Policy Significance Statement
This research offers policymakers a new framework for analyzing and understanding how practitioners use Algorithmic Decision Tools (ADTs) to inform their decision-making. The analysis also offers some suggestions around the conditions that appear to contribute to different approaches to using ADTs. At present, the literature offers very little in terms of helping us to understand what drives different use types. This research helps to fill that gap by offering insights into the different conditions that appeared to be associated with different use-types. Understanding which conditions support a constructive human–machine interaction, and which do the opposite, is critical if these tools are going to succeed in supporting better decision-making in the public sector.
1. Main Text
Understanding how administrators make decisions is critical to understand how and why governments function as they do. The recent introduction of algorithmic decision tools (ADTs) to public sector agencies is changing how decisions are being made, particularly by Street Level Bureaucrats (SLB)—public sector workers who interact directly with citizens (Bovens and Zouridis, Reference Bovens and Zouridis2002). While there is significant research on the impact of technology on SLB discretion (Ellis, Reference Ellis2011; Hoybye-Mortensen, Reference Hoybye-Mortensen2015; Busch and Henriksen, Reference Busch and Henriksen2018), insufficient research considers how SLBs use ADTs to inform their decision processes (Green and Chen, Reference Green and Chen2019). This paper contributes to this nascent field of research by exploring how an archetypal form of SLB—social workers—are using ADTs to guide their decision-making.
Previous research in the field of algorithmic decision-making has focused on either algorithm aversion, whereby practitioners ignore ADTs (Dietvorst et al., Reference Dietvorst, Simmons and Massey2015), or automation bias, whereby practitioners defer to ADTs (Skitka et al., Reference Skitka, Mosier, Burdick and Rosenblatt2000). Yet these approaches have neglected the ways in which practitioners combine ADTs with their own intuition and expertise. This article introduces the new concept of “artificing” to better understand such situations. The introduction of artificing bridges the previously posed divide between aversion and automation bias and offers a new framework for analyzing and understanding how practitioners use ADTs to inform their decision-making.
To validate this framework, explore the prevalence of each use-type, and develop possible explanations for why different use types might be more or less common under different conditions, the author carried out in-depth interviews with practitioners across three separate ADTs—one in the United States (US), and two in the United Kingdom (UK). One of the ADTs was a “rule-based” tool, which takes a sequence of human instructions and applies those rules to analyze real-world data (Fry, Reference Fry2018). The other tools examined are Artificial Intelligence (AI) tools, meaning that tool is given a goal, fed data, given feedback, and then left to achieve the specified end goal, without being programmed to do so (Bishop, Reference Bishop2006, 2).
Interviews revealed that artificing was the most common use-type, followed by aversion, while deference to ADTs was rare. In addition, common conditions were identified, which appear to influence how practitioners interact with ADTs: (a) understanding of the tool (b) perception of human judgment (c) seeing value in the tool (d) being offered opportunities to modify the tool (e) alignment of tool with expectations. These findings highlight the nuanced ways in which ADTs interact with human intuition and expertise, as well offering insights into which conditions support, and which inhibit, different use-types. Understanding this is critical if ADTs are to succeed in supporting better decision-making in the public sector.
2. Background
Over the last decade, ADTs have been introduced to support “more timely and accurate decision-making” (Shafiq, Reference Shafiq2016) and address known shortcomings of human decision-making, for example, inconsistency (Parada et al., Reference Parada, Barnoff and Coleman2007) and bias (Kahneman and Klein, Reference Kahneman and Klein2009; Gillingham, Reference Gillingham2016).
In the field of children’s services, actuarial decision-tools have already been used for several decades (Gillingham, Reference Gillingham2011; Cuccaro-Alamin et al., Reference Cuccaro-Alamin, Foust, Vaithianathan and Putnam-Hornstein2017). What is much newer is the use of AI tools, which draw on vast quantities of data and use computational methods to “identify… patterns in datasets that, previously, were too large and complex to analyse” (Gillingham and Graham, Reference Gillingham and Graham2017) and “learn” to calculate the relationship between those variables and a defined response variable (e.g., likelihood of child maltreatment).
Despite the increasing prevalence of algorithmic tools in a public sector context (Andrews, Reference Andrews2019; Vydra and Klievink, Reference Vydra and Klievink2019; Kuziemski and Misuraca, Reference Kuziemski and Misuraca2020; Vogl et al., Reference Vogl, Seidelin, Ganesh and Bright2020), little research to date has explored how they are used by SLBs in practice.
This study focuses on three ADTs being used across four sites,Footnote 1 summarized in Table 1.
Overview of ADTs being used across four interview sites.
Abbreviation: AI, Artificial Intelligence.
3. Conceptual Framework
In 1947, Herbert Simon introduced the idea that administrators make “boundedly rational” decisions (Reference Simon1997, 88), and “satisfice rather than maximise” (Reference Simon1997, 120). That is, they analyze the relevant and accessible facts in the time they have available and then use intuition to find their way to a satisfactory, yet not necessarily optimal, decision. Simon’s work influenced the thinking of Lipsky, who suggested that SLBs rely on intuition, simplification, and routinization of tasks to manage the complex task of making decisions (Reference Lipsky2010, 83).
