The Declining Significance of Classification

Rogers Brubaker

Abstract

Scoring and classification systems are ordering devices that make the social world legible and tractable. Both mediate between difference and inequality. But scoring systems organize and represent difference in gradational rather than categorical terms, through measurement rather than classification. This alters the relation between difference and inequality. Who gets what is linked not to who is what, as inferred from stable and publicly legible categorical identities, but to who does what, as recorded in variable and machine-readable behavioral traces. Scoring renders difference as performance; classification renders difference as identity. This comment therefore argues that individualized scoring algorithms and, more generally, new calculative technologies enabled by machine learning point to the declining significance of classification.

Greta Krippner and Daniel Hirschman’s “The Person of the Category: The Pricing of Risk and the Politics of Classification in Insurance and Credit” (2022) addresses crucially important questions about the ways in which the constitution of both groups and individuals—as well as the relationship between them—is mediated by changing technologies of classification, quantification, and analysis. The article is conceptually sharp, historically rich, empirically grounded, and theoretically illuminating. The wonderfully rich footnotes alone are worth the price of admission. The article represents the old Theory and Society at its best.

My comment is prompted by engagement rather than by disagreement. I find compelling both the concrete comparative historical argument about insurance pricing and credit scoring and the more abstract argument about the social and political implications of class-based and attribute-based systems. What I want to propose is therefore not a critique but a way of extending—and in part reframing—the paper’s concluding discussion of the broader significance of changing calculative technologies.

Krippner and Hirschman situate their argument on the terrain of the politics of classification, and they analyze the intellectual technologies developed and deployed in the credit and insurance industries as technologies of classification. But is the notion of classification perhaps being stretched too far? If “calculations of individualized scores” represent the “(digital) future” rather than the “(analog) past” (714), and if such scores entail the “dissolution of aggregates” and potentially the “dissolution of the individual as acting subject” (715), is it then helpful to conceive of such individualized scoring as a particular kind of “classification system” (714)? More generally, do artificial intelligence and machine learning, insofar as they “accelerate and amplify the trends we associate here with the early development of credit scoring,” represent the “future of the politics of classification” (714n)? Or might we instead understand the increased prominence of individualized scoring—and, by extension, of certain other calculative technologies used in algorithmic governance and supported by machine learning—as diminishing the significance of classification?

Classification requires classes. As Krippner and Hirschman’s paper clearly shows, the attribute-based systems that emerged in credit scoring, unlike the class-based systems used in the pricing of insurance, do not generate classes. The attribute-based logic enabled by multivariate statistics (and now by machine learning) generates scores, not classes. As the number of attributes taken into account in generating scores increases, Krippner and Hirschman suggest, attribute-based systems tend toward assigning each individual a unique score. Thus attribute-based systems do not classify individuals; they score them. Scoring, I would argue, is not a form of classification; it is an alternative to classification.

This might seem like semantic nitpicking, a distinction without a difference. Both scoring and classification, after all, are devices of ordering, ways of making the social world at once legible and tractable. In Bourdieusian language, both are principles of vision and division of the social world; both powerfully shape subjectivities; and both, as I will argue below, link difference and inequality. But they work in different ways and produce different consequences, for reasons well specified by Krippner and Hirschman. We know from a rich sociological, anthropological, and historical literature that classifications often become focal points of organization and sites of contestation and resistance; as Krippner and Hirschman remind us, they are at least potentially collectivizing technologies. Scoring systems, by contrast, “detach individuals from groups…altogether” (709); they are powerfully individualizing, even atomizing technologies (719), fostering an ethos of self-entrepreneurship (Fourcade and Healy 2017, 19–20; Brubaker 2020, 33–39).

Scoring and classes are, of course, not mutually exclusive. Scores may be grouped into classes and communicated in categorical rather than—or in addition to—numerical form. FICO groups its credit scores, for example, which range from 300 to 850, into five categories, labeled exceptional, very good, good, fair, and poor. But these categories are interpretative aids to the consumer, not operative categories for creditors. They allow consumers to know roughly where their score falls within the distribution of scores, but the thresholds between them have no general significance for the granting or denying of credit. Creditors do use cutoff values in determining the particular forms of credit to offer and at what prices, but these vary from creditor to creditor and from project to project (FICO 2023; FICO n.d.).

Krippner and Hirschman observe that creditors, like insurers, are “in the business of classification” (706). But in what sense is this the case? To be sure, creditors must draw distinctions and make “decisions about who should receive credit and what interest rate they should be charged” (706). And there is a sense in which such decision-informing distinctions entail classification: each distinction divides some population into at least two classes (in the purely logical sense of “class”) whose members are to be treated differently.

These local acts of line-drawing, however, are what one might call nonce-classifications: they create single-use, narrowly context-bound classes. Such nonce-classifications are far removed from the enduring, consequential, and massively cumulative systems of social classification on which the historical and social scientific literature on classification has focused. As the latter have changed—in the domains of sex, gender, and sexuality, for example; or race, ethnicity, and nationality; or medicine and psychiatry—the “space of possibilities for personhood” (Hacking 1986, 229) and the space of possibilities for group formation (as Krippner and Hirschman note) have changed as well. These systems of classification have been deeply implicated in the structures and processes of durable “categorical inequality” that link categorical difference to inequality, who is what to who gets what (Tilly 1998; Brubaker 2015). For that reason, they have been the focal points of what Bourdieu calls “classification struggles.”

Even if the nonce-classifications generated by particular lenders’ use of particular credit score thresholds are relatively inconsequential, the scoring systems used not only in the credit industry but in an increasingly wide range of domains are hugely consequential. And they certainly contribute to durable, cumulative forms of inequality by linking difference and inequality (Fourcade and Healy 2013). But systems of scoring construe difference in a fundamentally different way than systems of classification do: they organize and represent difference in gradational rather than categorical terms. Difference is constituted, then, through measurement rather than classification. This fundamentally alters the relation between difference and inequality. Who gets what is linked not to who is what, as inferred from stable and publicly legible categorical identities, but to who does what, as recorded in variable and machine-readable behavioral traces. Scoring renders difference as performance; classification renders difference as identity.

I am therefore reluctant to subsume scoring under the overarching rubric of classification as Krippner and Hirschman do, following an influential line of analysis developed by Marion Fourcade and Kieran Healy. For Fourcade and Healy (2013, 560), credit scoring techniques generate “classification situations”—a concept they propose by analogy with Weber’s “class situations”—that have “distinctive and class-like effects on life-chances and social identities.” Their analysis of the new forms of stratification engendered by the widespread use of credit scoring technologies is brilliantly illuminating, but it leaves the notion of classification unspecified. Fourcade and Healy characterize market institutions as “inveterate classifiers.” But their gloss on this—“they count, rank, measure, tag, and score on various metrics” (ibid, 562)—suggests that classification figures in their argument as an extremely general umbrella concept for all ways of organizing and apprehending the social world in a differentiating manner, gradational as well as categorical.¹ It is a fruitful move to bring into focus distinctive ways of “seeing like a market” (Fourcade and Healy 2017). But classification in a more specific sense—involving the establishment of a set of categories or classes and the assignment of instances to one category or another—is arguably much less central to seeing like a market (or to seeing like an algorithm) than it is (or was, in the age of “high modernity”) to seeing like a state.² The assimilation of scoring to classification blurs the difference between gradational and categorical ways of ordering the world and making it legible, calculable, and tractable.

