Automation scenarios: citizen attitudes towards automated decision-making in the public sector

ABSTRACT

This article explores citizen attitudes towards automated decision-making (ADM) in the public sector, addressing concerns related to social justice and transparency. ADM, used in diverse public services, such as benefit application processing, welfare fraud detection and tax calculation, has sparked public interest and scepticism. To shed light on this complex issue and make ADM more accessible for citizens, we presented three domain-specific scenarios in a population-representative survey in Estonia (n = 1,500), Germany (n = 2,001) and Sweden (n = 1,000). These scenarios involved job seeker categorisation, child welfare risk assessment and predictive policing through facial recognition. Drawing from this survey and adopting an exploratory approach, we analyse attitudes across responses to these scenarios and conduct a regression analysis, integrating individual variables such as age, gender, education, awareness, enthusiasm and trust in ADM systems. Our findings reveal differences in citizens’ attitudes based on welfare regimes and individual characteristics. This citizen-focused approach underscores the significance of involving citizens in the governance of ADM in the digital welfare state, transcending the traditional regulatory and stakeholder-centric perspectives.

KEYWORDS:

• Algorithmic automation
• scenarios
• citizen perspectives
• welfare
• welfare fraud
• predictive policing

This article engages with citizen attitudes towards automated decision-making (ADM) in the public sector. In recent years, ADM has been implemented in different forms in all kinds of societal spheres, including the public and welfare sectors (Kaun et al., 2024; Masso et al., 2023; Kaun, 2022; Madan & Ashok, 2023). Deployments of ADM in the public sector include automated decisions about benefit applications. predictive risk scoring of child welfare, artificial intelligence (AI)-powered video surveillance, and predictions of deviant behaviour in public spaces (for a mapping, see the report by Kaun et al. 2023). In connection with the increased implementation of ADM, concerns have been raised concerning the consequences of its use for social justice and inequality. Besides establishing high-risk areas in which the use of ADM is strictly regulated, policy work has focused on implementing measures to increase public transparency. One of the main objectives here is to expose ADM and its algorithms to public scrutiny and discussion as part of the democratic process. However, this has proven notoriously difficult, as technologies are increasingly complex and hard to understand for the general public. This article is an intervention in this field. It seeks feedback on social scenarios as a means of making ADM technology in the public sector palatable and relatable for citizens while considering the domain – and context-specificity of citizens’ attitudes.

Based on three domain-specific scenarios, this work takes an exploratory approach and details differences in attitudes towards algorithmic automation in the three welfare regimes of Estonia, Germany and Sweden with the help of a population-representative survey. The three domains chosen for exploration here have been conspicuous on the agenda, and there are several well-known ADM applications in place across the three countries. Therefore, we focus on employment services, child welfare and predictive policing.

It is notoriously difficult for citizens exposed to algorithmic automation by public sector institutions to evaluate the implications of these technologies. Only a minority of citizens are aware of automated decision-making in the public sector (Denk et al., 2022). In addition, the lack of transparency in the implementation process makes it difficult for citizens to develop informed positions. The issues explored in this article concern how attitudes towards ADM in different domains differ depending on the welfare regime context and individual factors.

To explore the welfare regime  – and domain-specific attitudes of citizens, we developed three scenarios that provided concrete, empirical entry points for the respondents. The three vignettes touch on the use of ADM to sort job seekers into different categories, the use of risk scoring for child welfare and the use of facial recognition for predictive policing. Here, we present the findings of this comparative analysis. We present descriptive findings, first, from a cross-regime comparison of attitudes towards ADM in the three scenarios, and second, from a regression analysis of individual variables (age, gender and education) combined with awareness of, enthusiasm for and trust in ADM systems to explain differences between the three welfare regimes and three domains. We conclude the article with an attempt to explain these disparities based on the respondents’ individual variables and attitudes towards welfare regimes.

Background: Why study ADM from a citizens’ perspective?