Practitioners in the field of children’s services represent a canonical form of SLB operating under conditions of bounded rationality. Practitioners are often making decisions under intense time pressure and with incomplete information (Kirkman and Melrose, Reference Kirkman and Melrose2014; Capatosto, Reference Capatosto2017) meaning it is not unreasonable to assume that they are satisficing (Lipsky, Reference Lipsky2010); in other words, analyzing the information that is available, in the time they have, and using that analysis, combined with their intuition, to come to a satisfactory, but not necessarily optimal, decision (Simon, Reference Simon1997, 119). This study proceeds on the assumption that SLBs satisfice.
Scholars differ on the role of intuition in satisficing behavior. Simon belongs to a school of thought known as Naturalistic Decision-Making (NDM) theorists, who argue that humans tend to rely on “expert intuition” (Kahneman and Klein, Reference Kahneman and Klein2009, 515) to satisfice. The NDM approach grew out of early research on master chess players (De Groot, Reference De Groot1978) and later decision made by firefighter commanders (Klein et al., Reference Klein, Calderwood and Clinton-Cirocco1986)—professionals who are able to win games, and successfully put out fires, without having to calculate all possible contingencies. According to NDM practitioners, humans draw on a “repertoire of patterns” (Kahneman and Klein, Reference Kahneman and Klein2009, 515), compiled over the course of their careers, to make intuitive, but highly informed and effective, decisions.
In contrast, proponents of the Heuristics and Biases (HB) school are skeptical about the notion of expert intuition (Kahneman and Klein, Reference Kahneman and Klein2009). They argue that human intuition tends to be characterized not by expertise, but by biases and heuristics (Moynihan and Lavertu, Reference Moynihan and Lavertu2012). This means that relying on intuition to guide decisions “…sometimes yield reasonable judgment and sometimes lead to severe and systematic errors” (Kahneman and Tversky, Reference Kahneman and Tversky1973).
The key difference between NDM and HB practitioners turns on whether expert intuition or HB is seen to be more dominant. NDM theorists believe that intuition is generally expert, while HB theorists believe the opposite to be true (Kahneman and Klein, Reference Kahneman and Klein2009). For this reason, NDM theorists tend to distrust algorithmic tools, while HB researchers “are predisposed to recommend the replacement of informal judgment by algorithms whenever possible” (Kahneman and Klein, Reference Kahneman and Klein2009, 523).
The introduction of ADTs in recent years has been justified, at least in part, by a desire to reduce the extent to which public administrators satisfice; that is, rely on intuition in their decision-making (Schwartz et al., Reference Schwartz, York, Nowakowski-Sims and Ramos-Hernandez2017). This is because the tools can very quickly and accurately capture, process, and spot patterns in the “buzzing, blooming confusion that constitutes the real world” taking into account hundreds of relevant factors, as opposed to the “just a few,” that humans are able to analyze (Simon, Reference Simon1997, 120). ADTs are also being embraced because studies demonstrate that algorithms tend to outperform humans in forecasting tasks (Grove et al., Reference Grove, Zald, Lebow, Snitz and Nelson2000).
However, while there is evidence that algorithms outperform humans at prediction tasks, most studies have compared pure clinical judgment to pure algorithmic judgment, rather than exploring what happens when human judgment (of which intuition forms a part) is used in conjunction with algorithmic tools. This insight, combined with the highly sensitive and complex nature of children’s services, the challenges of algorithmic bias and incomplete datasets, and—in some jurisdictions—legislative requirements,Footnote 2 mean that designers of ADTs in children’s services are not suggesting that the tools should completely replace human decision-making; instead, designers of ADTs suggest they will achieve optimal results when practitioners employ ADTs in combination with clinical judgment, and use them to augment, rather than supplant, decision-making processes (Cuccaro-Alamin et al., Reference Cuccaro-Alamin, Foust, Vaithianathan and Putnam-Hornstein2017; Celia, Reference Celia2018).
This approach means that an element of human intuition endures as part of the decision-making process, even after the introduction of algorithmic tools. The author calls this process “artificing”—the form of satisficing which persists following the introduction of ADTs.
Taking into account the discussion above, it appears that two forms of artificing likely exist: expert artificing—when insights from the tool are combined with expert intuition; and biased artificing—when insights from the tool are combined with HB.
It is also important to highlight that practitioners do not always use the tools as intended by their designers (Gillingham, Reference Gillingham2016). As such, there are likely additional ways—beyond artificing—that practitioners are using the tools. Previous literature suggests two additional use-types are likely.
First, practitioners could be largely or completely ignoring the tool; this is known as algorithm aversion (Dietvorst et al., Reference Dietvorst, Simmons and Massey2015). Second, practitioners could be deferring to the tool rather than using it to augment their own professional judgment—a phenomenon known as “automation bias” (Skitka et al., Reference Skitka, Mosier, Burdick and Rosenblatt2000). Automation bias manifests in both errors of commission (following incorrect algorithmic advice) and omission (failing to act because the algorithmic tool does not prompt the practitioner to do so) (Goddard et al., Reference Goddard, Roudsari and Wyatt2012).