Scoring is regularly paired with classification, the calculation of numbers with the drawing of lines. In the credit industry, as I suggested above, this happens when a creditor offers credit at a particular price to those scoring above a particular threshold on a particular metric. But even when they are thus paired—even when a continuous distribution of scores is segmented into discontinuous categories—the scoring and the classification remain distinct moments, produced by different intellectual technologies and undertaken by different actors. A “consumer reporting agency” or CRA (generally Experian, Trans Union, or Equifax) calculates the scores by applying a proprietary scoring formula (provided most often by FICO) to regularly updated data furnished by credit-card companies, mortgage lenders, and others.³ A lender (or landlord, insurance company, utility company, etc.) specifies the threshold and makes decisions accordingly, accessing the scores from one or more CRAs. Neither FICO nor the CRAs specify thresholds (except, as noted above, as rough consumer-facing guidelines as to where particular bands of scores fall in the overall distribution of scores). The scoring agencies are in the business of differentiation, but they are not in the business of classification; they do not establish classes. The creditors who allocate credit using score cutoffs do generate classes (in the minimal logical sense discussed above) of eligibles and ineligibles, yet the thresholds—and therefore the classes—vary from creditor to creditor and from configuration to configuration for a given creditor.⁴ This works against the reification of particular thresholds and the solidification of sets of people with credit scores within particular ranges into durable and recognizable classes (Rieder 2016, 49).⁵

In the closing pages of their article, Krippner and Hirschman suggest—correctly, in my view—that big data and machine learning are amplifying the scrambling of social identities introduced by attribute-based scoring systems (717).⁶ This arguably means that the burgeoning field of algorithmic governance as a whole—not just credit scoring—is fundamentally built upon gradational and individualizing rather than categorical and collectivizing ways of knowing and acting on the social world. Of course, lines must still be drawn, and thresholds specified.⁷ But to the extent that machine learning and superabundant behavioral data allow people to be apprehended, tracked, identified, targeted, nudged, and governed as individualized, continuously updated data-bundles rather than as stably identifiable members of enduring social categories, then those pre-existing categories lose much of their operative significance⁸ and classification (in the strong sense, pertaining to enduring classification systems of broad, cross-domain reach) becomes less important as a technology of governance.

This accelerating shift, as Krippner and Hirschman argue, has profound implications for group formation, the shaping of subjectivity, and the possibilities for collective action. It also has profound implications, as I suggested above, for the relation between difference and inequality. The mechanisms that mediate between social differentiation and unequal outcomes—and the ways in which social differentiation is made legible and actionable—are increasingly gradational rather than categorical.⁹ They depend on an infrastructure of quantification, commensuration, datafication, ubiquitous tracking, and machine learning, rather than on systems of social classification. Credit scores, with their proliferating off-label uses, are one key example of the gradational linkage between difference and inequality (Rona-Tas 2017). But so are the algorithmic systems used in the hiring, assessment, and termination of workers; in bail and sentencing decisions and in educational settings; and in the generalized metrics that purport to capture civic trustworthiness or civic merit emerging in China, known as the “social credit system” but going well beyond the domain of creditworthiness.¹⁰

Systems of algorithmic governance, like earlier procedures built on notions of “mechanical objectivity,” have been promoted in part for their putative blindness to the categorical distinctions on which formal or informal systems of categorical exclusion or differential treatment have rested.¹¹ These claims to procedurally guaranteed impartiality and objectivity have been widely and legitimately criticized; a large literature addresses various forms of algorithmic bias and pursues strategies for reducing or overcoming such bias.¹² I do not wish to enter into these debates. What I want to underscore is that even a perfectly impartial system of algorithmic governance, however defined, would not reduce the degree of inequality; it would simply change the form of the relation between difference and inequality.

The shift from categorical to gradational ways of construing difference may even intensify certain forms of inequality by enabling powerful, data-rich firms or states to capture additional value or tighten webs of control through finer-grained discriminations. As Krippner and Hirschman note, it is often claimed that data correlated with basic social categories like race or gender serve as proxies for those categories (Harcourt 2015). But from the perspective of a profit-maximizing firm, a data-hungry state, or any organization interested in prediction (in the domains of medicine, criminal justice, or child welfare, for example), categories have long served as proxies for previously unavailable quantitative data. Categories are lossy approximations; the shift from categorical to gradational data can therefore facilitate a tighter mapping of difference onto inequality. Extrapolating certain contemporary trends into the realm of social science fiction—yet a fiction close enough to our own realities to be worth paying attention to—one can envision a post-categorical “meritocratic” dystopia in which rewards and punishments, opportunities and restrictions are meted out in precise proportion to scores and predicted probabilities calculated from behavioral trace data. Such a “meritocracy” would be made all the more dystopian not only by the error- and bias-ridden nature of systems of algorithmic governance and the data upon which they rely, but perhaps even more by the predictive accuracy of such systems.¹³ The declining significance of classification in an age of algorithmic governance, then, is at best an ambivalent phenomenon, not an emancipatory development to be celebrated.

Competing Interests

The author has no competing interests to declare.

Sources Cited

Ajunwa, Ifeoma, and Daniel Greene. 2019. “Platforms at Work: Automated Hiring Platforms and Other New Intermediaries in the Organization of Work.” In Work and Labor in the Digital Age. Vol. 33, Research in the Sociology of Work, edited by S. P. Vallas and A. Kovalainen. Emerald Publishing Limited.

Brubaker, Rogers. 2015. Grounds for Difference. Harvard University Press.

Brubaker, Rogers. 2020. “Digital Hyperconnectivity and the Self.” Theory and Society.

Fannie Mae. 2025. “Eligibility Matrix.” https://singlefamily.fanniemae.com/media/20786/display.

FICO. 2023. “Understanding Fico Scores.” https://www.myfico.com/credit-education-static/doc/education/Understanding_FICO_Scores_5181BK.pdf.

FICO. n.d. “What Is a FICO Score?” https://www.myfico.com/credit-education/what-is-a-fico-score.

Fourcade, Marion. 2016. “Ordinalization: Lewis A. Coser Memorial Award for Theoretical Agenda Setting 2014.” Sociological Theory 34(3): 175–95.

Fourcade, Marion, and Jeffrey Gordon. 2020. “Learning Like a State: Statecraft in the Digital Age.” Journal of Law and Political Economy 1(1).

Fourcade, Marion, and Kieran Healy. 2013. “Classification Situations: Life-Chances in the Neoliberal Era.” Accounting, Organizations and Society 38(8): 559–72.

Fourcade, Marion, and Kieran Healy. 2017. “Seeing like a Market.” Socio-Economic Review 15(1): 9–29.

Hacking, Ian. 1986. “Making Up People.” In Reconstructing Individualism: Autonomy, Individuality, and the Self in Western Thought, edited by T. C. Heller and C. Brooke-Rose. Stanford University Press.

Harcourt, Bernard E. 2015. “Risk as a Proxy for Race: The Dangers of Risk Assessment.” Federal Sentencing Reporter 27(4): 237–243.

Huq, Aziz Z. 2019. “Racial Equity in Algorithmic Criminal Justice.” Duke Law Journal 68: 1043–1144.

John-Mathews, Jean-Marie, and Dominique Cardon. 2022. “The Crisis of Social Categories in the Age of AI.” Sociologica 16(3): 5–16.

Kellogg, Katherine C., Melissa A. Valentine, and Angéle Christin. 2020. “Algorithms at Work: The New Contested Terrain of Control.” Academy of Management Annals 14(1): 366–410.

Kleinberg, Jon, Jens Ludwig, Sendhil Mullainathan, and Cass R. Sunstein. 2019. “Discrimination in the Age of Algorithms.” Journal of Legal Analysis 10: 113–74.

Krippner, Greta R., and Daniel Hirschman. 2022. “The Person of the Category: The Pricing of Risk and the Politics of Classification in Insurance and Credit.” Theory and Society 51(5): 685–727.

Liu, Chuncheng. 2019. “Multiple Social Credit Systems in China.” Economic Sociology: The European Electronic Newsletter 21(1): 22–32.

Porter, Theodore M. 1995. Trust in Numbers: The Pursuit of Objectivity in Science and Public Life. Princeton University Press.

Rieder, Bernhard. 2016. “Big Data and the Paradox of Diversity.” Digital Culture & Society 2(2): 39–54.