Research on ADM has flourished in recent years (Araujo et al., 2020; Kuziemski & Misuraca, 2020; Lomborg et al., 2023). Especially with the implementation of the General Data Protection Regulation in the European Union, we have seen a steady increase in studies moving beyond the typical fields of computer science and into the social sciences and humanities, especially by focusing on the risks and ethics of implementing AI and ADM in the public sector (Ashok et al., 2022; Madan & Ashok, 2023). This research has generally focused on questions of governance and regulation (Binns, 2018; Borgesius, 2020; Coudert, 2010; Larsson, 2020; Wagner, 2019), values and ethics (Ashok et al., 2022; Masso et al., 2023; Madan & Ashok, 2023; Veale & Edwards, 2018) and, more rarely, the people involved in the ADM implementation process (Brown et al., 2019; Dencik et al., 2018; Lomborg et al., 2023). This last strain of research has mainly focused on people who develop ADM solutions (Henriksen & Blond, 2023), workers who maintain ADM infrastructures (Wu, 2023) and labourers who work at the interfaces of ADM solutions (Ranerup & Svensson, 2023).

Increasing researchers have focused on ADM use in the welfare and public sectors (Eubanks, 2018; Henriksen & Blond, 2023; Reutter, 2022) – particularly the shifts in discretion of case workers and their agencies in relation to ADM systems. Furthermore, studies have highlighted the socio-technical imaginaries of policymakers, the public sector and the tech businesses that cater to the public sector (Hockenhull & Cohn, 2021; Jørgensen & Søe, 2023).

Some research has continued earlier work on technology-related social change by examining citizens’ perspectives on ADM systems. For example, Helberger et al. (2020) explored attitudes towards fairness and ADM among citizens in the Netherlands. They found that most participants perceived ADM as fairer than human decision-making. However, there are important differences between age groups and educational levels in perceptions of the fairness of AI (Helberger et al., 2020). Furthermore, respondents commonly acknowledge the role of programmers in ensuring fairness. Araujo et al. (2020) examined citizens’ preferences regarding the governmental use of AI and reviewed studies that have engaged with what they call algorithmic appreciation by highlighting citizens’ perceptions of the issues beyond mere usefulness.

Hence, although the field of ADM research has been thriving, there is still a lack of research that takes an explicitly citizen-focused approach. There is also a dearth of comparative research across countries and domains that makes visible the context-specific factors of ADM use in the public sector. Therefore, based on an original survey instrument, this article engages with citizens’ awareness of, attitudes toward and experiences with ADM in the welfare sector from a comparative perspective (Hagendorff, 2020).

As we have seen in this literature review, there has been a strong focus on universal hard and soft laws that regulate and govern ADM in the public sector. Previous research has also shown that citizens distinguish between the different purposes for which ADM is deployed and domains it is deployed in (Kaun et al., 2024), which can have a crucial impact on how the public agencies and state representatives that implement ADM are perceived – for example, in terms of trust. If we further establish ADM as a political and not merely managerial topic (Reutter, 2022), citizen perspectives are crucial for understanding the further expansion of the digital welfare state and how ADM should be governed.

Theoretical approach

This study’s design was chosen to explore how the experiences of different welfare-state regimes might explain differences in attitudes towards ADM in the public sector. Following and extending an existing typology of welfare state regimes (Esping-Andersen, 1990, 2015), the chosen countries represent a social-democratic welfare state model (Sweden), a corporatist-statist welfare state model (Germany) and a ‘new democracy’ welfare state model (Estonia).

Comparative welfare research has a long tradition. While earlier studies followed Esping-Andersen’s distinction between welfare regimes based on differences in the principles upon which policies are implemented, including ideas of solidarity and equality, as well as on the relationship between welfare provision and the market (Pfau-Effinger, 2005), more recent studies have increasingly integrated citizens’ attitudes towards welfare policies (e.g., Van Hootegem et al., 2021; Svallfors, 1997) and considered cultural and ideological factors in the emergence of welfare regimes (e.g., Pfau-Effinger, 2005). Lim (2020) showed that citizens who embrace technological change often also support welfare measures, including unemployment insurance. Hence, there is an argument to be made for the social embedding of technological innovation, as the welfare state might foster positive attitudes towards new technologies. Thus, we studied three countries that use AI and automated decision-making in the public sector to different degrees. While Estonia and Sweden commonly implement digitalisation and use AI in general (Charles, 2009; Lember et al., 2018), Germany is less advanced in this regard. This theoretical reasoning is reflected in the variables tested, which we detail below.