4. Methodology
This was a theory-generating exploratory study, using a small number of participants. This approach privileged depth over breadth, and focused on revealing insights which are difficult to achieve with numeric data (Ospina et al., Reference Ospina, Esteve and Lee2018), although future research could build on these insights using survey approaches.
Thirteen practitioners were interviewed across four sites. As outlined in Table 1, two of the sites were actively using the tool, while two were not yet actively applying it as part of their everyday practice, meaning that practitioners’ answers are well-informed, but hypothetical. Eight interviews were conducted by phone (some video-link; some audio only), while five were conducted in person. Each interview lasted approximately 1 hr.
The interviews were designed to reveal how practitioners were using ADTs in their daily practice according to the conceptual framework outlined above. By conducting semi-structured interviews at length, the author was able to explore the complexity and nuance of both how and why practitioners were using tools in particular ways to support their decision-making (Adams, Reference Adams, Newcomer, Hatry and Wholey2015).
In addition to questions, each interview included three true-to-life vignettes, which were adapted for the different tools and were designed to offer insights into the way in which HB might play into the decision-making process. All subjects were relayed the three scenarios and asked to talk through how they would make a decision given that particular set of circumstances. This approach offers a unique way to explore how practitioners balance multiple factors when making complex decisions (Taylor, Reference Taylor2005; Stokes and Schmidt, Reference Stokes and Schmidt2012).
In order to generate insights into the conditions associated with different use-types, the author qualitatively coded all interviews to identify common conditions, and then assessed which conditions were present amongst artificing practitioners as compared with those who were ignoring or deferring to the tools.
5. Findings and Analysis
Practitioners were classified as demonstrating one or more use-types based on how they described themselves using the tool. Once classified within a category, the author analyzed the practitioner’s description of how they were using the tool as well as the factors that contributed to their approach to using the tool—whether directly stated or inferred.
In this section, the author analyses the prevalence of each use-type; describes what each looks like in practice; and offers possible explanations for why practitioners adopt different approaches to using ADTs under different conditions.
6. Artificing
Participants were classified within this category if they demonstrated a willingness to reconsider or change a decision based on the tool’s insights:
[The tool has] definitely led us to change a decision… (Tool 2, Interview 2)
Nine out of the 13 participants fell into this category.Footnote 3 While participants in this category indicated that they may change a decision because of the advice of the tool, they stressed that they would not always do so. Rather, they described taking the advice of the tool into consideration, combining it with their own judgment, and then coming to a decision based on a synthesis of that information. For example, one participant explained, “I’m not going to take what the tool says as gospel… I will explore [whatever the tool brings up], discuss it with the family and young person, and then come to a conclusion” (Tool 3b, Interview 3).
6.1. Distinguishing between expert and biased artificing
While biased artificing and expert artificing are presented as distinct use-types in the conceptual framework, in practice, these categories appear to converge in the vast majority of cases. Of the nine participants classified as “expert artificing,” seven also displayed elements of “biased artificing.” This came to light through the use of vignettes.
The most common form of bias was recall bias—where people allow their judgment to be strongly influenced by their ability to recall a similar example or event (Kirkman and Melrose, Reference Kirkman and Melrose2014). Participants were presented with the third vignette which describes the case of a young boy who is identified by the tool as low risk, apart from some concerns about poor school attendance. Practitioners were told that the previous week an almost identical case had been incorrectly assigned for no further action (the child was, in fact, being abused) and asked what they would do when presented with the current case.
A number of participants acknowledged that, even though the tool was identifying the child as low risk, they would treat it with greater caution as a result of the previous week’s case:
Would I consider what happened previously? Yes. You do. You do naturally. (Tool 3b, Interview 2)
While it is not possible to definitively exclude Bayesian updating as an influencing factor, the vignette was designed to elicit an emotional, rather than rational response from practitioners. Indeed, a number of practitioners acknowledged that, in response to this vignette, their decision was based on being “worried” or “overly sensitive” (Tool 2, Interview 4) and that the past experience would be “subconsciously” affecting their decision-making (Tool 3b, Interview 2). On this basis, the author’s interpretation was that this was an example of recall bias.
Another bias revealed through the vignettes was confirmation bias—the phenomenon whereby people only look for evidence that confirms their pre-existing views (Kirkman and Melrose, Reference Kirkman and Melrose2014). When presented with the third vignette, a number of practitioners explained that they would again treat the case as high-risk even though the tool was suggesting otherwise; not because of the prior week’s case, but because they believed that poor school attendance is associated with negative outcomes. For example, one participant explained, “my specialism is domestic abuse and… attendance is one of the factors [associated with domestic abuse]” (Tool 1, Interview 3). These practitioners were placing an irrationally high focus on school attendance as a risk factor because of pre-existing beliefs and preconceptions.
These findings highlight that the majority of practitioners who were artificing were drawing on both expert intuition and HB to supplement the tool’s insights, highlighting that biased artificing and expert artificing coexist more often than not.