Rona-Tas, Akos. 2017. “The Off-Label Use of Consumer Credit Ratings.” Historical Social Research / Historische Sozialforschung 42(1(159)): 52–76.

Scott, James C. 1998. Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed. Yale University Press.

Scurich, Nicholas. 2018. “The Case against Categorical Risk Estimates.” Behavioral Sciences & the Law 36(5): 554–64.

Tilly, Charles. 1998. Durable Inequality. University of California Press.

Whitman, Madisson. 2020. “‘We Called That a Behavior’: The Making of Institutional Data.” Big Data & Society 7(1).

Wynants, Laure, Maarten van Smeden, David J. McLernon, Dirk Timmerman, Ewout W. Steyerberg, and Ben Van Calster. 2019. “Three Myths about Risk Thresholds for Prediction Models.” BMC Medicine 17: 192.

Young, Michael Dunlop. 1958. The Rise of the Meritocracy, 1870–2033; An Essay on Education and Equality. Thames and Hudson.

The Duality of Categories and Attributes

Marion Fourcade and Kieran Healy

Abstract

Krippner and Hirschman’s “The Person of the Category” presents rich historical evidence to support an argument with which we largely agree. We suggest they can simplify the task before them, however, by not pre-judging the possibilities for mobilization under attribute-based systems and by recalling that categories and attributes are dual to one another. It is the continuous, pervasive, and above all plausibly behavioral character of data-collection that matters most to modern classification and scoring systems, rather than whether the information is statistically represented in category- or attribute-based form. We discuss the character of politics in systems of this kind, emphasizing the continued salience of social categories within such systems, the tendency of attribute-based methods to generate new political fault lines, and the role of regulation in governing the scope and limits of broadly “algorithmic” social control.

In “The Person of the Category,” Greta Krippner and Daniel Hirschman (2022, henceforth K&H) examine how algorithms work as technologies of classification. They contrast two empirical cases, insurance pricing and credit scoring, which they argue “present paradigmatic features of pre-digital and digital forms of classification, respectively” (K&H, 688). Their analysis is careful, and they are fully aware of how, historically, each case is neither a pure instance of its type nor historically insulated from the other. Still, they argue that the cases represent two meaningfully different and politically consequential ways of measuring and classifying people. In their view, insurance pricing is class-based. It works by “assigning individuals to membership in a group” (K&H, 689). This actuarial approach puts individuals in multi-dimensional categories like “Young, rural, woman driver”. Credit scoring, by contrast, is attribute-based. It works by “considering an individual’s values on a series of variables” (K&H, 689) and results in a score of some sort that attaches to them as an individual. “The critical distinction between these two systems of classification” they argue,

is that in class-based systems each individual is given the same value as all other members of the group to which she is assigned, whereas attribute systems attempt to give each individual her own unique value or score, as closely as this can be approximated (K&H, 689).

For K&H, this matters because, at root, a class-based system is easier for people to mobilize around for the purposes of collective action, whereas an attribute-based system is atomizing. The former kinds of categories—associated with a “pre-digital” era—are, in their terms, broadly “actuarial.” Actuarial categories are legible. Their components already exist out in the world. Identifying oneself through them is subjectively straightforward. For instance, “I am a young woman living in a rural area”. They also offer a basis for action: “People like me should not be discriminated against”. Attribute-based systems, meanwhile, are associated with the digital era. They are broadly “statistical” (though K&H do not use this term). This makes for much less in the way of actionable material for either self-identification or collective action (“My score is 627”).

K&H flesh out this idea with a comparative analysis of the insurance-pricing and credit-scoring cases. They make a plausible argument. It is true that, in general, a socially entrenched category such as “Unmarried Woman” is easier to identify with as a matter of subjective experience than a detached number like “627”. It is also easier to mobilize around the former than the latter when it comes to any kind of collective or political aim.

The remainder of this comment focuses on two linked points and their consequences. The first is that, as a general matter, the degree to which particular social categories are entrenched is a variable rather than a constant (Starr 1992). Category-based systems can produce empty or useless categories; attribute-based systems may throw up more opportunities for new identities than we might expect. The second is that, despite being different in their immediate presentation and often materially different in their immediate consequences, category- and attribute-based ways of presenting data are dual to one another (Breiger 1974). That is, they are aspects of the same thing, not fundamentally different ways of representing the world. If you have one, you can get to the other. While attribute-based systems are often more opaque than class-based ones in the way that K&H describe, they too have the potential to generate categories that may become bases for social identities or political action. Insofar as modern systems militate against group-formation or collective action in the way K&H have in mind, however, their “attribute-based” character is not to blame. The high-volume, large-scale, real-time and above all behavioral character of the data in these systems is far more consequential.

The Character of Categories

It is helpful to begin with a difficulty that K&H address early on in their article. They argue that “actuarial”, class-based systems are more amenable to group formation than attribute-based ones, especially when the classes are already socially salient. However, in order to make their argument, they must first (K&H, 690–694) clear the ground by dealing with an older but influential view about the insurance industry that claims actuarial systems cannot do this. Jonathan Simon’s critique of “actuarialism” developed the idea that the categories created by insurance companies and their ilk were “singularly sterile” (Simon 1988, 789); i.e., that they were “alive only in the imagination of the actuary who calculates and tabulates and not in any lived form of human association” (K&H, 691). This is awkward for K&H’s argument because, in their view, it is actuarial categories that allow for the identification of (and with) the “person in the category.” They want the attribute-based approach to be the sterile one.

K&H make a three-pronged response to Simon. First, they argue that many actuarial classifications draw on pre-existing and already chronically entrenched categories like “gender”, thus making it difficult to fully “sterilize” or “de-moralize” them. Second, they claim that even the more abstract or “artificial” categories in insurance tables “are still potential collectivities that can under particular circumstances be activated” (K&H, 693). And third, they assert that it is the more modern, attribute-based systems that provide a much better target for Simon’s original critique. Not only do such systems treat individuals as collections of attributes, they also routinely change which attributes are deemed relevant to their scores: “scoring technologies continually swap out predictors… such that there is no stable basis for constructing group membership” (K&H, 693).

Of these three points, the first seems uncontroversially true. To the extent that actuarial classifications incorporate existing categories that are not just recognizable but in fact central to social life and individual experience, they have more than enough raw material for future mobilization. The other two points are less convincing. Or rather, they are not necessary for the overall argument of the paper to succeed. The second point opens the door to the idea that there is variability in the degree to which even seemingly obscure or unlikely categories can be targets of mobilization or, in the longer term, form a basis for identity-formation in the manner of Ian Hacking’s process of “making up people” (Hacking 1986). K&H acknowledge this point (690–691) in the context of Simon’s effort to distinguish between “pre-actuarial” and “actuarial” classifications. Simon thought the former, though often pernicious in motive, did allow for the “space of possibilities for personhood” to expand, whereas the latter did not. K&H reject this distinction with respect to actuarial classifications but wish to re-introduce it for attribute-based ones. We suggest instead that there is no principled basis for deciding ex ante what kinds of classifications can and cannot form the basis of socially real or consequential categories. As K&H show, Simon was wrong to think that actuarial classifications were incapable of this. But in the same way, we contend, it is not necessary to argue that attribute-based classifications are similarly inept. It is better to be a more thoroughgoing Durkheimian on this point and say that, in principle, almost anything can be grist for the mill of human difference-making and boundary-drawing. Of course, this does not mean that every last feature or attribute will in fact seed some widespread social identity or successful political mobilization. But there is no point in tying our hands about it in advance.

