The Moral Affordances of Construing People as Cases: How Algorithms and the Data They Depend on Obscure Narrative and Noncomparative Justice

Enter Algorithms

That brings us to algorithms. At its most basic, an algorithm is simply a set of rules one follows to carry out an operation. Here, I focus on the sorts of algorithms that have most captured the attention of sociologists—those designed to allocate social and economic benefits and burdens, such as loans, jobs, police surveillance, educational opportunities, fraud investigations, and so on (Brayne and Christin 2021; Burrell and Fourcade 2021; Levy et al. 2021). While algorithms need not be computationally sophisticated, increasingly they are, harnessing techniques such as machine learning to draw on large amounts of personal data to detect patterns in human behavior and extrapolate those patterns into the future. These predictions then help people in positions of authority—loan officers, employers, police officers, admissions officers, fraud investigators—make decisions about who gets what.

For algorithms to work, people must be construed as cases. To an algorithm, a person is an entity with a set of features, and insight comes from comparing people with different sets of features to one another. In any given sample (or any sample useful to an algorithm), some people exhibit a behavior, and others do not. Some repay loans, and some default. Some succeed in college, and others drop out. An algorithm analyzes which patterns of characteristics have gone with various outcomes in the past and then, when new people come along, presumes their patterns will correlate with outcomes in the same way. A lender may, for example, use machine learning to identify a relationship between people moving money in and out of bank accounts and the likelihood of them paying their credit card bills on time. When courting a new customer, the lender examines a person’s bank deposits and withdrawals to see if the pattern is similar to those historically associated with nonpayment and if so, charges a higher interest rate or decides not to issue a card at all.

Behind all algorithms sit databases, which bring certain material constraints. Databases come in different architectural varieties—they may, for instance, be structured around hierarchies, networks, or attributes—but they all nonetheless depend on information rendered categorically (Bowker 2005; Couldry and Hepp 2016:132; Gillespie 2014). In the most commonly used “relational” databases, information is stored in sets of tables, with each column capturing a different type of information (e.g., name, age, occupation) and each row encoding a “relationship” between values for a particular “object” (e.g., person). Other types of databases use different organizational schemes, but the scheme itself is constant—common across the entire database. As Dourish (2017:108) explains, “arranging data into a common format makes each data item amenable to a common set of operations.” Data items can be sorted, collated, grouped, summarized, and—crucially for the argument I present here—compared.

Indeed, the insights algorithms provide to decision-makers always depend on comparison across cases. How a person is predicted to behave is a function of how other, somehow similar and somehow different, people have behaved in the past (Hacking 1990; Schauer 2003). Even as data become more fine-grained and individualized, the mechanics of algorithms still demand decontextualization and standardization so that information flows from one person can be easily compared to those from the next (Guseva and Rona-Tas 2001). This continues to be true when algorithms themselves determine which sorts of information about people to attend to. So-called unsupervised methods still depend on the processes of categorization and classification even if those processes become opaque to their human designers (Fourcade and Healy 2017a). ⁵

This is radically different from how one often uses narrative to understand a person and what the person is likely to do in the future. To make sense of cases, we gravitate toward comparing them, but with narrative, we can easily size up just a single person because that person moves through time and particular situations, demonstrating change and trajectory. For example, in Heimer’s (2001, 2009) study of the neonatal ward, parents of ill children had typically never seen such a sick infant; they had no store of cases to draw on, as doctors did. Yet parents watched their own child hour by hour, day by day, witnessing incremental change and the circumstances that went with it. Their knowledge was intensive rather than extensive, and for that reason, they at times had insights about their child’s health that doctors did not.

Now, that does not mean we cannot compare narratives. Indeed, we do all the time. Whose story is more believable, hers or his? How did this person succeed while that one failed? ⁶ What rendering people as cases uniquely enables is comparison at scale—figuring out similarity and difference not between two or three people but between dozens or hundreds or thousands or more. Comparing a small number of narratives is tractable because the basis of comparison can itself be emergent and reflexive, pivoting along with each story’s narrative arc and the meaning it produces. Yet this is exactly what gums up the works if we try to compare stories many at a time.