6.2. What does artificing look like?
There does not appear to be a uniform way of “artificing.” Some participants in this category described using the tool to help them prioritize cases (Tool 1, Interviews 2 and 4); one person described using the tool like the advice of an experienced colleague (Tool 3a, Interview 2); and a number described using the tool to trigger earlier conversations (Tool 1, Interview 4; Tool 2, Interview 2; Tool 3b, Interviews 2 and 3). Many practitioners described using the tool as a check-and-balance:
If it’s a case I’ve decided we don’t need to respond on, and then I see a risk score of 20, it makes me think again a little bit about it. Maybe not make me change my mind… but it helps me… go back through and take a look at it again… are we still comfortable with the decision we’ve made? (Tool 2, Interview 4)
A number of participants also described the tool as helping to guard against complacency:
When you’ve worked in social work for a long time, you build a high tolerance for neglect and unacceptable behaviors. Tools help you to be more reflective, so it’s no longer just your view. (Tool 1, Interview 2)
Public sector organizations, who are being advised by tool designers to encourage artificing, might find this variability confusing, because it makes artificing harder to define and model as the preferred use-type. However, it does not appear to be possible, or desirable, to attempt to codify artificing for two reasons.
Firstly, AI tools are not as good as predicting rare events and unusual combinations of circumstances as they are common ones, because there are less data available to train them on (Church and Fairchild, Reference Church and Fairchild2017; Pryce et al., Reference Pryce, Yelick, Zhang and Fields2018; Munro, Reference Munro2019). For this reason, it may be appropriate for practitioners to determine the extent to which they rely on the tool in any given case, depending on whether the circumstances presenting are commonplace or appear to be more unusual. Avoiding a tight definition gives practitioners licence to exercise this flexibility in their approach. And determine the extent to which they rely on a tool in any given case.
Secondly, a key reason that designers recommend that the tool be used in conjunction with professional judgment is to ensure that uniquely human skills, such as empathy and intuition, continue to inform the decision-making process (Munro, Reference Munro2008). Being overly prescriptive in how to use the tool is likely to undermine practitioners’ ability to express the very humanness that is seen to be an important counterbalance to the tool.
Thus, artificing is a dynamic process which varies from practitioner-to-practitioner, and case-to-case.
Interviews revealed five conditions that appeared to be present when practitioners were artificing. These emerged inductively on the basis of the interviews through a process of qualitative coding.
6.3. Solid understanding of the tool
Practitioners who were classified in this category demonstrated a solid understanding of the tool, meaning they:
• were able to identify many of the tool’s data streams;
• demonstrated a basic understanding about how the tool works
• understood the strengths of ADTs, for example, their ability to “pick up on patterns which may not have been picked up before” (Tool 3b, Interview 3); and
• were able to specify the response variable that the tool was using to inform its analysis (relevant for AI tools only).
Interviews did not reveal why these practitioners had a strong understanding of the tool. The training that practitioners received in how to use a tool did not appear to be a critical factor and neither did the tool interface. There were several examples of practitioners using the same tool—having received the same training—who exhibited vastly different understandings of the tool—some very strong, and others very weak. Future studies could focus on this question.
It is also important to note that, in addition to understanding the strengths of these tools, participants in this category also recognized their limitations. For example, one practitioner explained:
The tool is going to let you know what school attendance is, but it’s not going to let you know the reasons behind that… (Tool 3b, Interview 2)
However, artificers did not see those limitations as a reason to reject the tool; rather, limitations were seen as a reason to use the tool in conjunction with their “professional curiosity” (Tool 3b, Interview 1) or expert intuition:
It’s about using the data but putting the human factor in as well. Using the two together. (Tool 3a, Interview 2)
There are three reasons why practitioners in this category might have been willing to tolerate the tools’ limitations.
Firstly, a number of participants in this category described being able to make adjustments to the tool. Dietvorst et al. (Reference Dietvorst, Simmons and Massey2018) have reported that people will use imperfect algorithms if they can (even slightly) modify them. This is discussed as a distinct condition below but appears to be an important reason why practitioners in this category were willing to work with the tool despite knowing its shortcomings.
Secondly, many practitioners in this category felt confident to adapt how they were using the tool in order to overcome its limitations. For example, one practitioner working with the non-AI tool recognized that the data were out of date and recognized this as a shortcoming. However, she had designed a way of using the tool which compensated for this shortcoming, meaning that she was happy to continue using the tool. She used the tool to prioritize children and then flagged these priority children with partner organizations 2 weeks in advance to give them time to gather up-to-date data on the child ahead of their meeting (Tool 1, Interview 4). This suggests that when people can adapt how they’re using the tool, they will be more tolerant of its limitations.
Finally, it may be that participants in this category were willing to tolerate algorithmic error because they did not see themselves as having to make a stark choice between their own judgment and the advice of the tool. In Dietvorst et al.’s studies (Reference Dietvorst, Simmons and Massey2015, Reference Dietvorst, Simmons and Massey2018), people were asked to choose between the prediction of a human or a machine. Perhaps when people are able to identify the third option—humans working together with machines—they are more tolerant of algorithmic error because they feel more confident that they will be able to compensate for the tool’s deficiencies.