Method

The analysis is based on a survey conducted in Estonia, Germany and Sweden between 18 October and 9 November 2021 by the market research company Kantar Sifo, which has departments in these three countries. The survey was distributed to those aged 18–75 years, including 10,118 people in Estonia, 12,506 in Germany and 6,083 in Sweden. The response rates were 15% in Estonia (n = 1,500), 16% in Germany (n = 2,001) and 17% in Sweden (n = 1,000), which correspond to the typical response rates in online panels (Pedersen & Nielsen, 2016), as opposed to face-to-face surveys (Szolnoki & Hoffmann, 2013) or studies with paid crowdsourced respondents (Eklund et al., 2019). Additionally, the Estonian sample was expanded by 500 participants to ensure that the sample represented the two main national groups, Estonian and Russian speakers (see Table 1). In sum, while the response rates discussed above cannot be considered high, we argue that they are satisfactory relative to the rates typically reported for this type of survey (e.g., Nulty, 2008).

Table 1. Sample structure of the survey (%).

		Estonia	Germany	Sweden	Total
Gender	Male	51	49	51	50
	Female	49	52	49	50
Age	18–34	26	39	39	34
	35–49	37	38	29	35
	59–75	37	23	32	31
Total	%	100	100	100	100
	N	1500	2001	1000	4501

Our analysis primarily focused on three scenarios presented to the survey participants in their local languages, which were then translated for publication purposes. Vignettes and scenarios are often used in qualitative and quantitative research to engage with experiences, especially those of vulnerable populations (Cheah et al., 2023), and to explore sensitive topics (Aujla, 2020). As Aujla (2020, no page number) argues, ‘The vignette, also known as a scenario or situation, is a short story with hypothetical characters used in both quantitative and qualitative studies to elicit participants’ perspectives on difficult topics’. While previous methodological discussions of vignettes and scenarios’ strengths and weaknesses have primarily focused on how they encourage research participants to approach topics that are difficult to talk about in a depersonalised and open manner (e.g., Schoenberg & Ravdal, 2000), we employed the method to trigger attitudes towards technologies that are often perceived and presented as black boxes that are hard to understand. Through specific and relatable scenarios, we encouraged our survey participants to take a specific position on ADM systems while opening up the purported black box of algorithms in the welfare sector.

Vignettes can take the shape of visual materials, including pictures, figures or drawings, as well as textual forms, including fictious and real-life stories. In quantitative research, vignettes and scenarios are often used in experimental settings – for example, in the constant-variable-vignette method detailed by Wason et al. (2002). This method is designed to examine, for example, respondents’ moral judgments of sensitive issues, such as immoral behaviour (Cheah et al., 2023). We worked with text-based scenarios designed to be plausible and similar to real-world situations while not exactly representing existing cases (see the scenario description in Appendix 1).

The scenarios were tested in a qualitative pretest with four Swedish participants. The participants engaged with the vignette descriptions and were then interviewed about the clarity of the descriptions. Based on the interviews, the formulations of the vignettes were slightly adjusted. Before moving into the field, the whole questionnaire was pretested with a small subsample and deemed robust. Based on the pretests and the results from the survey, we consider vignettes and scenario-based questions fruitful means of exploring attitudes towards complex technological systems.

The scenarios developed reflect established discussions of ADM systems in public sector domains related to core welfare services, such as social benefits and child welfare, as well as the arena of control (policing). Hence, they constitute high-risk domains with special importance in terms of their implications for vulnerable populations. The exact figures for failure rates and models on which the risk scores are based, however, are fictitious. Moreover, they are not explicitly placed in a national context. The aim was to provide scenarios that would be realistic enough but not represent specific or potentially controversial cases of ADM use.