Importantly, narrative is not synonymous with singularity. At times, we care about the stories of individuals’ lives precisely because they capture the experiences of other people in similar positions (Ewick and Silbey 1995; Polletta 1998; Steinmetz 1992). Protagonists can have proper names, or they can be archetypes (e.g., a single mother, a conservative voter, a Silicon Valley executive). Narrative thus works at two levels. Stories can be about specific individuals, or they can invoke cultural tropes about particular types of individuals. ⁷ Rather than a lack of generality, what makes understanding people in narrative distinct from understanding them as cases is the way that narrative bakes in chronology and context and characters who must grapple with both—with meaning emerging from this entanglement. These features then carry important consequences.

The temporal flow of narrative, in which one event leads to the next, means stories are well-suited to exhibiting why things happen (Bruner 1991; White 1980). As Polkinghorne (1988:21) writes, if we want to understand a man buying life insurance and we have rendered him as a case, our explanation may go something like this: “because he is a [W]hite male, in the 40-to-50 age category, and those in this category are, in 70 percent of cases, also in the category of people who buy life insurance.” The explanation is unsatisfyingly sparse, telling us precious little about how one state of being translates into another. It gives no hint—as narrative would—of what has transpired in the man’s life (becoming a father, witnessing a colleague die suddenly) or motivated him to action (thinking of himself as a provider, receiving an insurance solicitation in the mail).

That does not mean people rendered as cases cannot be arranged temporally. Narrative, as discussed earlier, does not have a unique claim on one thing happening after the next. Yet to arrange cases longitudinally, they in fact must be arranged. As Manovich (1999:85) writes, a “database represents the world as a list of items which it refuses to order.” Someone can come along and impose a temporal order on those items, but such sequencing—and the meaning that follows—comes from without, not within.

Moreover, because narrative often reveals (or at least suggests) a protagonist’s inner mental life, including emotions and intentions, it helps assign responsibility, blame, and praise (Bruner 1986; Harding et al. 2017; Stone 1989; Wagner-Pacifici 2010). In the example of the man purchasing life insurance, we draw different conclusions when the buyer is worried about his family than when he is contemplating how he might fake his own death. Indeed, even small changes to the interplay between narrative arc, characters’ actions, and the surrounding situation can lead to very different judgments about intentionality and control and, in turn, how we feel about and understand the person at the center of the story (Cerulo and Ruane 2014; Monroe and Malle 2017; Scott and Lyman 1968).

All of which is to say, construing people as cases and analyzing them through case comparison allows for a different sort of understanding than does making sense of people as actors in a narrative. Each mode of organizing cognition foregrounds different types of information, making people legible in different ways and enabling different styles of reasoning. While not all scholars of categorization and classification directly engage with narrative as the epistemic “other,” they nonetheless tend to describe the processes they study as ones that render people as entities that are somehow flat, thin, stripped down, or otherwise incomplete—albeit incredibly well-suited to modern forms of bureaucratic, quantitative, and computational governance.

The Missing Link to Moral Evaluation

What I argue has gone less theorized and that the comparison to narrative puts into sharp relief are the moral ramifications of rendering people as cases. That is not to say the literature on categorization and classification ignores the issue. In one sense, it is rife with moral matter. When powerful institutions, such as states or corporations, categorize people in some but not other ways, they make certain social experiences visible while obscuring others (Bowker and Star 2000; Lampland and Star 2009; Lengwiler 2009; Minow 1990; Starr 1992). This then patterns access to material and symbolic resources, at times reinforcing patterns of historical (dis)advantage, hence the high stakes of battles over which categories ought to be used (Bourdieu 1984; Starr 1992; Tilly 1998). ⁸ Categories can also themselves be value-laden, or “categories of worth,” whether they appear in official classifications or underpin broader cultural understandings of good and bad (Durkheim 1965; Lamont 2000; Mohr 2005; Zerubavel 1997). The moral valence of such categories makes them powerful in motivating social action—consider the politics of the “deserving” and “undeserving” poor (Somers and Block 2005; Steensland 2010).

