Policy Significance Statement

This work’s policy implications are for decision-makers and practitioners of data-driven innovation in the public sector committed to ethical practice. We join those who argue for flexible and reasonable policymaking that considers the context. We suggest that a virtue-based approach can support their efforts to connect abstract and unchanging principles with contextual and shifting demands of day-to-day practice.

2. Virtue and the Smart City

We next briefly describe our understanding of two key concepts of Aristotle’s ethics,Footnote ² virtue and phronêsis, which will help us sketch out a theoretical frame to translate ethical principles into more practical terms.

Our choice to apply Aristotelian language and concepts to the ethics of innovation and technology is by no means unprecedented. Virtue ethical approaches to these domains have been explored and convincingly defended in recent years, especially by Vallor (Reference Vallor2016) and others (such as Blok et al., Reference Blok, Gremmen and Wesselink2015; Sand, Reference Sand2018; Barford, Reference Barford2019; Costello, Reference Costello2019; Reijers, Reference Reijers2020; Frigo et al., Reference Frigo, Marthaler, Albers, Ott and Hillerbrand2021; Astola et al., Reference Astola, Bombaerts, Spahn and Royakkers2022). As Barford summarizes: “the human qualities studied by virtue ethics—e.g., personal character and habits of acting and choosing with practical wisdom so as to promote human flourishing and not only one’s personal goals or those of one’s employer, to judge and to prioritize—are required successfully to engage in values-based ethical design [of ICT systems],” p. 1.

Having said that, there is a frequent criticism leveled at virtue ethical theories regarding their alleged inability to effectively guide action (see, e.g., Louden, Reference Louden, Crisp and Slote1997). However, as Sand (Reference Sand2018) responds to this criticism, virtue ethics can successfully contribute to addressing problems of applied ethics by generating moral knowledge. Further, as Sand also points out, effective action guidance is a more general problem for all normative theories when applied to wicked problems and not specific to virtue theories. Therefore we, alongside Annas (Reference Annas2011) and Vallor (Reference Vallor2016), find it useful to apply analysis inspired by virtue ethics in the context of technological innovation.

In Aristotelian thinking, virtue is a deliberate practice—a way of acting— that enables its possessor consistently to excel in their function. While virtues are more than mere “simple dispositions to engage in certain behaviors stereotypical of the virtues” (Russell, Reference Russell and Polansky2014, p. 203), it is critical to note that they do come about as the result of habituation (êthos) or “the repeated performance and practice of the actions typical of the virtues” (Basilio, Reference Basilio2021, p. 531). Further, there is a distinction between virtues of character and intellectual virtues. The exercise of specifically virtues of character is what enables one to perform one’s function well, while the exercise of intellectual virtue enables one to arrive at the truth. It is important to stress here that this notion of virtue is one crucially involving activity and not of only being (explored by, e.g., Broadie, Reference Broadie1993). Merely possessing a virtue is not sufficient, rather it must be accompanied by activity.

Being virtuous in one’s practice is a deliberate balance. For Aristotle, virtue of character is a mean state that is destroyed by the extremes of deficiency or excess. For example, courage is the virtue (mean) that lies at the right distance from foolhardiness (excess) and cowardice (deficiency). It refers to the right balance of drive and caution appropriate for the situation at hand. Virtue does not necessarily mean being equidistant from the extremes, and the appropriate position shifts from situation to situation. It is a matter of judgement, or what Aristotle refers to as the intellectual virtue of phronêsis, practical reason, to determine the appropriate balance or mean between deficiency and excess for one’s specific situation.

A city’s ethical principles for data or AI use can be understood as representations of the (aspirational) virtues of the smart city. The successful exercise of phronêsis can then be understood as finding the right balance and a way to apply the abstract ethical principles in the concrete practices and processes. So understood, it becomes imperative for our aim—which is to bridge the gap between principle and practice—to consider in more detail the exercise of phronêsis, to which task we turn next through employing a dialogic case study.