6.4. Acknowledgment of human limitations and bias
Another feature which was present when participants were artificing was an acknowledgment of the limitations of human decision-making.
The tool was seen to be something which could help overcome these limitations. In particular, there was an acknowledgment that the tools could process a lot more information than practitioners were able to. One participant described the tool as being able to “see what has happened over a long period of time.” As a result, he explained, the tool may be able to identify and highlight relevant history which is not apparent to the practitioner who has limited information, and limited time in which to process it (Tool 3b, Interview 3).
Practitioners also identified the existence of cognitive biases, and the role that ADTs could play in helping to address these:
When you see the same families come through again and again you start to think, “oh, it’s this family again… and we’re just going to screen it out”… but the tool is a reminder to stop and think… am I being biased here? (Tool 2, Interview 4)
6.5. Value of the tool clear
Building on the finding above, practitioners who were artificing saw significant value in the tool; they felt that the tool would help them to “do a better job” (Tool 3b, Interview 3) by helping them to work faster and reduce the chance of missing something.
A number of practitioners in this category also explained that the tool provides them with a greater sense of confidence, both because the tool makes more information available, but also because the tool is more reliable than the advice of another professional. As one participant explained, without the tool, her data-gathering relies on the advice of other professionals, which is often given over the phone. This creates a sense of unease because “they can turn around and say, ‘I never said that’” (Tool 2b, Interview 1). In contrast, if the tool provides a particular data-point about a child, this feels more concrete.
6.6. Opportunity to modify
A number of practitioners who were artificing had been given the opportunity to provide feedback on the tool and suggest how it might be improved. This ability to provide feedback on the tool likely contributed to their willingness to use and trust it.
A good illustration of this was borne out by conversations about the non-AI tool, where the data which the tool depended upon were often out-of-date. Where a practitioner felt that her feedback that the data were out-of-date was ignored, she ignored the tool as a result (Tool 1, Interview 1). In contrast, where another practitioner using the same tool was invited to be involved in a pilot to improve data currency, she was comfortable using the tool, and was, in fact, artificing, despite being aware of its shortcomings (Tool 1, Interview 4).
6.7. Tool verifying experience and expectations
Finally, a number of participants in this category suggested they were comfortable using the tool because it was largely confirming their own views and expectations:
When I’m going to score someone, I can look at the history and say, this person’s going to be [very high]. And usually I’m right. (Tool 2, Interview 4)
Unlike many of the other conditions listed above, this does not appear to be a sound justification for using the tool; rather, it appears to be a form of confirmation bias (Munro, Reference Munro2008, 21).
7. Algorithm Aversion
Participants were classified within this category if they described the tool as having no impact on their decision-making:
If I assign something, it’s because I believe it meets criteria and everything else for me is null and void. I won’t change my mind… I would never un-assign something because of a [low] score [on the tool]. (Tool 2, Interview 1)
In addition, participants in this category suggested that their practice had not changed, or they didn’t anticipate it changing, as a result of the introduction of the tool. For example, one participant observed that if a practitioner in her team reported concerns about a child’s school attendance, her approach would not change, even if the tool indicated that attendance was 100%
…in all honesty, I’d still want to pick up the phone. I’d still pick up the phone [and call the school]. (Tool 3b, Interview 2)
Practitioners in this group described relying more on traditional sources of information like the referral report, case-note systems, and existing databases to inform their decision-making, with the tool sitting at the bottom of the hierarchy.
Five out of the 13 practitioners fell into this category.Footnote 4 Interviews revealed a number of common conditions that appeared to be present when participants displayed algorithm aversion.
7.1. Poor understanding of the tool
Participants in this group had a poor understanding of the tool. Many identified knowing very little—if anything—about the data streams that the tool was drawing on to generate its output:
I don’t know if it draws [data from] other systems… I think it’s just ours. (Tool 2, Interview 3)
A number of practitioners within this category also misunderstood how the tool functioned, suggesting that it drew on static historical data whereas, in fact, the data are regularly updated.Footnote 5 For one participant, this misapprehension that the data were purely historical was a key reason why she wouldn’t change a decision based on the insights provided by the tool:
I don’t want to make a decision simply based on everything from the past. (Tool 2, Interview 3)
Participants in this category also had very little comprehension of how the tool worked. For example, when asked whether she knew how the tool generates a risk score, one participant answered, “I have no clue” (Tool 2, Interview 3).