That said, nearly all the scholarly focus is on how different values and moral understandings are embodied by one classification system versus another. There is much less exploration of what it morally means to categorize people in the first place, to construe them as entities that are properly differentiated and defined by their attributes (whatever those may be). The attention is on understanding what it means to render people as this sort of case or that sort of case—not on understanding the moral implications of the cultural form itself, of transforming a person who is not a case into one who is.

On this count, the quantification literature is more developed. Espeland and Stevens (2008:431–32), for example, talk at length about the “ethics of numbers,” including the tendency of quantifying processes to “narrow our appraisal of value and relevance to what can be measured easily, at the expense of other ways of knowing.” It is not simply that the way we quantify embodies particular values, although it does (Fourcade 2011; Salais 2016). It is also that the reduction of difference to that which can be unambiguously counted and otherwise mathematically manipulated is, in and of itself, a judgment about how certain types of (incommensurable) features do not matter (Espeland 1998; Espeland and Stevens 1998).

Yet quantification is not the same thing as categorization. Before we can count things up, we first must fit them into categories—categorization logically and technically precedes quantification (Fourcade 2016; Hacking 1986; Hirschman, Berrey, and Rose-Greenland 2016). This means we cannot necessarily assume lessons learned from the study of quantification apply to categorization because quantification is the subsequent process. The question of what moral stakes come with categorization and classification (i.e., the processes that render people as cases) as such remains largely unexplored.

Affordances and the Work They Do

I posit that rendering a person as a case or in narrative affords different forms of moral reasoning. Drawing on ecological psychology and design studies, I define affordances as features of an environment or object that steer attention and behavior. A chair, with its horizontal platform at knee level, affords sitting. A door with a handle that juts out in the shape of a sideways “U” affords pulling, whereas one with a plate that sits flush against the glass affords pushing. One certainly can stand on a chair (or throw it across a room) or push a U-shaped handle forward, but these are the less likely acts.

Working on problems of visual perception, Gibson (1979) was the first to popularize the concept of affordance as a way to link an animal’s environment to its actions while downplaying the intermediary step of mental representation. ¹¹ Norman (2002) later picked up the term, applying it to everyday objects and technologies in part to explain why people often instinctually use such items incorrectly—because their designers have created them with the wrong affordances. Affordance means slightly different things in each tradition (Chong and Proctor 2020; Davis 2020), but what is consistent is a particular unthinking quality. Through affordances, the world around us guides us to certain actions even in the absence of deliberate reasoning and active decision-making.

Affordances thus have real-world effects. In one example, Norman (2002:9) notes that when British transportation authorities built bus shelters out of glass, vandals smashed them, and when they built shelters out of wood, vandals drew on them with graffiti. Those setting out to commit crime did so in different ways because each material invited a different set of actions. Yet, importantly, affordances are nondeterministic. In fact, the ability of an affordance to shape behavior is often culturally and institutionally mediated (Davis and Chouinard 2016). Door handles afford opening doors, but in certain times and places, women still do not, waiting instead for the man to do so.

This, I argue, is also the nature of the relationship between cultural forms for rendering people (as cases or actors in narrative) and types of moral reasoning. Systems of modern governance—including but not limited to algorithms—depend on capturing people in reduced, rationalized, unambiguous formats, such as in rows of spreadsheets with n columns of characteristics. This, in effect, is a design decision, which makes it easier to think in some directions than in others. ¹² This is akin to what information scholars Dourish and Mazmanian (2013:101) call “the material consequences of representational practices.” As they write, “the particular forms that information takes . . . shape the questions that can easily be asked of it, the kinds of manipulations and analyses it supports, and how it can be used to understand the world” (Dourish and Mazmanian 2013:100).