K&H’s third point against Simon brings the central issue into view. They say that the fragmentation of subjects Simon complains about is “more fully realized” by attribute-based systems not only because they generate scores rather than groups, but because the basis for calculating those scores is continually shuffled around or swapped out. As we argue in Fourcade and Healy (2024; see also Fourcade 2021), not only do modern systems routinely readjust their scoring methods, thereby putting individuals at risk of sudden reclassification for reasons that will likely be opaque to them, those methods and the classification-situations they create vary across markets. The “same” constellation of attributes can produce different kinds of experience in different market settings. However, the shuffling and swapping-out of attributes is mostly enabled by advances in information technology. It is not an intrinsic feature of an attribute-based approach. If early insurance companies had been able to easily shuffle and swap their categorical configurations, they probably would have. That they did not is more a matter of technological capacity rather than a feature of class-based methods.

Personal Data Makes Categories Dual to Attributes

We are left with the question of whether attribute-based systems really do have the features that K&H describe. These systems represent people as clusters of attributes and generally use those attributes to calculate individualized scores. Those scores really are, to begin with, less conventionally “accessible” or “legible” than a group-based classification. Nevertheless, we should not reify the distinction between attribute-based and class-based methods. They are dual to each other in much the same sense that persons and groups are dual to one another. The classic statement of this idea, which goes back to Georg Simmel, is Breiger (1974).¹⁴ The insurance classes and categories K&H discuss are just aggregates of individual characteristics, or attributes. By the same token, scores attributed to individuals can be characterized as the result of membership in cross-classified categories.

This is perhaps easiest to show by way of example. Conveniently, and not coincidentally, the long tradition of studying and modeling the relationship between individual attributes and cross-classified tables of counts often draws on examples from the insurance industry.

Consider Table 1, which presents counts of car accidents occurring across a number of categories of person. The data are taken from the sample datasets accompanying Agresti (2012). For our purposes, the details about the data and its limits are less relevant than its form. The table presents a simple cross-classification of categories that is directly analogous to the more complex example shown in K&H’s Figure 1 (reproduced on page 30 below), in which cells in the insurance pricing table are labeled with codes designating what sort of person is in each group. For example, code 8291 corresponds to “Unmarried female, with driver training, aged 20, driving for pleasure use or farm use”, all within the broader category of “Youthful operator, not eligible for Good Student Credit”. Each cell is associated with a “factor” (e.g., 1.20) that corresponds to some adjustment to the insurance premium charged to people falling into this cell. This premium, for its part, would be constructed from a table of counts of accident rates like our Table 1. The more accidents that people in a particular cell are involved in, the higher their insurance premium will be.

To mimic K&H’s Figure 1, we can construct a “factor” that captures the relative degree of risk for each group by fitting a model to the cells in Table 1. We want to capture variability in accident or injury rates. We are interested in the relative risks of car accident injuries for our different groups.¹⁵ Again, the details are irrelevant here, but we can write a model (either a logistic regression on individual accidents or, equivalently in this case, a Poisson model of the cell counts in the table) estimating how much accident risk is associated with each group. We would end up with something like Table 2. Here, the column labeled “Factor” is our estimate of riskiness, which would serve as the basis for adjusting insurance premiums and is analogous to the one shown in K&H’s Figure 1.

So, we can take a table of injury rates cross-classified by different kinds of driver and construct a risk score for each category based on differences in accident rates between groups. However, should we wish, we can equivalently represent this model as a series of attributes. Regression coefficients attaching to variables would then be summed to produce a predicted value, or score, for any given individual. In order to get a score for any combination of attributes, we tell the model what attributes we want, and we get some number out the other side. If we wanted, we could convert these coefficients into a more user-friendly look-up table of the sort shown in K&H’s Figure 2 (reproduced on page 31 below), so that a scorer would just have to add up numbers to get a risk score interpretable on a scale we would also provide.

The point is that these are just two different ways of looking at the same thing. We can move between them as we wish. If we want to emphasize the classes, we can look at the data as a table of cross-classified categories and the risk factors associated with each group. If we want to emphasize individual attributes, we can look at the same data as a series of personal scores. Table 3 shows ten randomly selected individuals from our original data and their personalized predicted scores. (We have added a measure for Age to our model, just to show that we can extend the approach to attributes with more than two categories.) In a model like this, there will be as many unique scores as there are unique combinations of attributes across the individuals observed in our data.

A reader might object that modern attribute-based systems, like credit scoring methods, are far more complicated than this toy example; that they involve hundreds or thousands of attributes; that they are changing all the time; and that individuals under these circumstances are in no position to establish their categorical membership. These are all perfectly reasonable responses. But it remains that there is nothing in either the data or the modeling that makes one approach an intrinsically “class-based” and the other an intrinsically “attribute-based” way of seeing the world. In a quite literal way, they are the same thing seen from two different perspectives.

Moreover, categorical representations can be complex, too. K&H’s Figure 1, for example, shows just a small part of a table with a total of 161 cells or classes.¹⁶ That is a complex classification that would be difficult to grasp all at once. It probably contains quite a few cells that are very sparsely populated, that is, category combinations that are not especially common or typical in terms of social identities. Indeed, this sort of complexity—the way actuarial cross-classifications tend to expand multiplicatively—is what led to Simon’s critique of the sterility of actuarial schemes in the first place. By the same token, the “attribute-based” credit scorecard shown in K&H’s Figure 2 consists of numbers that add up to a score. But each of those numbers is derived from a categorical variable. That is what the table of cells is. Certainly, the specific score one ends up with on that scorecard does not correspond directly to a social category, in the sense that different combinations of attributes could sum up to the same score. But the same is true of the class-based table in K&H’s Figure 1, just from the other side: Code 8392 (Married male with driver training, aged 19, drives to work) has the same score as Code 8222 (Unmarried Female, without driver training, aged 18, drives to work).

For this reason, it can be a little misleading to speak of the “unique individual score” generated by an attribute-based system and set it against a seemingly much simpler adjustment factor associated with a category system. While it is true that, in an “attribute” view, every unique individual has a score of their own, it is not true that every individual’s score is therefore unique. Hundreds of millions of unique individuals in the United States have a FICO credit score of their own. But a FICO score is simply an integer between 300 and 850. Millions of individuals have the same score.

In practice, even the fine-grained character of attribute systems tends to be employed in a coarsened manner. Both class- and attribute-based systems tend to make-up people (or, less dramatically, define classes of customers or identify market segments). Most obviously, in the case of credit scoring, the gradations of points-based scoring schemes—which, in the end, remain discrete rather than truly continuous—are typically collapsed into much simpler classifications (“Good/Acceptable/Bad”; “Prime/Subprime”). Those classifications look much more amenable to individual identification and political mobilization.

We see this all the time, with greater or lesser degrees of plausibility and with greater or lesser degrees of success. Categories derived from market research may become shorthand for kinds of person; classes developed by survey researchers (“Soccer Moms”) loop back to become the basis for electoral competition; and simple designations like “Good Credit” or “Bad Credit” become recognizable in the market despite the complexity of the underlying scoring technologies and the heterogeneity of people thus classified. These social processes play out in both “category-based” and “attribute-based” systems, again because these are just two views of the same underlying thing. It is more important to think about what the politics of classification looks like across these systems and what it looks like in terms of the processes of data collection and application, than to focus on an invidious distinction between them.

The Demoralization of Politics or The Politics of Demoralization?

The person is what ties category and attribute together, what makes them dual to one another. Modern attribute-based credit scoring systems (and their descendants) have far more dials to turn and switches to flip than the original round of class-based actuarial methods, and they can do it all much faster. But it is a mistake to think one approach precludes the other. Again, they are dual to one another, in both method and substance. If there is a real difference between the older way and the new, it is not the distinction between class and attribute. Rather, it is the tendency—itself the outcome of both political demands and technological change—for newer systems to present themselves as collecting fundamentally behavioral data and to have the capacity to do this in something like real time. Much of the moralizing power of our current engines of eigencapital derives from their claim to capture what people do as a basis for describing the sort of people they are. What matters is the ability to implement a broadly behavioral conception of social data¹⁷ and to capture and represent it as a steady flow of information about choices and decisions, regardless of whether this flow is represented in terms of group categories or individual attributes.