7.2. Belief in superiority of human judgment
Unlike artificers, who highlighted the limitations of human decision-making, practitioners in this category saw human judgment as being superior to ADTs:
If I’m having a face-to-face with somebody and they’re telling me… this is the information, I’m going to trust that, rather than technology. (Tool 3b, Interview 2)
This was in part because participants in this category saw observation as being a uniquely human skill, which a machine could not emulate:
Observations… are fundamental…looking around the home environment [you can assess]… How does it seem? Have they got electricity on? Have they got food in the fridge? Is the carpet sticky?… It’s what you see and how people respond to questions; how quickly they’re talking, their eye contact… that sort of thing… that may not show up on a database. (Tool 3a, Interview 1)
Practitioners were also reluctant to use the tools because they believed that the tools could not incorporate qualitative data into their analysis (Grove and Meehl, Reference Grove and Meehl1996). In fact, both of the AI tools incorporate case-note information to inform their output. This misperception that the tools work with purely quantitative data was seen as a significant shortcoming:
Often a home visit tells you a lot about a family’s situation. Is it messy? It is cluttered? Is it clean? Is it overly clean?… How do the parents present? It’s those types of observations that you make… [that] inform your lines of inquiry…and you don’t always get that information, I wouldn’t have thought, from a data-source. (Tool 3a, Interview 1)
Unlike their peers in the “artificing” category, who accepted that the tool would naturally have some limitations, practitioners in this category saw the limitations of the tool as being unacceptable. In fact, this group appeared to hold the tool to impossibly high standards (Dawes, Reference Dawes1979)—standards that far exceed what humans might be reasonably expected to achieve:
I just don’t know that it’s 100% accurate every time. (Tool 2, Interview 1)
This phenomenon of holding machines to a higher standard than humans is again explored in the work of Dietvorst et al. (Reference Dietvorst, Simmons and Massey2015) whose study showed that humans display “greater intolerance for error from algorithms than from humans. People are more likely to abandon an algorithm than a human judge for making the same mistake” (124).
In addition, a number of participants in this category described having had negative experiences with the tool previously, which likely contributed to their mistrust of the tool. One participant told a story about visiting a school to discuss the circumstances of a child who had been flagged as high risk, only to be told “that the child had died. And that’s very embarrassing” (Tool 1, Interview 1).
The fact that these negative experiences resulted in practitioners rejecting the tools supports the findings of previous studies, which demonstrate that people tend to avoid algorithms after seeing them make mistakes.
7.3. Value of the tool not clear
In contrast to artificers, practitioners in this category did not see the tool as adding any value to their practice:
If we didn’t have [the tool], I’m not sure it would make a lot of difference to our decision-making… I’m not sure it would be sorely missed. (Tool 1, Interview 1)
Practitioners described feeling comfortable and confident with their practice, and did not see the tool as having anything to offer:
I feel that our approach and our assessment is quite holistic. We can cover the areas we need to cover there and then. I’m not sure what [the tool] would add. (Tool 3a, Interview 1)
The fact that practitioners in this category did not tend to acknowledge the limitations of human decision-making, but rather, focused on the superiority of human judgment, meant that the value of the tool was far less clear than it was for artificers.
7.4. No opportunity to modify
Unlike practitioners who were artificing, practitioners in this category felt that they had no opportunities to modify, or provide feedback, on the tool’s performance. This was a particular issue for the non-AI tool, where data currency was identified as a significant shortcoming. One practitioner felt very frustrated by the data being out-of-date, and felt she had no means of recourse:
We’re repeatedly feeding back the frustrations… [but] I can’t see it changing at all, which is unfortunate. (Tool 1, Interview 1)
This suggests that the inability to provide feedback on the tool’s performance can contribute to algorithm aversion.
7.5. Tool not verifying experience and expectations
The final condition that appeared to contribute to algorithm aversion was the tool not verifying practitioners’ previous experience or expectations. One participant described having mixed feelings about the tool precisely because there had been a number of times where the score was vastly different to her own assessment:
…there were sometimes where it has pulled information and gave (sic) a family a pretty high score that just really didn’t make sense. (Tool 2, Interview 3)
Rejecting the tool because it does not conform to expectations is another example of confirmation bias. While confirmation bias is characterized by a tendency to look for information that supports existing beliefs, is also characterized by having “a blind spot when it comes to evidence that contradicts these views” (Munro, Reference Munro2008, 21).
8. Automation Bias
Participants were classified within this use-type if they described relying entirely on the tool to make their decisions. Notably, only one participant was classified in this category.
The one observed instance of automation bias was, interestingly, by someone using the non-AI tool (generally speaking, fears around automation bias have largely focused on AI tools). Early in the interview, the participant claimed, “even if a family had five flags, or no flags, that wouldn’t be the thing that guides my decision” (Tool 1, Interview 3). Nevertheless, the discussion elicited through the vignettes suggested that she would defer to the tool in certain instances.
The practitioner was presented with the first vignette, which describes a scenario where the tool indicates that a child is high risk, but the practitioner’s previous personal interactions with the child had indicated there were no causes for concern. When asked what she would do in this situation, the practitioner responded that she would investigate immediately:
With five flags, that’s very significant. There’s something going on for that child. (Tool 1, Interview 3)
Here we see that, despite saying that the tool wouldn’t be the key factor in her decision-making, when presented with a scenario where the tool is raising five risk-flags, the practitioner explains that she would investigate. She accepts the information offered by the tool, without tempering it with her own experience with the child.
It is important to note that this participant also exhibited signs of artificing. Given this, it may be that the deferential behavior was driven by more risk aversion than true automation bias.