When people are rendered in a way that affords comparison, decision-makers are more likely to compare them for the purposes of knowing things about them—and, I add, for normative evaluation. This does not mean that standards of comparative justice are the only ones applied, just that they are the more salient ones, the ones people more easily reach for. Affordance is not fate, and there is still plenty of room, even with people construed as cases, for concerns about noncomparative justice to surface. But the information infrastructure itself does not facilitate this.

Moral Reasoning and the Cultural Form People Take

To return, then, to the main issue: How does rendering people as cases afford one mode of moral reasoning while posing informational difficulties for another?

Rendering people as cases reduces the information available about them, but in a specific way. Through categorization and classification, people retain only attributes that are common across individuals. When a bureaucratic worker logs clients into a database, the fields may take on different values, but the fields themselves do not change. A person may be tagged as married, single, or divorced, but marital status as a category of knowing is held in common. People are still, ontologically speaking, idiosyncratic, but in decision-making systems that capture them as cases, axes of similarity and difference are standardized. This homogenization makes it easy to compare people to one another—including when adjudicating whether standards of comparative justice, such as equality and proportionality, have been met. If we want to know whether men and women are treated equally, it helps to have each person flagged as either a man or a woman. If we want to know whether low- and moderate-income taxpayers are audited at higher rates than rich ones, it is useful to have individuals assigned to discrete income brackets.

Trying to get a handle on a comparative justice issue like social equality with people rendered in narrative would be much trickier. The modularity of cases paves the way for comparison at scale, but the entanglement of narrative’s component parts (plot, character, setting) complicates matters, as does how a story produces meaning—in emergent, reflexive, and nonlinear ways. A story is more than the sum of its parts (even omitted information can be critical), and it functions as an integrated whole. As Somers (1994:616) writes, unlike categorical thinking, “narrativity precludes sense-making of a singular isolated phenomenon” (see also Ricoeur 1980). To reduce a story to a set of standardized dimensions along which it might be compared with others would be to destroy the essence of narrative. Moreover, narrative’s complicated relationship with temporality undermines simultaneous understanding (Wagner-Pacifici 2010), which is key for comparison.

By contrast, for many standards of noncomparative justice, such as desert and merit, understanding people as cases informationally falls short. Desert, for example, is a standard by which people rightly receive benefits or burdens according to the moral worth of their past actions (Feinberg 1970; Feldman and Skow 2019; Rachels 1991; Sher 1987). Students who master the material deserve good grades, and people who commit violent acts deserve prison sentences. Adjudicating whether such standards have been properly applied often requires knowing the sorts of things case-based renderings are bad at conveying, such as what led to observable behavior. Students who ace an exam because they stole the answer key actually do not deserve good grades, and people who hurt others in desperate acts of self-defense (arguably) do not deserve prison sentences, or at least not long ones. Construing people as cases strips away information about intention and motivation, context about the constraints people face, and connections between disparate events—all of which is often necessary for determining what people can be fairly held accountable for in a noncomparative sense.

This type of information is much more readily supplied by narrative, given the way plot, character, and setting are hardwired into the cultural form. Indeed, when faced with people construed as cases (e.g., as they exist in administrative documents and other records), decision-makers often actively reach beyond the written page and turn to narrative to fill in informational gaps—whether that means soliciting stories from the individuals they are evaluating or simply imagining what those stories are likely to be (Espeland 2015; Lynch 2017). Through narrative, it is easier to see and appreciate the role of social structure, differences in subjective understanding, and life trajectory—what in moral evaluation adds up to consideration for extenuating circumstances, personal redemption, and so on.