The first scenario concerns the use of ADM by unemployment services. The scenario describes the proposed automation of the evaluation of hiring potential among unemployed people, with answers solicited using a 20-item questionnaire. Based on the scores, the job seekers were divided into different groups and assigned different kinds of training and support accordingly. We refer to this as the employment services scenario.

The second scenario focuses on the prediction of child welfare – specifically, the probability that a child will be exposed to harm in their family. The ADM system assigns a risk score based on variables encompassing the background of the family (e.g., history of harmful behaviour and drug abuse, educational and income level) as well as their housing situation (area and crime rates in the neighbourhood). The calculated risk score would automatically trigger specific remedies, including a home visit and interview with social workers. This is referred to in the remainder of the paper as our social services scenario.

The third scenario is concerned with predictive policing with the help of facial recognition. People who trigger a hit in the database would be automatically stopped, searched and questioned. However, the system has a comparatively high rate of false positives. The respondents were queried about their perception of whether and how the system should be deployed. We refer to this as our predictive policing scenario.

After each scenario description, we provided our respondents with three different statements about each scenario and asked them to express their attitudes by way of a four-point scale (1 = completely disagree, 4 = completely agree). For each of the three scenarios, and in each of the three countries, the internal consistency of the three specific response items was determined acceptable, with Cronbach’s alpha values ranging from .69 to .73 (Hair, 2010). With satisfactory consistency in place, we created three separate indices – one for each scenario – by dividing the summed scores for the response items in each scenario by the total number of response items for each scenario. Much like the method described above, these scenario indices ranged from 1 to 4 and formed the basis for our dependent variables.

The analysis was performed in two ways. First, we assumed a high-level perspective, comparing the means of the three scenario indices across the three studied countries. We report these results through a series of boxplots combined with jitter plots. Beyond allowing us to assess the central tendencies across indices and countries, this mode of presentation provides us with insights into the spread of our respondents (each represented as differently coloured dots in the forthcoming three figures) across the index scales. The boxplots are further complemented with black dots that indicate the mean for each country on each scale. Combining traditional boxplots – which represent the distribution of the data, the outliers and the medians, with points indicating the mean, and with jittered points indicating the spread of respondents across the scale – allows for a deeper understanding of the basic structure of our data.

Second, we assessed the influence of a series of independent variables on the aforementioned scenario indices through a series of multiple regression analyses. Given the comparably novel and still-emerging phenomenon of ADM in our three case countries, we opted for a more open, exploratory approach (as suggested by Stebbins, 2001) and assessed the influence of a series of independent variables on our indices, as discussed above. Specifically, we separate our analyses by country and begin our exploration by including the following sociodemographic measures:

• Gender: Represented as a dichotomy in which 0 = female and 1 = male.

• Age: In years.

• Education: To more fully represent the differing educational systems of the three countries, the German respondents were asked to gauge their level of education on an ordinal scale from 1–7, while the Estonian and Swedish respondents did the same using an ordinal scale from 1–5.

Beyond these sociodemographic variables, we also included three indices as independent variables. These indices were first introduced in a previous research paper (please see Kaun et al. 2024 for full details) and are described below.

Awareness index: The respondents were prompted by three statements regarding their rights as providers of data to be used in ADM. Employing a three-point scale (1 = not aware, 2 = somewhat aware, 3 = very aware), we asked the respondents to assess the degree to which they were aware of their rights as specified by the statements. Cronbach’s alpha for the three statements was calculated at .84. The index itself was constructed by summing the scores for each respondent and dividing that score by the number of included measures.

Suitability index: We included six items to test the degree to which the respondents agreed that ADM would be a suitable administrative technique in different settings. Employing a five-point Likert scale (1 = totally disagree, 5 = totally agree), we asked the respondents about their feelings about the use of ADM in a series of contexts, such as state administration, school supervision, municipal administration and health care. The Cronbach’s alpha value for the six items was .86. As in the previous indices, the items were summed, and the total was divided by the number of items, creating what we refer to as our suitability index.

Trust index: Given the role of ADM in diverse societal institutions, we included a series of items to gauge respondents’ trust in such entities. A series of five-point Likert scales were used (1 = don’t trust at all, 5 = completely trust) for the respondents to gauge their trust in institutions such as the state, the government, parliament, the police and the court system. Cronbach’s alpha for the full set of trust variables was .911. As with the previous indices, the final trust index was constructed by summing the scores on the items and dividing the result by the number of items included Figures 1–3.

