherently, crucial values like justice, privacy, transparency, accountability, and autonomy for citizens risk being sidelined. Without coherent discussions on current and future applications, these technologies may perpetuate harm, effectively trapping groups of citizens in negative stereotypes embedded in their design. Our findings thereby confirm concerns raised in earlier literature that digitalization can lead to dehumanization, bias, and a lack of empathy in public service provision (Alston, Reference Alston2019; Vogl et al., Reference Vogl, Seidelin, Ganesh and Bright2020; Ranchordas, Reference Ranchordas2021), underscoring the urgent need for more political engagement and value-sensitive approaches to data-driven governance.
Political leaders play a crucial role in steering digital transformation and thereby safeguarding values in discussions about data-driven technologies (Verhoeven 2019 p.7, in VNG, Reference van den Berg, Schaefer, Muis, de Graaf, Banning and Klein2021, VNG, 2022, Gasco-Hernandez et al., Reference Gasco-Hernandez, Nasi, Cucciniello and Hiedemann2022). However, in line with earlier research (Benfeldt et al., Reference Benfeldt, Persson and Madsen2018; Gasco-Hernandez et al., Reference Gasco-Hernandez, Nasi, Cucciniello and Hiedemann2022), our research shows that both aldermen and council members indicate that most of them currently lack sufficient knowledge and insight to engage in meaningful political discussions about the design and impact of these technologies. Time and knowledge constraints, due to the part-time nature of council positions, make it understandable that technical aspects are often deferred to experts. Yet, despite recognizing these limitations, they do not articulate plans to address them, suggesting that appeals to limited time and knowledge may serve as a convenient rationale for depoliticizing the topic of data-driven technologies. Additionally, audit offices function only to a limited extent, as issues with data-driven technologies are typically brought to light through their critical reports, often only after harm to citizens has already occurred. Eventually, the recurring problems with the responsible use of algorithms that occurred in our document analysis led in 2022 to the introduction of a new governance approach for sensors, data applications, and AI. While the instruments of this approach are a step in the right direction and provide a solid foundation for further progress in ensuring responsible use of algorithms, future evaluation and research are needed to determine their effectiveness in preventing future problems.
In line with prior studies, we argue that it seems insufficient to only reflect on the limited time, knowledge, and responsibilities of council members and aldermen, as these problems ask for a more rigorous revision of the governmental norms, habits, and structures (Centre for BOLD Cities, 2023; van Est et al., Reference van Est, de Bakker, van den Broek, Deuten, Diederen, van Keulen, Korthagen and Voncken2018; Das et al., Reference Das, Faasse, Karstens and Diederen2020; Belhaj, Reference Belhaj2020; VNG, Reference van den Berg, Schaefer, Muis, de Graaf, Banning and Klein2021, 2022; Jasanoff, Reference Jasanoff2005; Winner, Reference Winner2017). Our analysis shows that current governmental norms and habits, which treat data-driven technologies primarily as tools and confine discussions to the technical sphere, are inadequate for ensuring the responsible use of data-driven technologies in work and income.
While our systematic analysis offers insight into how municipal discussions on digitalization have evolved over time, it captures only part of the institutional context. Relying solely on textual documents and video recordings limits our understanding of politicians’ perspectives on their own intentions, and on informal interactions and discussions. Although we recommend interviews for future research, it is important to take into account potential inaccuracies due to memory limitations or political motivations, particularly with sensitive topics or scandals.
Our understanding of the institutional context is also constrained by incentive bias, as strategic behavior may shape what is formally documented (Grimmer et al., Reference Grimmer, Roberts and Stewart2022). While transparency makes documents accessible, it may also incentivize local politicians to shift sensitive discussions to informal settings. This absence of contextual information can hinder a complete understanding of informal interactions and discussions outside formal meetings, and should be explored in future research.
Our findings are context-specific and should be interpreted with awareness of their limited transferability. Rotterdam’s pragmatic, innovation-driven culture and its working-class history likely influence both the development and reception of data-driven technologies. Additionally, the absence of direct mayoral elections in the Netherlands shapes the political accountability landscape, with potential implications for how data-driven policy is debated and contested. These contextual factors limit the direct transferability of findings to cities with different political systems and civic cultures.
Future research should examine the media’s dual role as a watchdog and a catalyst for public debate during data-related scandals, such as SyRI and the Benefit Fraud Analytics Project. On the national level, media coverage during the childcare benefits scandal not only exposed government misconduct but also faced criticism for amplifying public outrage in ways that influenced policy responses and interventions (Parlementaire Enquêtecommissie Fraudebeleid en Dienstverlening, 2024). However, the media’s interactions with local politicians during municipal-level data-driven technology scandals remain under-researched.
Finally, we highlight the need to investigate how (un)reliable knowledge shapes local political debates. Our findings suggest that inadequate knowledge about data-driven technologies and reliance on research of questionable reliability, such as in the risk profile foreign assets case, can lead to unjust and unequal treatment of citizens, underscoring the need for trustworthy knowledge in municipal policymaking.
Data availability statement
The original documents and recordings are publicly available through the municipal government information system websites of the municipality of Rotterdam:
Documents and recordings up to February 2022: https://rotterdam.raadsinformatie.nl/dashboard
Documents and recordings from March 2022 onward: https://gemeenteraad.rotterdam.nl/
Metadata that support the findings of this study are available from the corresponding author upon reasonable request. However, sharing the metadata in a structured and accessible format is not feasible, as it would require significant time and resources beyond the project’s scope and risks misinterpretation without proper contextual understanding. The document analysis covers eight years of municipal records and video recordings, with metadata dispersed across various formats, systems, and project phases. Much of the metadata is embedded in internal working documents and context-specific annotations that are difficult to separate from internal project documentation.
Acknowledgements
We kindly thank the members of the The Hague Centre for Digital Governance, Leiden University, the members of the Health Care Governance group, and especially Renee Michels for her review, at the Erasmus School of Health Policy & Management, Erasmus University, the participants of the SPA annual conference 2024, and participants of the Data for Policy Conference 2024 for their valuable comments and discussions on earlier versions of this article.
Author contribution
Conceptualization: M.K; L.O. Data curation: MK. Formal analysis: MK. Investigation: MK. Methodology: M.K. Project administration: MK. Supervision: L.O; K.P; LVZ. Visualization: M.K. Writing-original draft: M.K. Writing-review & editing: L.O; K.P; LVZ. All authors approved the final submitted draft.
Provenance
This article was submitted for consideration for the 2024 Data for Policy Conference to be published in Data & Policy on the strength of the Conference review process.
Competing interests
The authors declare none.
Open access
Comments
No Comments have been published for this article.