Key to this process is how narrative conveys both how and why things happen. In adjudicating noncomparative justice, an important question is often how much of a role a person has had in bringing about events, whether the person did so intentionally, and if the person could have acted otherwise (Lempert and Sanders 1986). Narrative excels at providing thick causal understanding, the sort that refers back to both proximate and ultimate causes, and integrates not only observable events but also people’s thoughts and feelings—stories come with a point of view. Indeed, some scholars of narrative argue that the establishment of cause and effect is inherent to the form, part of what makes narrative different from other kinds of temporal ordering and what gives stories a sense of meaning and closure (Polkinghorne 1988; Polletta et al. 2011; White 1980; see also, Knight and Reed 2019:248). In this way, narrative is well-suited to the assignment of responsibility, not just in a mechanical sense but also a moral one.

Saying that rendering people in narrative affords adjudications of noncomparative justice is not to say that every noncomparative situation requires narrative’s fullest trappings. Some noncomparative standards, such as rights, may depend on only the sparsest of biographical detail, such as having been born in a particular country. Again, we are in the world of affordance, not one-to-one correspondence. That said, in practice, narrative is often what helps decision-makers get noncomparative justice exactly right rather than only approximately so. Consider, for example, the students above who do well on an exam. At first pass, understanding the students as decontextualized, detemporalized cases may appear to work just fine. All we need to know is that students who fall into the category “all answers correct” deserve an A. But this is not true for the cheaters. The stories of how people wind up in the categories they do often reveal moral distinctions that classification obfuscates (Kiviat 2021). Indeed, this is where the instability of narrative’s temporal boundaries can come in handy. Starting the story a bit earlier or letting it play a bit longer can cast important new light on the person in question.

None of this is to say that narrative is a better tool for adjudicating justice claims overall. For identifying comparative (in)justice, construing people as cases clearly works better. Indeed, narrative’s relative informational inattention to points of comparison across different people can mean that through storytelling, we reintroduce bias. Enmeshed in narrative particulars, it can be hard to see the forest for the trees. Moreover, narrative’s showcasing of mental states, including motive, can help decision-makers claim that what seems to be a comparative injustice actually is not (e.g., an employer did not intentionally discriminate, and therefore, the resulting allocation of jobs is not problematic) (Lempert and Sanders 1986:390). The goal is not to dismiss thinking about people as cases but, rather, to show that for a full normative evaluation of the allocation of benefits and burdens, we need both cultural forms.

Theorizing Algorithmic Fairness

As algorithmic decision-making has spread far and wide, so have ideas about what it means for algorithms to allocate fairly. Organizations increasingly use computational predictions about how people will behave to assign benefits and burdens. These practices and their rhetorical defenses have been shaped by academic ideas about how to identify injustice. There is now a large body of scholarship on “algorithmic fairness,” which chiefly sits in computer science but also features the work of legal experts and philosophers. Here, I take this scholarship as an empirical phenomenon of interest in its own right—as an instance of a group of people embarking on a task of moral evaluation. ¹³

The algorithmic fairness literature largely arises from the discipline of computer science, and it bears that field’s hallmarks. The goal is generally to find a solution to a problem. Each member in a population experiences an outcome (e.g., a loan is granted, bail is denied) and the task is to figure out if by a particular definition of fairness, those outcomes have been allocated fairly. In approaching this problem, computer scientists use their usual tools and methods, including abstraction, formal logic, and calculation. Computer scientists inhabit an intellectual world where the information infrastructure must accommodate computation (Dourish and Mazmanian 2013; Malik 2020; Selbst et al. 2019:59). This means people necessarily get rendered as cases.

In this setting, conceptions of comparative fairness dominate. The nearly universal starting point is the Aristotelian notion of treating like cases alike. In one influential article, Dwork et al. (2012:215) write: “We capture fairness by the principle that any two individuals who are similar with respect to a particular task should be classified similarly.” In another article, Friedler, Scheidegger, and Venkatasubramanian (2016:2) point to the field’s “guiding informal understanding of fairness” as the precept that “similar people should be treated similarly.” Canvassing dozens of fair machine learning experts in academia, industry, and government, Chouldechova and Roth (2018) identify two main types of fairness definitions: those that require parity between individuals and those that require parity between groups (e.g., that members of a protected class fare no worse than others, on average). In all instances, equivalence is key. Fairness is defined by how some person or group of people is treated relative to another. The definitions on offer unilaterally reflect comparative justice.