Figure 1. Employment services scenario index values. A higher score indicates a more positive attitude towards the scenario.

Figure 2. Social services scenario index values. A higher score indicates a more positive attitude towards the scenario.

Figure 3. Predictive policing scenario index values. A higher score indicates a more positive attitude towards the scenario.

Results

For the employment services scenario, the German respondents’ attitudes emerged as more positive (mean [M] = 2.47, standard deviation [SD] = .6) than those of their Swedish (M = 2.35, SD = .64) and Estonian (M = 2.27, SD = .67) counterparts. A Kruskal – Wallis H test indicated significant differences between the reported means (p < .000), and Dunn’s test revealed that all means were significantly different from each other (p < .000 across all comparisons). While significant, the effect size as measured by means of a Kruskal–Wallis effect size test (e.g., Kassambara, 2023) emerged as .03, which is small if we follow the guidelines suggested by Cohen (1988). With this caveat in mind, the results nevertheless indicate that our German respondents thought of ADM in employment services as more efficient, objective and just compared to their Estonian and Swedish counterparts.

In the social services scenario, the Swedish respondents had the most positive attitude (M = 2.9, SD = .65) towards the child welfare scenario. The Estonian respondents were not quite as enthusiastic in this regard (M = 2.85, SD = .71), and our German respondents expressed limited agreement on average (M = 2.79, SD = .66). A Kruskal – Wallis H test indicated significant differences between the reported means (p < .000), and Dunn’s test revealed that, while the difference between Sweden and Estonia failed to reach statistical significance (p = .005), significance was reached when comparing the index value reported for Germany to the values from the other countries (p < .000 across all such comparisons). We calculated the effect sizes following the rationale previously discussed, once again finding what must be considered a small effect size (.01). Nevertheless, compared to our German respondents, the results indicate that our Swedish and Estonian respondents considered the use of ADM in child welfare protection to be more helpful, equally as efficient and more just.

Finally, for our predictive policing scenario, the Estonian respondents proved more positive (M = 2.83, SD = .58) than their German (M = 2.71, SD = .62) and Swedish (M = 2.67, SD = .67) counterparts. A Kruskal – Wallis H test indicated significant differences between the reported means (p < .000), and Dunn’s test specified that, while the mean index value for Estonia was found to be significantly different from the values reported for Germany and Sweden (p < .000 across both comparisons), the difference between Germany and Sweden in this regard emerged as insignificant (p = .418). Once again, we see small effect sizes (.01), suggesting further explanatory potential beyond the country comparisons provided here. With this caveat, our Estonian respondents considered the use of facial recognition in crime prevention to be more efficient and objective, and they were less concerned about false accusations.

We next assess the influence of the independent variables on the scenario indices mentioned above. The regressions are presented visually, with green and red indicating positive and negative coefficients, respectively.

While the variable of gender did not yield any significant results, age emerged as a positive significant predictor across all countries, suggesting that comparably older respondents would be more positive towards the employment services scenario. Similarly, the influence of education, which was significant but negative, suggests that across all countries, respondents with higher levels of education are likely to be more sceptical about the specified scenario. Being aware of citizen rights in relation to ADM does not appear to have had any significant influence on the employment services scenario index; perhaps unsurprisingly, the opposite can be said about the estimates reported for the suitability index. Finally, the trust index, which gauged the degree to which the respondents felt that they can trust different societal actors, was negative for all three countries and significant for Germany and Estonia. Granting the insignificant result emanating from the Swedish data, this finding nevertheless suggests that those respondents who expressed higher levels of institutional trust were more likely to feel sceptical about our employment services scenario.