The Duality of Politics

What should we make of the apparent neutralization of actionable categories in systems like credit scoring? One interpretation would style it as a victory for civil rights. Historically underserved groups have fought long battles, which Krippner herself (2017; 2024) has extensively documented, to be evaluated as individuals outside of the socially entrenched categories that, historically, directly and intentionally determined their life-chances, mostly for the worse. In that sense, the widespread implementation of behavioral scoring can be seen as a real political accomplishment, a fulfillment of a core demand of civil rights: to be judged by what one does, rather than who one is.

A second interpretation is more skeptical. Perhaps algorithmic systems are deliberately engineered to conceal ongoing discriminatory practices rather than eliminate them. In this way of looking at things, risk-scoring methods are tortured until they produce apparently non-discriminatory predictions, the better to preserve advantage for some over others. The specific attributes selected for use in models, and the relative weight given to them, are not neutral technical choices but careful decisions calibrated to maintain social hierarchies while appearing neutral and objective.

In a sense, both perspectives are correct. Attributes and categories are dual to one another, and this duality extends to their politics. The move to escape the constraints of the data’s apparent structure—whether “downwards” in the direction of more fine-grained individuality or “upwards” in the direction of a greater degree of group solidarity—is an expression of the tendency for people and their social lives to overflow whatever organizational and institutional matrix is imposed upon them (Fourcade and Healy 2024, 61–66, 260). Whether they foreground individual-level attributes or category-based properties, we should not restrict our theory of their politics to those foregrounded elements. Indeed, we should not restrict ourselves to the formal structure of the data at all. Within every entrenched identity is a collection of persons wanting to be treated as individuals. Inside every individual-level bundle of attributes is a social identity waiting to emerge and act on its own behalf. These dual struggles reflect efforts, flawed but ineradicable, to balance individual justice with collective redress. At any point in time, they may take different forms, depending on the character of legal instruments, the scope and availability of data, and the state of statistics. But they are two faces of the same political coin.

Competing Interests

The authors have no competing interests to declare.

Sources Cited

Agresti, Alan. 2012. Categorical Data Analysis. 3rd edition. Wiley.

Angwin, Julia, and Terry Parris, Jr. 2016. “Facebook Lets Advertisers Exclude Users by Race.” ProPublica, October 28. https://www.propublica.org/article/facebook-lets-advertisers-exclude-users-by-race.

Benjamin, Ruha. 2019. Race After Technology: Abolitionist Tools for the New Jim Code. Polity.

Breiger, Ronald L. 1974. “The Duality of Persons and Groups.” Social Forces 53: 181–90.

Browne, Simone. 2015. Dark Matters: On the Surveillance of Blackness. Illustrated edition. Duke University Press.

Carruthers, Bruce. 2022. The Economy of Promises. Princeton University Press.

Cheney-Lippold, John. 2017. We Are Data. NYU Press.

Corbett-Davies, Sam, Emma Pierson, Avi Feller, and Sharad Goel. 2016. “A Computer Program Used for Bail and Sentencing Decisions Was Labeled Biased Against Blacks. It’s Actually Not That Clear.” Washington Post, October 17. https://www.washingtonpost.com/news/monkey-cage/wp/2016/10/17/can-an-algorithm-be-racist-our-analysis-is-more-cautious-than-propublicas/.

Dressel, Julia, and Hany Farid. 2018. “The Accuracy, Fairness, and Limits of Predicting Recidivism.” Science Advances 4(1): eaao5580. https://doi.org/10.1126/sciadv.aao5580.

Fourcade, Marion. 2021. “Ordinal Citizenship.” British Journal of Sociology 72(2): 154–173.

Fourcade, Marion, and Kieran Healy. 2024. The Ordinal Society. Harvard University Press.

Gandy, Oscar H. 2009. Coming to Terms with Chance: Engaging Rational Discrimination and Cumulative Disadvantage. Routledge.

Hacking, Ian. 1986. “Making up People.” In Reconstructing Individualism: Autonomy, Individuality and the Self in Western Thought, edited by Thomas C. Heller, Morton Sosna, and David E. Wellberry. Stanford University Press.

Harcourt, Bernard E. 2015. “Risk as a Proxy for Race: The Dangers of Risk Assessment.” Federal Sentencing Reporter 27(4): 237–43.

Kapoor, Sayash, and Arvind Narayanan. 2024. AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How To Tell the Difference. Princeton University Press.

Kiviat, Barbara. 2019. “The Moral Limits of Predictive Practices: The Case of Credit-Based Insurance Scores.” American Sociological Review 84 (6): 1134–58.

Kiviat, Barbara. 2021. “Which Data Fairly Differentiate? American Views on the Use of Personal Data in Two Market Settings.” Sociological Science 8: 26–47.

Krippner, Greta R. 2017. “Democracy of Credit: Ownership and the Politics of Credit Access in Late Twentieth-Century America.” American Journal of Sociology 123(1): 1–47.

Krippner, Greta R. 2024. “Gendered Market Devices: The Persistence of Gender Discrimination in Insurance Markets.” American Journal of Sociology 130(3): 595–643.

Krippner, Greta R., and Daniel Hirschman. 2022. “The Person of the Category: The Pricing of Risk and the Politics of Classification in Insurance and Credit.” Theory and Society 51 (5): 685–727.

Lockhart, Jeffrey W. 2023. “Gender, Sex, and the Constraints of Machine Learning Methods.” In The Oxford Handbook of the Sociology of Machine Learning, edited by Christian Borch and Juan Pablo Pardo-Guerra. Oxford University Press.

Mittleman, Joel. 2022. “Intersecting the Academic Gender Gap: The Education of Lesbian, Gay, and Bisexual America.” American Sociological Review 87(2): 303–335.

Norris, Davon. Forthcoming. “Epistemological Slights or How Racial Capitalism Gets Away with Being Racial.” In Marion Fourcade, Greta Krippner, and Sarah Quinn (eds.) Political Economy Rebooted. Duke University Press.

Norris, Davon. 2022. How All Data Became Credit Data: The Logic of Scoring and the Limits to Racial Inclusion in an Algorithmic Age. Ph.D. Thesis, The Ohio State University.

Patterson, Orlando. 2000. “Taking Culture Seriously: A Framework and an Afro-American Illustration.” In Culture Matters: How Values Shape Human Progress, edited by Lawrence E. Harrison and Samuel P. Huntington. Basic Books.

Sadowski, Jathan. 2024. “Total Life Insurance: Logics of Anticipatory Control and Actuarial Governance in Insurance Technology.” Social Studies of Science 54(2): 231–256.

Schoon, Eric W., David Melamed, and Ronald L. Breiger. 2024. Regression Inside Out. Strategies for Social Inquiry. Cambridge University Press.

Simon, Jonathan. 1988. “The Ideological Effects of Actuarial Practices.” Law & Society Review 22(4): 771–800.

Starr, Paul. 1992. “Social Categories and Claims in the Liberal State.” Social Research 59(2): 263–95.

Underwood, Barbara D. 1979. “Law and the Crystal Ball: Predicting Behavior with Statistical Inference and Individualized Judgment.” The Yale Law Journal 88(7): 1408.

Wachter, Sandra. 2020. “Affinity Profiling and Discrimination by Association in Online Behavioral Advertising.” Berkeley Technology Law Journal 35(2): 367–430.

The Social Life of Scores

Greta R. Krippner and Daniel Hirschman

Abstract

We are grateful for the opportunity to clarify our argument with the aid of Brubaker’s and Fourcade and Healy’s generative readings of “The Person of the Category.” With respect to Brubaker’s critique, we maintain that scoring does achieve a kind of classification, albeit one that is distinct from the processes of group formation explored in a now-influential body of sociological scholarship. With respect to Fourcade and Healy’s argument, we defend the importance of attribute-based versus class-based systems for shaping political action by distinguishing between cultural and computational requirements of decision technologies.