Applying these principles is no easy task, and there is substantial debate about which equivalencies are the important ones. A survey of leading machine learning outlets returned 20 prominent standards, including statistical parity (equal outcomes across groups), predictive parity (equal chances of being correctly assigned to the good outcome), and balance for false negatives (equal chances of being incorrectly assigned to the bad outcome) (Verma and Rubin 2018; see also, Berk et al. 2021; Mehrabi et al. 2022; Mitchell et al. 2021). ¹⁴ In practice, the stakes can be high. In one often-cited case, the makers of a bail-setting algorithm touted their product as racially unbiased because Black and White defendants showed equivalence on one outcome (i.e., one definition of fairness)—even though, as others later showed, Black defendants fared unequally in different ways (Angwin et al. 2016; Dieterich, Mendoza, and Brennan 2016). There is thus great dynamism in the algorithmic fairness literature and all the more so given that different metrics often cannot be satisfied simultaneously (Corbett-Davies and Goel 2018; Hellman 2020; Kleinberg, Mullainathan, and Raghavan 2017).

And yet, much is missing. Notions of noncomparative justice—of ensuring each individual gets what they deserve or have a right to—are practically nonexistent. If judges send all jaywalkers to prison for 10 years, then they are treating everyone equally and yet arguably also unfairly given how greatly the punishment outweighs the crime. The algorithmic fairness literature has no metric to detect such an injustice. Nor does it have a way to articulate what might be unfair about an algorithm that uses information on preexisting medical conditions to allocate health insurance, one that draws on social media data to determine who should get a loan, or one that uses occupation and education to price car insurance—all contentious practices, commonly called “unfair” in the broader public sphere. Algorithmic fairness, as currently theorized, is thus blind to what Fazelpour and Lipton (2020:8) call “the non-comparative harm of making decisions on irrelevant grounds.” ¹⁵

Moreover, the literature lacks a way to parse which predictions are themselves unjust. Consider, for example, universities that allocate scholarship money based on algorithmic predictions about the minimum amount of aid it will take to entice a particular student to enroll (Aulck, Nambi, and West 2019; Engler 2021). Were we to intuit that price sensitivity is an unfair standard for allocating resources in the domain of higher education, there would be no way to formalize this in the algorithmic fairness literature. Nor would there be a way to articulate that a prediction of any sort may be an unjust way to distribute a benefit or burden. In linking allocations to behavior, decision-makers can look to the past or to the future; to evidence of how people have behaved or to predictions about how they are likely to do so going forward (Jencks 1988). Many noncomparative standards of justice, such as merit and desert, rely on how people have behaved (e.g., whether they worked hard, acted with intent, achieved high marks), while beliefs about future behavior are irrelevant (Underwood 1979). In the algorithmic fairness literature, this goes entirely uncommented on.

So, why? Why exclude noncomparative justice concerns in favor of comparative ones? I propose that the answer, or at least an important part of the answer, is precisely because people have been construed as cases. ¹⁶ For algorithms to work mechanically, it is necessary to render the people they are designed to act on as cases. Scholars seeking to build tools to test the fairness of how these algorithms work therefore begin with this view of people—as database-bound objects with attributes that can be sorted and slotted according to similarity and difference (Dourish 2017). This information infrastructure then affords comparison both analytically and morally. Scholars of algorithmic fairness solve the problems they are able to with the tools they have, and the tools they have make comparative forms of justice the more tractable ones. The “material consequences of representational practices” are not only epistemological in nature (Dourish and Mazmanian 2013:101) but normative as well.