3. Dialogic Case Study Approach

Our goal in this paper is to illustrate how a virtue-based approach can bridge the gap between high-level principles for the city as a whole with the hands-on work of data scientists and others as they engage in the practices of data-driven innovation. In what follows, we aim to illustrate how virtues operate on a practical level, namely with regard to the individuals—in our cases, data scientists—whose actions make up parts of the practices that are data-driven innovation. To study this, we explore two data-driven innovation cases being developed by the data and analytics team in the city of Helsinki.

To study these cases, we use a bidirectional dialogic case study approach (Rule and John, Reference Rule and John2015). Our purpose here is to comment on the theory using the two concrete example cases and more generally to show how to bridge the gap between theory and practice. With this approach, the cases and theory engage in two ways: by both building the theory from the cases and by testing it with the cases—enabling a two-way commentary.

The first case involves advanced data analytics developed to provide better quality services for the long-term unemployed. The context for this case is a large pilot project in Finland in which the responsibility of employment services is being transitioned from a national agency to the level of individual municipal governments.Footnote ³ This case involves personal data collected about individuals and their use of employment and other services provided by the city. The purpose of the algorithm developed is to find a personalized path for job-seekers towards employment or other occupation. The innovative aspect in this case is in aiming to include, in addition to simple employment services use data, also data from complementary sources that the city controls. This combination would allow for more personalized and predictive service provision, with the stated goals including better served city residents.

The second case involves developing a machine learning algorithm to enable third-party event organizers more effectively to connect with residents and visitors of the city. This case involves event description data and the purpose of the algorithm in this case is to automatically and meaningfully classify and categories the events. The innovative aspect here is that, instead of a more typical big data approach where access to resources largely determines success, the aim is more equitable and non-discriminatory treatment of both the two groups, event organizers, and potential audiences. For example, on the organizer side the goal includes providing equal visibility to different parties regardless of their characteristics such as size of marketing budget. Meanwhile, on the audience side, the goal includes deliberate exposure to events––search results and recommendations––that is, accurately tailored to an individual’s specific interests given that those interests may appear as negligible niche outliers in a mass data analysis approach and, as a result, get filtered out of view.

For the purposes of this theoretical commentary, we relied on a single informant in the city of Helsinki, a data scientist representing the project coordination and data science perspectives, with whom we had three, approximately 90-min interview sessions. We also used project documentation and publicly available online materials to gather additional perspectives. In addition, the authors are former employees of the city of Helsinki with first-hand experience and knowledge of the data governance, management, strategy, and data policy perspectives both in general and regarding these specific cases.Footnote ⁴

To analyze the case materials, we used an iterative approach. In between each interview session, we re-examined our case materials and iteratively developed an updated synthesis bridging the practices evidenced in cases and theories. In particular, we analyzed the examples of algorithmic choices described by the informant and mapped them using Vallor’s (Reference Vallor2016) technomoral virtues. We then engaged again in a dialogue with the informant to gather additional information and missing perspectives. After three iterations, our analysis saturated and resulted in the identification of the four common themes as presented in Table 1.

Table 1.

Four data-driven innovation virtues linking technomoral virtues and technological choices

It is important to note here that cities are typically rather conservative organizations and are also highly regulated by national and/or European Union legislation. This is especially the case with regard to the use of (personal) data and its combination from distinct registers. While our chosen cases may at first sound like fairly simple or ordinary data science projects, they in reality require novel and innovative approaches involving technologies, service design, legal interpretation, recruitment strategy, procurement standards, overall policy, and management in the city government.

4. Phronêsis for the Data Scientist

We next turn into the results of our study. Overall, we find that data scientists, as practitioners of data-driven innovation, must engage their practical reason, phronêsis, to make ethically charged decisions or, in the language of virtue, to find the appropriate balance between deficiency and excess, in the context of those decisions.