It’s a tremendous thrill—and a daunting challenge—to respond to comments on our paper by three scholars whose work has deeply shaped our own. Rogers Brubaker’s (2005; 2015; 2016) field-defining writings on classification, especially his insistence that we should consider “groupness” as contingent and variable rather than thinking of “groups” as fixed and bounded entities, inspired our approach to how insurers and creditors constitute groups differently and why this might matter for understanding the changing contours of political action in the digital age. Similarly, we have closely followed Marion Fourcade and Kieran Healy’s fruitful collaboration since they first began exploring credit scoring as producing “classification situations” (Fourcade and Healy 2013) and later broadened these insights into a general theory of “ordinalization” as a social process and even a form of society (Fourcade 2016; Fourcade and Healy 2024).

Before engaging Brubaker’s and Fourcade and Healy’s astute comments, we first briefly summarize the argument of our original paper for those who have not read it. Broadly, our concern in the paper is to consider how the proliferation of new scoring technologies (popularly, if confusingly, referred to as “algorithms”) across decision domains in society has reshaped potentials for collective action. To get leverage on this problem, we contrast scoring technologies used in determining access to credit to older decision technologies rooted in insurance markets. Insurance pricing relies on what we refer to as a class-based decision technology: individuals are assigned membership in a group defined by a set of characteristics shared with others and then treated as average or typical members of that group. The fact that one is joined with others who hold a few salient characteristics in common both constitutes the group and (potentially) facilitates collective action. By contrast, credit scoring is attribute-based: one’s creditworthiness is evaluated by tallying values on a series of variables (or attributes) deemed important for determining the likelihood that a loan will be repaid. While multiple individuals may share the same score, each would-be borrower arrives at this score through a unique combination of values on the variables scored. As such, the group constructed by the credit score lacks the sociological foundation of the insurance risk class: one does not necessarily share characteristics in common with those who achieve the same score. For this reason, we expect that collective mobilization will be more difficult to achieve where scoring technologies predominate.

To illustrate this argument, we examine the distinct histories of contestation around the use of gender as a classifier in insurance pricing and credit scoring—paradigmatic examples of class-based and attribute-based systems of pricing risk, respectively. Insurance markets in the United States witnessed sustained contestation over gender discrimination beginning in the 1980s. Our analysis shows that the class-based nature of the pricing system allowed gender difference to “show through” other variables used to price the risk of accident. As a result, challenges to gender discrimination in insurance markets persisted, even after gender was legally prohibited as a pricing variable in some states and some lines of insurance (as in the Pennsylvania auto insurance case we examine in our article). Contestation around gender discrimination in U.S. credit markets, by contrast, was short-lived, dissipating shortly after credit scoring was introduced into these markets in the mid-1970s following the passage of a federal anti-discrimination law.¹⁹ Here we argue that the way the credit score effectively scrambles group memberships—because individuals who achieve the same score do not share characteristics in common other than the score itself—militates against sustained collective action.

In commenting on our paper, Brubaker’s main point is that classing and scoring are fundamentally different operations, thus it is misleading to consider scoring as a form of classification at all. Fourcade and Healy critique the paper from another direction, invoking Breiger’s (1974) notion of “duality” to suggest that classes and attributes are aspects of the same thing, “two faces of the same political coin,” as they write. In this regard, Fourcade and Healy suggest that what matters for political mobilization is not whether a decision technology is organized around classes or attributes, but the increasing reliance on data that are voluminous, continually available, and behavioral in nature. Accordingly, our goal in this reply is to attempt to navigate between Brubaker’s and Fourcade and Healy’s differing critiques in order to clarify ambiguities in our own argument about how these two classification technologies enable and disable different kinds of politics. We will ultimately defend our position that both classing and scoring are forms of classification in a certain sense, while acknowledging the way in which our presentation in the paper creates unnecessary confusion around this point. We will also stand by our view that the technologies themselves—that is, their class- or attribute-based nature—do matter for political action, but in ways that Brubaker and Fourcade and Healy help us to articulate more sharply than in our original paper.

Let’s start with Brubaker. He points to our (repeated) assertion that credit scoring aims to give each individual a unique score as meaning that scoring should not be treated as a classification technology. “Classification requires classes” Brubaker argues, and since credit scores do not constitute classes but place each individual in their own cell (or at least aspire to do this), this is at best a single-use “nonce-classification,” a logical operation and not a sociological one. This argument is laid out with precision and care, and on its own terms it is airtight. But this argument hinges on what exactly it means to give each individual a “unique score,” and here we want to revisit and recast our original claims. Upon reflection, we think this formulation (which Fourcade and Healy also call attention to) is misleading. If we clarify the reasons why, we can recover the way in which there is a kind of classification happening even in scoring models, albeit a classification that is not likely to result in the formation of groups in a sociological sense, per our original argument and consistent with Brubaker’s critique.

Let’s remember how these two forms of pricing—class-based pricing in insurance and attribute-based pricing in credit—work in practice. (For convenience, we’ve reproduced the original figures from our paper replicating the insurance pricing table and early credit scorecard in Figures 1 and 2 below.) To price risk, insurers group individuals by identifying characteristics believed to be associated with risk (here, of an automobile accident) and then observe the prior loss history (the actual number of accidents) for this group. On this basis, insurers form an estimate of expected future claims, and they share this cost equally across group members who are assumed to represent equivalent risks to the insurer. Creditors approach the same problem—determining the risk of default on a loan—very differently. Rather than placing each individual into a group with like others, applicants for credit are assigned points on a number of variables, each of which predicts the likelihood of repaying a loan. Tallying up these points results in a score that will be above or below the threshold for receiving credit (set differently by each creditor, as Brubaker notes). Again, the critical idea here is that while an individual applicant for credit might have the same total score as others also applying for credit, they arrive at this score by a different combination of values on variables scored. On this basis, we write in our paper that while “two individuals with the same credit score may share an outcome . . . they do not share a social experience” (Krippner and Hirschman 2022, 708; italics added).

Figure 1

Class-Based Pricing (Insurance Pricing Table).

Source: Government Accounting Office (1979).

Figure 2

Attribute-Based Pricing (Credit Scorecard).

Source: Lewis (1994).

The actual mechanics of these two classification systems matter here. When we talk of individuals in insurance markets occupying a class “together,” there is quite literally a cell (i.e., a risk class) in a table that “holds” those individuals (see Figure 1). A decision about how to price risk is made for each such cell in the table (even if different cells may be assessed equivalent rates, as Fourcade and Healy correctly observe). Superficially, the credit scoring table looks analogous, but as we note in our paper, each cell in the credit scoring table does not hold a group of individuals, or even a single individual. It holds an attribute (a value on a variable used to predict creditworthiness; see Figure 2).²⁰ As these attributes are summed to arrive at a numerical score that determines whether access to credit is granted or denied, the cell in the credit scoring table is not strictly analogous to the cell in the insurers’ risk pricing table: it does not define the point where the decision (to grant or deny access to credit) is made.

What is important here vis-à-vis Brubaker’s critique is that if we are thinking about classification not in terms of what is contained in each cell of the scoring table but in terms of how the decision to grant or deny credit is made by summing across these cells, the credit score does arguably achieve a kind of classification. That is, any given numerical score will be held by multiple individuals who may in this sense constitute a “class” (albeit one with very different characteristics than the class constituted by insurance pricing). Put differently, what is “unique” here is not an individual’s outcome on a decision variable (i.e., a score determining access to credit), but the particular path by which they arrive at this outcome (or score). This, we believe, is consequential for political mobilization: a shared path is more likely to facilitate political action than a path that is idiosyncratic to a particular person. By the same token, the statement we make in the paper that individuals scored by creditors “may share an outcome, but they do not share a social experience” no longer appears quite correct. We meant to suggest that the outcome (i.e., access to credit) is not attached by the credit score to a shared social position (i.e., as defined by age, gender, marital status, and so on) in the same way that is true for insurance pricing. But the fact of gaining or (especially) being denied access to credit is itself a social experience—one that is, in almost all instances, shared with others. And this social experience may in fact constitute a group (e.g., “subprime borrowers”) that might be mobilized politically,²¹ although we think this is less likely to the extent that this experience is not linked to a fixed location in social space.