While gender did not emerge as a significant predictor of attitudes towards our employment services scenario, as shown in Figure 4, it yielded negative (all countries) and significant (for Germany and Estonia) coefficients in relation to the social services scenario, as shown in Figure 5. Given the way this variable was coded, this result suggests that male respondents tended to be less positive about this scenario than their female counterparts. Age, then, produced a result similar to the one displayed in Figure 4; save for Sweden, we see significant estimates across all countries, again showing that older respondents expressed more positive attitudes towards the social service scenario as well as the employment services scenario.

Figure 4. Results of a multiple linear regression conducted to predict index values in the employment services scenario. R2 (Adj. R2)  – DE: .14 (.13), EST: .28 (.28), SE: .23 (.22). * = p ≤ 0.05, ** = p ≤ 0.01, *** = p ≤ 0.001.

Figure 5. Results of a multiple linear regression conducted to predict index values in the employment services scenario. R2 (Adj. R2)  – DE: .25 (.24), EST: .24 (.23), SE: .18 (.17). * = p ≤ 0.05, ** = p ≤ 0.01, *** = p ≤ 0.001

The education level variable failed to reach significance. While the awareness index behaved in the same way as in the employment services scenario (Figure 4), a different result can be discerned in relation to the social services scenario, as displayed in Figure 5. Specifically, the awareness index emerged as negative (for all countries) and significant (for Germany and Estonia), suggesting that the more concerned the respondents about their rights in relation to ADM, the more negative their attitudes about the social services scenario. The two remaining variables in Figure 5 produce results similar to those displayed in Figure 4, showing that those respondents who expressed enthusiasm towards ADM efforts were more positive towards the scenario discussed here, while those with greater trust in societal institutions are likely to be less enthusiastic in this regard.¹

While the last two scenarios saw age emerge as a positive and mostly significant predictor across all three countries, age in the predictive policing scenario functioned as a negative significant predictor in two out of the three case countries. Thus, in Estonia and Sweden, older respondents appear to be more sceptical towards predictive policing than their younger counterparts Figure 6.

Figure 6. Results of a multiple linear regression conducted to predict index values in the predictive policing scenario. R2 (Adj. R2)  – DE: .15 (.14), EST: .14 (.13), SE: .11 (.10). * = p ≤ 0.05, ** = p ≤ 0.01, *** = p ≤ 0.001.

The results for the education variable provide another interesting discrepancy when compared to the estimates reported for previous scenarios. Specifically, while more education resulted in more negative attitudes towards the employment services scenario, education in relation to predictive policing had a positive effect across all countries – but only reached significance in the German and Swedish cases. As such, German and Swedish respondents with higher education levels emerged as more positive in this regard than those respondents with less education. The awareness index emerges as negative for all countries – significantly so for Sweden – suggesting that across all three scenarios, higher awareness in the sense defined above leads to lower index scores. As above, the remaining two independent variables appear to function in the same way: finding ADM suitable leads to higher scenario index scores, while expressing trust in societal institutions leads to lower scenario index scores.

Discussion

Our analysis points to important differences between the three countries when it comes to attitudes towards ADM in three different domains. This finding confirms our initial theoretical expectations, which were developed based on previous studies. Each country came out on top for one scenario: Germany for the job seeker categorisation scenario, Sweden for social services and Estonia for predictive policing. Hence, there seem to be important domain-specific differences in citizen evaluations of ADM between welfare regimes. While German respondents were more positive towards automation in employment services, Swedes saw an opportunity to prevent harm to children, and Estonians felt best about using ADM to prevent crimes. These findings can be explained by specific welfare structures in the three countries, which manifest in attitudes towards employment automation, social welfare and crime prevention.

In Germany, there is a relatively long tradition within public discussions about the social norms relevant to implementing digital solutions in the public sector (Schmidt & Weichert, 2012). Furthermore, a corporatist-statist welfare regime will be less focused on distributive logic than a social-democratic regime. ADM – we assume – is hence perceived by citizens as likely to be used primarily for the purposes of controlling and surveilling welfare distribution rather than actually providing care. This potential concern about surveillance and control through ADM systems is reflected in the stronger preference for their use in a less invasive scenario with milder existential implications – namely, the job seeker scenario. In contrast, the Swedish respondents’ preference for ADM to be used in the social service scenario may be linked to the historically strong regard for children’s rights in this nation. In addition, Sweden represents a social-democratic welfare regime with historically expansive welfare institutions, including social services, that have earned high trust among the citizens.