This is not to say that computer scientists and their colleagues do not have intuitions about noncomparative (in)justice. It is possible to think outside the form in which data arrive. Indeed, in the next section, I show an example of just that—of a group that deliberately reaches beyond records and databases to get at noncomparative standards such as desert. The fact that the algorithmic fairness literature does not do this or even comment on its omission is telling. Davis and Chouinard (2016) underscore that affordances are culturally and institutionally mediated, and most scholars of algorithmic fairness inhabit an intellectual world where the reduction of complex social phenomena, such as lived human experience, to discrete classification and formal logic is taken for granted. It is easy to be blind to standards of noncomparative justice when one is entrenched in institutions dependent on understanding people as cases and comfortable with the assumption that case-based renderings are sufficient representations. Affordance, as noted previously, is not fate, but it may be more so in the hands of some users.

Next I turn to a different domain, that of public policy, where the reduction of social phenomena to cases is not taken for granted and where questions about the fairness of algorithmic allocation at times take a very different form.

Debating Credit Scoring Algorithms

Scholars often hold up credit scores as the paradigmatic algorithmic prediction (e.g., Fourcade and Healy 2017b; Pasquale 2015). In their classic iteration, credit scores are constructed from information in consumer credit files—records that construe people as cases defined by their debts and other money-related attributes—and are designed to predict the likelihood that a person will repay borrowed funds. Today, organizations besides lenders, such as insurance companies and landlords, also consult credit scores to make allocative decisions (e.g., what premium to charge, whether to take on a person as a tenant), and scoring algorithms increasingly include data from beyond credit files (Consumer Financial Protection Bureau 2012). Regulators, lawmakers, and other public policy actors have watched these developments with interest, especially with the progression from easy-to-understand scorecards to more computationally sophisticated and opaque algorithms. The fairness of credit scoring has been debated in public policy venues on countless occasions (Kiviat 2019b; Krippner 2017; Krippner and Hirschman 2022).

In these debates, there is plenty of concern for comparative justice. Policymakers, consumer advocates, and others routinely raise the issue that even though credit scores do not directly rely on attributes such as sex, race, ethnicity, and geography, using such scores in decisions about lending, insurance, and housing might nonetheless lead to problematic differences across groups defined by these and other socially meaningful traits. ¹⁷ Arguments about disparate impact, proxy discrimination, and so on rest on comparisons between people understood to be different sorts of cases (e.g., men vs. women). The comparative fairness questions that preoccupy the algorithmic fairness literature reflect real-world concerns, and the materiality of credit scoring algorithms and the credit bureau databases they depend on lend themselves to answering those questions. With borrowers rendered as cases, it is possible to compare millions of them at a time, revealing patterned social inequalities. Credit records and algorithms may be proprietary, but credit bureaus have shared enough information over the years for policy actors to raise alarm at such differences.

Yet public policy debates about credit scores also revolve around noncomparative fairness. This often takes the form of contestation over the sorts of information that can rightly be fed into algorithms—that is, the aspects of a person’s life that organizations should use in shaping the distribution of socially important resources. Consider, for example, long-standing pushback—from state attorney generals, the Consumer Financial Protection Bureau, and others—against including medical debt in credit scoring algorithms (Best, Bosco, and Wu 2019; Carrns 2022; Consumer Financial Protection Bureau 2022). Here, the issue is noncomparative in that policymakers question whether having (unpaid) medical debt is a legitimate standard for making a decision about what sort of credit, insurance, or housing a person receives. Irrespective of how it may help predict future behavior, medical debt is seen as inappropriate grounds for making such decisions—a perfect example of the sort of “non-comparative harm” Fazelpour and Lipton (2020:8) find missing in the algorithmic fairness literature.