Practitioners working in children’s services are (unsurprisingly) notoriously risk averse (Kirkman and Melrose, Reference Kirkman and Melrose2014). It is highly plausible, then, that the reason the participant relied so heavily on the tool in this particular instance was because the tool was suggesting that there were significant risks. It would perhaps be easier to confidently classify behavior as “deferential” where a practitioner decided—against his or her own best judgment—to not investigate a child because of a low-risk score. There were no examples of this.
9. Relationship Between Practitioner Personal Characteristics and Use-Type
As part of the interview process, participants were asked a range of questions about their personal characteristics to explore whether there was any association between these features and practitioners’ approach to the tool. As illustrated in Table 2, and recognizing the very small sample sizes, there were no meaningful associations. Unlike previous studies, which have identified that “novices” are more likely to defer to decision-tools, while experts are more likely to artifice or ignore the tool (Fook, Reference Fook, Ryan and Hawkins2000; Gillingham, Reference Gillingham2011; Kirkman and Melrose, Reference Kirkman and Melrose2014), here there was no apparent association between years’ experience and use-type.
Practitioners’ personal characteristics and relationship to use-type.
Abbreviation: AI, Artificial Intelligence.
a Single subject also displayed elements of artificing.
In addition, practitioners’ sex, tool type, age, qualification level, attitudes toward technology, seniority, and field of work all appeared to have very little bearing on use-type.
As mentioned above, the limitations of this study mean that these conclusions should not be overstated. Nevertheless, they provide an interesting counterpoint to previous studies, which would benefit from further exploration.
10. Discussion
The findings above show that practitioners are using the tools differently, supporting findings that practitioners do not necessarily use the tool as instructed (Gillingham, Reference Gillingham2016; Gillingham and Graham, Reference Gillingham and Graham2017). This discussion will explore the conditions associated with different approaches, as well as the implications of different use-types.
11. Conditions Associated with Different Use-Types
There are five observed conditions that influence a practitioner’s approach to using the tool.
The findings also show that the conditions present when people are artificing are an inversion of the conditions present when people exhibit algorithm aversion, shown below in Figure 1.
Conditions affecting use-type—a spectrum.
While these conditions have been presented as distinct, many of them are synergistic. For example, someone who recognizes the flaws in human decision-making is more likely to value the tool. In contrast, someone who believes that human decision-making is infallible is unlikely to see the tool’s value.
It is also important to acknowledge that there were observed inconsistencies in the conditions that were present when people were artificing or ignoring the tool. What the findings above have attempted to do is draw out common, rather than universal, conditions which help explain why some participants were artificing while others were ignoring the tool. The cross-over and inconsistencies highlight that this is an area which would benefit from further research.
In addition, given the small number of interviews, future studies could also seek to verify the veracity of these conditions and explore whether other conditions—for example, decision-type—also influence the way practitioners use the tool.
There were also instances of people displaying elements of more than one use-type, with one participant displaying elements of artificing and aversion, and one participant displaying elements of artificing and automation bias. This highlights that while there might be some practitioners who sit consistently at either the left or the right of every arrow, others may show more variability. Further research is needed to explore this in greater depth.
Given that only one participant demonstrated elements of automation bias, the conditions which support this use-type were not observable from this study. Future studies could explore whether the same conditions observed for artificing and aversion are present when people are exhibiting automation bias.
12. Implications of Different Use-Types
12.1. Artificing
Nine out of 13 practitioners were expert artificing, meaning that they were using the tool as recommended by tool designers. However, of these nine practitioners, seven were also biased artificing, meaning they were drawing on elements of expertise, together with elements of HB, to inform the intuitive element of the decision-making process.
This finding challenges a core claim used to justify the introduction of these tools; namely that they will minimize bias and increase objectivity in human decision-making (Shlonsky and Wagner, Reference Shlonsky and Wagner2005; Church and Fairchild, Reference Church and Fairchild2017; Schwartz et al., Reference Schwartz, York, Nowakowski-Sims and Ramos-Hernandez2017; Pryce et al., Reference Pryce, Yelick, Zhang and Fields2018). It also highlights that ADTs should not replace investment in professional education programs designed to minimize bias and develop professional expertise (Fook et al., Reference Fook, Ryan and Hawkins2000) amongst children’s services practitioners.
12.2. Aversion
More than one in three practitioners were ignoring the tool. This is significant because “even the most predictive, psychometrically sound instrument, if not used, will be ineffective in the field” (Shlonsky and Wagner, Reference Shlonsky and Wagner2005).
This finding also has implications for evaluating and understanding the tool’s impact. Tool impact assessments which assume that people are using the tools as instructed will produce misleading results if people are, in fact, ignoring the tool. For example, the authors of the recent impact assessment of the Allegheny Tool in the United States by Stanford University note that 39% of practitioners were not referring cases for further assessment when mandated to by the tool. Despite this, the authors did not take potential algorithm aversion into account when conducting the evaluation (Allegheny County Department of Human Services, 2019), meaning the results give an incomplete understanding of the tool’s effects in practice.
Incorporating the framework of artificing, automation bias and aversion into future impact assessments would help to paint a clearer picture of the true impact that tools are having.