Estonians are clearly the most enthusiastic about ADM’s use in predictive policing. The relatively supportive general attitudes towards ADM and predictive policing in Estonia can be explained by the fact that the use of algorithmic systems (Männiste & Masso, 2018, 2020) and other digital technologies (Tammpuu & Masso, 2018) in public administration plays a significant role in the nation’s identity and branding (Masso et al., 2020). Therefore, the results of this study suggest that, in Estonia, ADM technology could become a source of renewed trust in public administration. This technology is perceived as linked to the values of fairness and justice, which can be considered especially crucial for deterring crime and conferring punishment. Moreover, ADM could significantly shape the meanings and understanding of crime, security and safety in Estonia, contributing to the formation of security-related social norms and values.

Regarding the individual factors and their significance in diverging attitudes, gender plays an interesting role. While gender played an insignificant role in the unemployment and predictive policing scenarios, it emerged as a negative (for all countries) and significant (for Germany and Estonia) predictor in the social services scenario. This suggests that male respondents were more sceptical about this scenario than their female counterparts. Here, we can speculate that women still often have the main care responsibility for children, especially in Germany and Estonia (as compared to Sweden), in the family and beyond. Thus, perhaps they see potential for ADM to be used to identify parents and families in need of support.

Turning to age, a higher age largely led to higher values in the scenario indices – except with predictive policing, where the reverse was true. This means that the older the respondents, the more negative their attitude towards predictive policing in all three countries. We assume here that the age effect changes when respondents are confronted with a more uncertain and intrusive scenario. Whereas the meaning and public understanding of ‘employment’ and ‘social services’ have been more stable over time compared to those of 'police work' (Masso et al. 2024), these shifting meanings might be visible in the comparison of different generations.

In contrast, education was mainly a negative predictor in the employment and social service scenarios – the more education, the smaller the index value. This means that highly educated respondents did not view these scenarios as suitable. For the predictive policing scenario, however, we see a different picture. Education here was a positive predictor, which means that the higher the educational level, the higher the index value. We can speculate that highly educated respondents, who presumably have stronger digital skills, do not see themselves as being at risk of wrongful examination by the police.

Hence, we suggest that attitudes towards predictive policing might reveal class-based differences among our respondents. What is consistent across the scenarios and welfare regimes is the influence of trust. The more trust our respondents expressed for societal institutions, the more negative they were towards the ADM scenarios. Hence, trust is important across domains and welfare regimes (Kaun et al. 2024).

Conclusion

The data point to important differences in welfare regimes and domains within which ADM is deployed. These divergences are especially relevant to governing and regulating ADM applications based on citizen preferences. Citizens make important distinctions between ADM technologies and the domains in which they are implemented. These differences in attitudes depend both on the specific welfare regime and individual factors such as age, gender, education and societal trust. Hence, our findings show that we need to combine individual and structural factors to explain the different attitudes towards ADM discovered in digital welfare research. The results also indicate the need for domain-specific approaches to ADM. This combination of structural and individual variables demands more attention in future research, both within fields that focus specifically on ADM and in comparative welfare research.

Our study has important limitations that should be mitigated in future studies. First, the focus was on three domains that have been broadly discussed in the public. More mundane cases and scenarios would broaden the empirical basis for studying diverging attitudes towards ADM systems. Second, population-representative survey studies come with specific methodological constraints, including openness towards unexpected attitudes and experiences. Similarly, much like other survey-based research, our study had rather low response rates. While this could be seen as a validity issue, these response rates are typical of surveys, as noted above, and we think of our findings as a starting point for further scrutiny rather than an end point in themselves. For instance, future studies engaging with differences in citizens’ experiences should combine qualitative and quantitative approaches to better understand how and why citizens’ attitudes and experiences differ. Third, future research should include more contextual variables that operationalise the different welfare regimes – for example, the range of social welfare policies, universal access to welfare and expenditures for welfare in relation to BNP. This would allow us to update earlier typologies of welfare systems.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by Östersjöstiftelsen.