Policymakers invariably express this by reaching beyond credit records and bringing narrative into the debate. In explanations across domains, proponents of eliminating medical debt from credit scores focus on how this debt is rarely incurred intentionally. A person becomes ill or has an accident and faces no choice but to receive high-priced medical treatment—in some situations, such as being rushed to the hospital, a person may not even be conscious. Arguments against using medical debt also highlight the role of social structure (e.g., a lack of universal health care) and other actors, such as insurance companies that are responsible for medical costs but may fail to pay bills fast enough to keep bad debt off credit reports. In other words, how a person comes to be represented as a particular sort of case (e.g., one with delinquent medical debt) is key to understanding the moral valence of being categorized that way. Notably, none of this relies on making comparisons between individuals or groups. The claim of injustice arises from a person—even just one, single person—being held accountable for something that is not their fault. ¹⁸

Narrative works so well in articulating such noncomparative qualms because narrative captures how events have unfolded (or may have unfolded or may unfold) in a person’s life. As a cultural form, narrative foregrounds factors such as a person’s agency in making events happen and the role of external forces and actors, which can be key details in deciding whether a person deserves the treatment they receive (Feinberg 1970; Rachels 1991; Sher 1987). Moreover, the ambiguity of narrative’s temporal boundaries allows us to peer into both the past and the future. Credit bureau records can be arranged chronologically, to be sure, but doing so is an exercise in sequencing, not sensemaking. Only with narrative’s entanglement of character, setting, and plot does meaning emerge.

Indeed, stories about what gives rise to high or low credit scores pervade debates about scoring algorithms. For example, in hearings about the use of credit scores in insurance pricing, legislators routinely testify that it is unfair for rates to go up for senior citizens whose low scores come from decades of prudently avoiding borrowing (Kiviat 2019c). In telling the story (of either a specific senior or seniors writ large), a sparse credit history takes on new meaning: It becomes the mark of a life virtuously led. Likewise, consumer advocates question the increasing inclusion of utility bill payments in credit scores, concerned that in the winter, poor people may let skyrocketing heating bills temporarily slide and later be penalized for that decision (Wu 2012). Here, it is not the past but the imagined future that confers meaning on the present. A person who intends to repay ought to be given special moral status, or at least that is the normative position a certain narrative lets us consider.

In these situations, people use narrative to inject information about actors’ motivations and intentions, future plans (e.g., making the utility company whole once spring comes), and external constraints and structural forces. These details—the sort stripped away when people are rendered as cases—then shed light on whether algorithms disadvantage people when they do not deserve to be. In other words, narrative helps sort out whether the standards people are held to are fair.

This is not to say that filling in the details of noncomparative justice is easy. There can be much disagreement about which standards are the right ones. For example, in debates about credit scores in car insurance pricing, two standards perpetually clash: one that looks to the future and says drivers predicted to file insurance claims rightly pay more and one that also looks to the past and says among those predicted to be costly, only those who have demonstrated irresponsible behavior ought to face a surcharge (Kiviat 2019c). For those backing the second standard, high rates should not follow if a low credit score is not a person’s fault or if the link between scores and claims is driven by socioeconomic standing—for example, poor people both struggle to keep good credit and consistently file claims when bad things happen because they cannot pay for claims out of pocket. Different actors in the debate mobilize different stories, but whatever story they tell, the point is that it is a story. ¹⁹ The discursive terrain on which moral evaluation takes places is one of narrative.

Yet to underscore an earlier point, this does not mean narrative is a better tool for moral evaluation overall. Rendering people in narrative lets us see some moral fault lines, and rendering them as cases allows us to see others. In fact, at times, the two cultural forms work in concert. Consider, for example, that simply wanting allocative systems to treat people equally says nothing about which people ought to experience equivalent outcomes. Why, for instance, is it so important in the United States to test credit scoring (and other) algorithms for different outcomes by race? Because of the history (i.e., the story) of hundreds of years of subjugation and marginalization, in which powerful people and institutions limited the agency and life chances of Black Americans. Rendering people as cases may help make inequalities visible, but narrative is what deciphers which inequalities are the important ones to nullify in the first place.