Merely exercising one’s phronêsis does not, of course, guarantee optimal or even good results, nor is it intended to do so. Further, at times individuals may be constrained by their institutional environment in a way that hinders acting upon what one deems appropriate. While phronêsis is helpful when speaking about the ethics of individuals, more work is needed to understand the ethics of organizations that comprise more than just individuals (for a similar conclusion, see Mittelstadt, Reference Mittelstadt2019, p. 555.)

Analysis of the two cases studied led to the identification of four common themes, described as examples of translational or bridging virtues of data-driven innovation (see the middle column of Table 1). These virtues can be considered as specific instances of one or more of Vallor’s more general technomoral virtues, which are mentioned in the leftmost column. Vallor’s still fairly generic virtues turned out to be most useful as an approximation and a stand-in for a set of data ethics principles at the city. Finally, via these more specific, bridging virtues of data-driven innovation, we are able to make much more direct and immediate connections to the particular technical and algorithmic choices that the data scientist faces on a day-to-day basis (visually the middle column is the bridge between the two ends). Examples of these choices present in our cases are given in the right-most column.

It bears emphasizing again here that each row of this table represents an illustrative example of the kind of bridge we seek, and each cell in the middle column likewise is an example that we happened upon in the course of this particular research effort. Neither the set of rows or the set of what we call “virtues of data-driven innovation” in the middle column are to be considered exhaustive or comprehensive––nor indeed are the virtues mentioned not in need of (further) moral justification. Our purpose here is only to visualize the missing layer between principles and practices and to show how the space between the leftmost and rightmost columns might be filled out in the instance of this specific data scientist and these two specific cases. Next, we briefly discuss each of the four virtues of data-driven innovation mentioned in the middle column.

4.4. Service to individuals in society

When designing and developing public services, one needs to weigh the interests of the individual against those of society or larger community. For a public-sector body, resources are limited and individually tailored care and services are not often possible within the bounds of those resources. The task in these cases is to ensure that most people are adequately served while no one, regardless of the complexity of their needs, is left behind.

In terms of the general technomoral virtues, this relates to holding on to the moral whole (perspective), in that one must retain the big picture of societal well-being while caring for specific people and their needs, and skillful adaptation to change (flexibility), in that the right approach to any given case will have to be dynamic are react to changes in the environment and society as well as emerging individual needs.

Excess in this virtue is a kind of erasure of individual experience and difference, while deficiency is prioritizing the individual above all other concerns. For this virtue, which could be described as the right consideration for the collective with respect to the individual and vice versa, we did not observe a single common mechanism for the data scientist to deploy in exercising their phonêsis as was the case for the three preceding virtues. We did, however, observe two particular actions taken.

The first was to parametrize the loss functionFootnote ⁵ in a clustering algorithm (see the code excerpt from our case in Figure 1). By choosing different values for the parameters, a data scientist can balance between a preference for larger or smaller clusters. In practice, preferring larger clusters means providing a more generic outcome that approximates the need of more people at once and preferring smaller clusters produces more precisely targetable groups.

Figure 1.

Code excerpt for balancing between community and individual utility.

The second action we observed was in response to the observation that algorithms often tend to favor the general masses and that, as a result, rare and exceptional instances are not found. This is a fundamental problem in all reasoning based on statistical analysis, but it takes various forms in different contexts and so there are multiple approaches possible to address it. The action we observed was to add randomness. By increasing randomness, a data scientist can balance toward niche categories by equally suggesting some unexpected and rare instances.

To conclude this section, some summary remarks. In the course of our work, we saw that our informant was continuously exercising their phronêsis in addition to their technical skills and knowledge to be able to do their work. We conceptualize this as their activity of attempting to find the appropriate balance between deficiency and excess, that is, navigating the landscape of virtue with pronêsis. In the first case study, we noticed that a balance was being sought between sufficient anonymization of personal data of the people using employment services and the usefulness of the data in terms of information richness and utility. In the second case, the balance to be struck involved the generalizability of the ML-based classification model and how well the model finds specific solutions.