Notes on contributors

Anne Kaun

Anne Kaun is Professor at the Department of Media and Communication Studies, Södertörn University, Sweden, and a Wallenberg Academy Fellow studying the democratic implications of automated decisions making, artificial intelligence and digitalisation more generally in the welfare sector [email: [email protected]].

Anders Olof Larsson

Anders Olof Larsson is Professor at the Department of Communication, Kristiania University College, Norway. He is studying the use of online interactivity and social media by societal institutions and their audiences, journalism studies, political communication and methodology, especially quantitative and computational methods.

Anu Masso

Anu Masso is Associate Professor at the Ragnar Nurkse Department of Innovation and Governance, Tallinn University of Technology, Estonia. Her research interests include big data, social datafication, spatial mobility, social diversity, algorithmic governance, data justice, and research methods.

Notes

1 A positive coefficient indicates that as the value of the independent variable increases, the mean of the dependent variable also tends to increase. A negative coefficient suggests that as the independent variable increases, the dependent variable tends to decrease.

 

Appendix

Scenario Items

Each scenario consists of three items. In each of the scenarios, two of the items are positively phrased, meaning that if the respondent provided a high score, this would indicate a positive stance towards that item. The third item in each scenario was negatively phrased, meaning that if the respondent provided a high score, this would indicate a negative stance towards that item. Before the analysis, the negative items were reverse coded to facilitate index construction.

Scenario 1

A civil servant has the task of making a decision about the placement of a job seeker in different programme options for job market training. S/he is making the decision after one face-to-face meeting that included answering a 20-item questionnaire. Public employment services are now considering automating the process and sorting job seekers automatically after they have filled out a survey at home, scoring the seekers based on their replies without them ever meeting a civil servant. The job seekers will undergo specific training programmes based on their scores. To what extent do you agree with the following statements? Please mark one answer per line.

Table

	Completely agree	Rather agree	Rather disagree	Do not agree at all	Don’t know, difficult to say
1. Sorting job seekers automatically will save on costs	4	3	2	1	88
2. Sorting job seekers automatically will be more objective	4	3	2	1	88
3. Sorting job seekers automatically is biased and unfair	4	3	2	1	88

Scenario 2

A municipality uses an automated system to calculate the probability that a child will be exposed to harm in a family. If a family is assigned a high-risk score, social services will interview the parents and evaluate the family situation. The calculations are based on a) individual characteristics, such as level of education, income and history of harmful behaviour in the family, as well as drug abuse, and b) the living area and crime rates in that area. The predictions are 75 percent accurate. This means that in 25 out of 100 cases, the predictions are wrong, and either a family is falsely accused of neglect, or a child is not detected as being in the risk zone although they are harmed. To what extent do you agree with the following statements? Please mark one answer per line.

Table

	Completely agree	Rather agree	Rather disagree	Do not agree at all	Don’t know, difficult to say
1. Risk scoring like this will help prevent harm	4	3	2	1	88
2. Harm prevention will be easier and more efficient	4	3	2	1	88
3. Families that have not done anything wrong will be falsely accused of harming their children	4	3	2	1	88

Scenario 3

The police are deploying algorithms to predict where crimes might occur in the future. With the help of data analytics based on online activities as well as AI-based facial recognition, the police aim to identify criminals and prevent crime before it happens. The predictions are 50 percent accurate. This means that in 50 out of 100 cases, the predictions are wrong, and a person is either falsely accused of committing a crime or not identified as a criminal. A person without a criminal background or prior history is stopped in the street of his local neighbourhood because the facial recognition AI matched him with the criminal database. He is arrested and interviewed for several hours before he is released. To what extent do you agree with the following statements? Please mark one answer per line.

Table

	Completely agree	Rather agree	Rather disagree	Do not agree at all	Don’t know, difficult to say
1. The use of AI for facial recognition is a good way to prevent crime	4	3	2	1	88
2. People who are identified using AI facial recognition should always be thoroughly investigated	4	3	2	1	88
3. If people are wrongly identified and scrutinised, they should be compensated	4	3	2	1	88