We can express this in a way that we think clarifies matters by considering whether the class in question exists independently of the decision technology or is produced by it.²² Generally speaking, classification in insurance constitutes a group formed by social categories that exist in the world independently of the technology that produces a decision. In credit markets, by contrast, classification constructs a group (defined by the score) that is produced by the decision technology itself. To be sure, the creditor’s score could, in theory, rely on the same pre-existing social categories that construct the insurance risk class as variables to be scored.²³ But unlike in the insurance case, the attributes used in scoring models are not held constant across individuals, hence they do not define membership in the group. Accordingly, we take Brubaker’s point that scoring technologies do not involve classification in the sense explored by a now-expansive literature in sociology examining durable forms of social difference (e.g., Tilly 1998; Brubaker 2005, 2016; Massey 2007; Loveman 2014; Mora 2014; Rodriguez-Muñiz 2021; Currah 2022). But this is the case not because each individual gets their own “unique score”—as we wrote in our paper—but because the score forms a group detached from the legible social categories that construct class-based pricing.

Here we might simply adopt Brubaker’s (2015) distinction between “categorical” and “gradational” ways of organizing social difference—treating only the former as a “classification”—and call it a day. While we find useful the conceptual distinction between categorical and gradational forms of difference, we aren’t quite ready to let go of thinking of the score as a classification. This is because, as we’ve suggested, the credit score may potentially constitute a class (by collapsing a grade into a category)—a “subprime borrower” is not a “nonce-classification” in the sense that Brubaker means. But to emphasize our original point, and in full agreement with Brubaker, neither is the credit score the same kind of object as the insurer’s risk class. In our view, these are distinct because these two technologies constitute groups in very different ways. To reiterate the key distinction, the insurance pricing table forms a group from social categories that exist independently of the technology that produces a decision (e.g., age, gender, marital status, etc.), whereas the credit score produces a group (if it does) by constructing categories that exist only by virtue of the decision technology itself (e.g., “prime” and “subprime borrower”). Critically, this means that class-based systems will tend to draw on categories that represent hardened forms of social difference. By contrast, attribute-based systems will construct categories de novo that will be more tenuous as vehicles for self-identification and consequently social struggle.²⁴

This brings us to Fourcade and Healy, who make two main points, one which we largely cede, and another which we will attempt to rebut at least in part. The point we cede (as evidenced in our response to Brubaker): there is more variability around the politics generated by attribute-based systems of classification than we allow in our paper. Here, our effort to crystallize the differences between the two classification technologies we examine gave an overly deterministic flavor to our argument. We meant to suggest not that attribute-based decision technologies make collective mobilization impossible, but that they make collective mobilization less likely for all the reasons we have just elaborated. Although we want to avoid a deterministic reading of our argument, we also want to underscore that less likely is still an important patterning to the politics of the digital age. In this regard, while there is a great deal of variation in how political struggles unfold in a society organized around scoring technologies (which Fourcade and Healy elucidate in their rich discussion), there are also strong tendencies toward depoliticization that we can and should examine.

The second issue that Fourcade and Healy raise falls under their discussion of “duality.” Here we have several responses. On its own terms, the idea of duality is very appealing. It captures something profound about the nature of insurance, a long-neglected object of study in sociology that is indeed “dual” in how it deals with individuals and collectives (Ewald 2020). We also appreciate the broader point that attributes and classes can be considered “dual” to each other: attributes can be expressed as classes, and classes as attributes (a point Brubaker also observes in passing). But this observation does not mean that classes and attributes are simply “aspects of the same thing.” The act of “conversion” from attribute to class (or from class to attribute) is critical, and it depends in important ways on the nature of the technology used to produce a decision.

We can go a bit further here. Fourcade and Healy make this argument most forcefully by presenting an extended mathematical illustration in which they show very precisely how classes can be expressed as attributes and how attributes can be aggregated into classes. As a purely logical exercise, the demonstration is irrefutable. On another level, however, Fourcade and Healy conflate the mathematics of classes versus attributes with what we might think of as their phenomenology. What can be done with a mathematical calculation is very different than what can be done in the world, where conventions, norms, knowledge infrastructures, rules and regulations, and other sociological and legal features shape social practices. Insurance is particularly loaded down by such constraints; as a result, it is perhaps not surprising that although insurers could in practice convert classes into attributes, they typically do not. Fourcade and Healy write that “if early insurance companies had been able to easily shuffle and swap their categorical configurations, they probably would have.” But as we know from Dan Bouk’s (2015) work on the history of risk-making practices in the United States, insurers at the turn of the twentieth century already had available technologies that allowed them to price individual risk factors (i.e., attributes). However, insurers applied these techniques only in very limited ways because they were at odds with deeply entrenched cultural understandings, professional norms, and regulatory rules that constituted the risk class as the basic organizational (and operational) unit in insurance markets (Krippner 2019; 2024).

More surprising, we can observe something similar at play in insurance markets today, notwithstanding new machine learning technologies that have vastly advanced capacities to predict and price individual risks. It is often taken for granted that machine learning and artificial intelligence have transformed pricing in insurance, effectively ushering in a new regime of “data” that has displaced probabilistic (i.e., class-based) methods with fully individualized (i.e., attribute-based) understandings of risk (Ewald 2012). But notwithstanding considerable hype around “personalized” pricing in these markets, a robust literature describes “the revolution that did not happen” in applying machine learning to insurance pricing (Francois and Voldoire 2022; cf., Meyers and Hoyweghen 2018, 2020; McFall 2019; Barry and Charpentier 2020; Cevolini and Esposito 2020, 2022; Jeanningros and McFall 2020; Francois 2025). The reasons for this failed revolution are multi-faceted (see Charpentier and Vamparys 2025), but broadly they reflect basic incompatibilities between the notion of risk pooling (which requires the formation of a group) and individualized pricing reliant on behavioral data. Accordingly, when machine learning is introduced into insurance markets, it tends to be layered on top of traditional actuarial (i.e., class-based) techniques: individuals are first assigned to a class defined using standard demographic variables, and then behavioral data may be applied to make incremental adjustments to this base price. Effectively, to the extent risks are individualized in insurance, it is through the creation of classes that are more finely segmented, not though the overthrow of the class altogether (Barry 2020).

For this reason, we are reluctant to accept Fourcade and Healy’s suggestion that what matters for our argument is not whether data are “statistically represented” in class or attribute form but instead the volume, velocity, and especially behavioral quality of these data. While we certainly agree that the volume, velocity, and behavioral quality of data in modern scoring systems is disabling for political contestation, we do not see this as fully separable from the class or attribute form of the data. This is because whether data are organized around classes or attributes is not a simple matter of how they are “represented”—these are objects in the world that have a tangible reality, not mere representations of that reality (cf., Panagia 2021). In this regard, class-based systems tend strongly toward reliance on a finite number of stable categories, not only for computational reasons but more critically for cultural ones (Krippner 2024).²⁵ The insurance risk class is derived from traditions of mutualism in which the risk pool constitutes a “community of fate” established between members of a class connected to each other through socially legible characteristics. The attributes scored by creditors create no such community as each individual navigates a singular trajectory through the scoring table. As a result, attribute-based systems are not constrained to select a few characteristics that connect scored individuals in sociologically meaningful ways (although scores may become sociologically meaningful after the fact, as we have noted). Attribute-based systems are, in this regard, more amenable to the “churning” associated with the large volume and high velocity of data used in machine learning applications, with predictable results in terms of the likelihood of collective action.