13. Automation Bias
As highlighted above, automation bias did not feature strongly as a use-type. A key reason may be that deference is being explicitly discouraged in tool training. One practitioner described that she and her colleagues had been, “directed that we shouldn’t be basing our decision solely off the tool” (Tool 2, Interview 1).
However, while automation bias is not yet an observed feature in how the majority of practitioners are using these tools, a number of practitioners expressed concerns that deference will emerge as an issue once tools become more embedded in practice:
[I’m nervous about reaching a point where] we take what’s written down [by the tool] as gospel and stop speaking to people. (Tool 3b, Interview 1)
This is certainly something to continue monitoring, as an overly deferential approach is something to be avoided because, while ADTs are very good at certain things—for example, processing and synthesizing vast quantities of data and spotting patterns—they also have significant limitations (Shlonsky and Wagner, Reference Shlonsky and Wagner2005). In particular, they:
• are not good at predicting rare events (Pryce et al., Reference Pryce, Yelick, Zhang and Fields2018);
• “cannot be programmed to predict for every single event that may occur at any point in the future” (Elish, Reference Elish2019, 10);
• are often trained on incomplete data (Munro, Reference Munro2019);
• are often trained on biased data, resulting in discriminatory tools (Munro, Reference Munro2019).
For these reasons, it is critical to cultivate an environment where SLBs feel empowered to question, scrutinize, and challenge ADTs; treating tools less like an objective master and more like any other source of power (Fry, Reference Fry2018).
Having described the implications of the different use-types, I will now explore the conditions that appear to support different approaches to use the tools.
14. Conclusion
This paper has examined how practitioners working in children’s services are using ADTs to support their decision-making. Drawing on Herbert Simon’s theory of satisficing, together with NDM and Heuristic and Biases theories, the author introduces a new conceptual framework for understanding how SLBs use ADTs.
Interviews reveal that artificing and algorithm aversion are the most common use-types. Only one practitioner displayed elements of automation bias. However, there are fears that deference may become more prevalent as tools become more established. This is something to continue monitoring, as keeping a “human in the loop” (Elish, Reference Elish2019, 20) is critical given the known fallibilities of ADTs.
Interviews also revealed that, while the distinction between expert and biased artificing is conceptually quite neat, in practice, the lines are blurred. When practitioners rely on intuition to guide decisions, they are—in most cases—drawing on a combination of biased and expert intuition. This highlights that HB endures as a feature of decision-making despite the introduction of ADTs.
The existing literature offers very little in terms of helping us to understand what drives different use types (Dietvorst et al., Reference Dietvorst, Simmons and Massey2015). This research helps to fill that gap by offering insights into the different conditions that were present when people were artificing and ignoring the tool. Understanding which conditions support a constructive human–machine interaction, and which do the opposite, is critical if these tools are going to succeed in supporting better decision-making in the public sector.
There are a number of key limitations to this study. Firstly, the insights gleaned from this research are based on semi-structured interviews only; an ethnomethodological approach (Garfinkel, Reference Garfinkel1984) would offer deeper insights. Secondly, the number of subjects is small. This was due to the nascent nature of the tools, as well as a general reluctance by Local Authorities to be interviewed due to intense media scrutiny of the use of ADTs in children’s social care (see e.g., “Thurrock Council slammed for using analytics to profile residents and predict traits”, Thurrock Gazette, Reference Gazette2018; Pegg and McIntyre, Reference Pegg and McIntyre2018). However, given the nature of this study was theory generating, rather than theory testing, the limited number of subjects was seen to be acceptable.
This study also lays foundations for future research. While artificing is being promoted by tool designers as the optimal way to use the tool, there is, as yet, no empirical evidence to demonstrate the superiority of this approach. The framework outlined in this study could be used to empirically validate the hypothesis that a combination of human and machine intelligence results in the best outcomes for families and children.
Funding Statement
This work received no specific grant from any funding agency, commercial or not-for-profit sectors.
Competing Interests
The author declares none.
Author Contributions
Thea Snow was the sole author of this article, responsible for conceptualization, methodology, analysis, and writing.
Data Availability Statement
The Appendix (uploaded separately) to this paper contains interview protocols, questions, and vignettes used to conduct the interviews. In order to conduct in-depth and frank interviews, interviewees were assured anonymity both for themselves and for their organizations and work sites, and so research ethics concerns preclude making transcripts available.
Supplementary Materials
To view supplementary material for this article, please visit http://dx.doi.org/10.1017/dap.2020.25.
Acknowledgments
The author would like to acknowledge a number of people whose support made this research possible. Firstly, the designers of the algorithmic tools who not only contributed their insights, but also set up practitioner interviews, without which, this study could not have happened. Also, the practitioners who gave up their time for interviews. Thank you to a number of social work professionals—Tarryn Klotnick, Julie Masters, and Julie Rooke—who offered feedback and guidance in developing the interview questions and vignettes. And finally, thank you to Associate Professor Daniel Berliner (London School of Economics and Political Science) for his support and feedback throughout the entire process.
Open access
Comments
No Comments have been published for this article.