There is an important caveat to our argument. The features that we attach to class-based systems may be especially marked in insurance, an archaic social institution in which the risk class has a special, almost mystical, significance (Krippner 2024). Thus, while insurance pricing has been resistant to transformation through engagement with “big data,” the other examples we give in the paper of class-based systems of valuation—marketing and political polling—appear much more permeable to machine learning techniques (see Cheney-Lippold 2011, 2017; Moor and Lury 2018; Brubaker 2023). Even if our argument about class-based pricing applies only to insurance and not more broadly, we still think that our main purpose of illuminating the characteristics of scoring technologies that tend toward disabling political action will have been well served by finding a comparison case that represents the opposed tendencies in their most distilled and potent form.

There is one point on which Brubaker, Fourcade and Healy, and we ourselves converge: the rise of scoring technologies is a deeply ambivalent phenomenon, promising relief from the categorical (class-based) forms of inequality that stratified pre-digital societies, only to reinscribe these inequalities in new, perhaps more pernicious forms (cf., Farrell and Fourcade 2023; Norris Forthcoming). In the face of such challenges, the conceptual clarity provided by Brubaker’s and Fourcade and Healy’s probing analyses has much to offer the study of politics in the brave new world of the score.

Acknowledgements

We thank Junchao Tang for helpful comments on an earlier version of this essay.

Competing Interests

The authors have no competing interests to declare.

Sources Cited

Barry, Laurence. 2020. “Insurance, Big Data, and Changing Conceptions of Fairness.” European Journal of Sociology 61(2): 159–84.

Barry, Laurence, and Arthur Charpentier. 2020. “Personalization as a Promise: Can Big Data Change the Practice of Insurance?” Big Data & Society 7(1): 1–12.

Bouk, Dan. 2015. How Our Days Became Numbered: Risk and the Rise of the Statistical Individual. University of Chicago Press.

Breiger, Ronald L. 1974. “The Duality of Persons and Groups.” Social Forces 53: 181–90.

Brubaker, Rogers. 2005. “Ethnicity without Groups.” In Remaking Modernity, edited by Julia Adams, Elisabeth S. Clemens, and Ann Shola Orloff. Duke University Press.

Brubaker, Rogers. 2015. Grounds for Difference. Harvard University Press.

Brubaker, Rogers. 2016. Trans: Gender and Race in an Age of Unsettled Identities. Princeton University Press.

Brubaker, Rogers. 2023. Hyperconnectivity and Its Discontents. Polity Press.

Cevolini, Alberto, and Elena Esposito. 2020. “From Pool to Profile: Social Consequences of Algorithmic Prediction in Insurance.” Big Data & Society 7(2): 1–11.

Cevolini, Alberto, and Elena Esposito. 2022. “From Actuarial to Behavioural Valuation: The Impact of Telematics on Motor Insurance.” Valuation Studies 9(1): 109–139.

Charpentier, Arthur, and Xavier Vamparys. 2025. “Artificial Intelligence and Personalization of Insurance: Failure or Delayed Ignition?” Big Data & Society 12(1): 1–13.

Cheney-Lippold, John. 2011. “A New Algorithmic Identity: Soft Biopolitics and the Modulation of Control.” Theory, Culture, and Society 28(6): 164–181.

Cheney-Lippold, John. 2017. We Are Data: Algorithms and the Making of Our Digital Selves. NYU Press.

Currah, Paisley. 2022. Sex Is as Sex Does: Governing Transgender Identity. NYU Press.

Ewald, Francois. 2012. “Assurance, Prévention, Prédiction: Dans L’univers du Big Data.” Institut Montparnasse. https://www.institutmontparnasse.eu/wp-content/uploads/2012/11/Collection_recherches_n_4.pdf.

Ewald, Francois. 2020. The Birth of Solidarity: The History of the French Welfare State. Duke University Press.

Farrell, Henry, and Marion Fourcade. 2023. “The Moral Economy of High-Tech Modernism.” Daedalus 152(1): 225–235.

Fourcade, Marion. 2016. “Ordinalization.” Sociological Theory 34(3): 175–195.

Fourcade, Marion, and Kieran Healy. 2013. “Classification Situations: Life Chances in the Neoliberal Era.” Accounting, Organizations and Society 38(8): 559–72.

Fourcade, Marion, and Kieran Healy. 2017. “Seeing Like a Market.” Socio-Economic Review 15(1): 9–29.

Fourcade, Marion, and Kieran Healy. 2024. The Ordinal Society. Harvard University Press.

Francois, Pierre. 2025. “The Old Regime (of Mutualisation) and the Revolution (of Big Data).” The British Journal of Sociology Special Issue: Insurance and Social Theory: 1–8.

Francois, Pierre, and Theo Voldoire. 2022. “The Revolution that Did Not Happen: Telematics and Car Insurance in the 2010s.” Big Data & Society 9(2): 1–12.

Gamerdinger, Alexander. Forthcoming. “Moral Authority Over Risk Classifications: How Data Professionals Shape the Uneven Algorithmization of Life Insurance.” Socio-Economic Review.

Government Accounting Office. 1979. Issues and Needed Improvements in State Regulation of the Insurance Business: Report to the Congress. Washington, D.C.: Government Printing Office.

Jeanningros, Hugo, and Liz McFall. 2020. “The Value of Sharing: Branding and Behaviour in a Life and Health Insurance Company.” Big Data & Society 7(2): 1–15.

Krippner, Greta R. 2017. “Democracy of Credit: Ownership and the Politics of Credit Access in Late Twentieth-Century America.” American Journal of Sociology 123(1): 1–47.

Krippner, Greta R. 2019. “Bearing Risks, Being Watched.” Public Books, February 26. https://www.publicbooks.org/bearing-risks-and-being-watched/.

Krippner, Greta R. 2024. “Gendered Market Devices: The Persistence of Gender Discrimination in Insurance Markets.” American Journal of Sociology 130(3): 595–643.

Krippner, Greta R., and Daniel Hirschman. 2022. “The Person of the Category: The Pricing of Risk and the Politics of Classification in Insurance and Credit.” Theory and Society 51(5): 685–727.

Lewis, E. M. 1994. An Introduction to Credit Scoring. Athena Press.

Loveman, Mara. 2014. National Colors: Racial Classification and the State in Latin America. Oxford University Press.

Massey, Douglas S. 2007. Categorically Unequal: The American Stratification System. Russell Sage Foundation.

McFall, Liz. 2019. “Personalizing Solidarity? The Role of Self-Tracking in Health Insurance Pricing.” Economy and Society 48(1): 52–76.

Meyers, Gert, and Ine Van Hoyweghen. 2018. “Enacting Actuarial Fairness in Insurance: From Fair Discrimination to Behaviour-based Fairness.” Science as Culture 27(4): 413–438.

Meyers, Gert, and Ine Van Hoyweghen. 2020. “Happy Failures: Experimentation with Behaviour-Based Personalisation in Car Insurance.” Big Data & Society 7(1): 1–14.

Moor, Liz, and Celia Lury. 2018. “Price and the Person: Markets, Discrimination, and Personhood.” Journal of Cultural Economy 11(6): 501–513.

Mora, Cristina G. 2014. Making Hispanics: How Activists, Bureaucrats, and Media Constructed a New American. University of Chicago Press.

Norris, Davon. Forthcoming. “Epistemological Sleights or How Racial Capitalism Gets Away with Being Racial.” In Political Economy, Rebooted, edited by Marion Fourcade, Greta Krippner, and Sarah Quinn. Duke University Press.

Panagia, Davide. 2021. “On the Possibilities of a Political Theory of Algorithms.” Political Theory 49(1): 109–33.

Poon, Martha. 2012. What Lenders See: A Pre-History of the FICO® Credit Scoring System. Ph.D. Thesis, University of California, San Diego.

Rodríguez-Muñiz, Michael. 2021. Figures of the Future: Latino Civil Rights and the Politics of Demographic Change. Princeton University Press.

Tilly, Charles. 1998. Durable Inequality. University of California Press.