Recorded on September 22, 2025, this video features a talk by Davon Norris, Assistant Professor of Organizational Studies and Sociology (by courtesy) and Faculty Associate at the Stone Center for Inequality Dynamics at the University of Michigan.

Professor Norris’s research is broadly oriented to understanding how our ways of determining what is valuable informs patterns of inequality with an acute focus on racism and racial inequality. Often, this means he studies the history, construction, and operation of various ratings, scores, and rankings whether that be at the government level (i.e., government credit ratings) or individual level (i.e., consumer credit scores). Other work that comes out of this interest in valuation processes further probes questions related to finance and the role of credit and debt in shaping inequality.

His research has been published in outlets such as Social Forces, Socio-Economic Review, Social Problems, and Sociological Forum, and has received awards from the Future of Privacy Forum and American Sociological Association. His work has been funded by the American Sociological Association. Davon received his Bachelor of Science in Accounting (2014), Master of Arts in Sociology (2018) and Ph.D. (2022) in Sociology all from The Ohio State University.

This talk was presented as part of a symposium series presented by the UC Berkeley Computational Research for Equity in the Legal System Training Program (CRELS), which trains doctoral students representing a variety of degree programs and expertise areas in the social sciences, computer science and statistics.

The event was co-sponsored by Social Science Matrix, the Berkeley Economy and Society Initiative (BESI) Tech Cluster, the Berkeley Institute for Data Science (BIDS), and the UC Berkeley Department of Sociology.

(upbeat instrumental music)

[DANYA LAGOS]

Hello, everyone. Thank you all so much for coming to today’s talk in the series of the fall colloquium for the sociology department, in collaboration with the Social Science Matrix, the Berkeley Institute for Data Science, and the Computational Research for Equity in the Legal System Training Program.

So a big thank you to all of our co-sponsors. It is my distinct pleasure to introduce Professor Davon Norris.

Professor Norris is a three-time Buckeye, having received a BA in Accounting, an MA and a PhD in Sociology, all from the Ohio State University. In what must be a very fun and interesting dynamic as a Buckeye in Ann Arbor, Professor Norris is an assistant professor of Organizational Studies and Sociology at the University of Michigan.

Professor Norris is an economic sociologist who examines the tools that determine what is valuable, worthwhile, or good, and how they are implicated in patterns of inequality, particularly racial inequality. And his research has been published in a number of our discipline’s leading venues, including the Nature of Human Behavior, Social Forces, Socio-Economic Review, Social Problems, and Sociological Forum. And we’re gonna play by sociology rules today for our interdisciplinary audience work.

I’ll take stack, so feel free to all sit towards the front And I’ll face the audience when it’s question time And I’ll take notes to call on you all. Let’s give a warm Berkeley welcome to Professor Norris.

(applause)

[DAVON NORRIS]

Yeah, thank you. Good afternoon, everyone. Thank you all for being here, and thank you, Danya for that introduction.

I wanna give a shout-out to Marion, CRELS, BIDS, and all of the other letters and acronyms that are a part of the reason why I’m here today. And I also wanna give a shout-out to Alina for handling all the logistics and making sure that I actually got here today despite an earthquake.

(audience laughing)

So as mentioned, I’m Davon Norris, and before I actually get into my talk, I have this personal tradition, you might even call it a ritual that I do before talks. So I love sneakers, and because I love sneakers, I always seek out a reason to buy a new pair.

So something I started doing with talks like this is buying a pair of shoes that I think encapsulates something about the school, the place, or just what comes to mind when I think about where I’m giving a talk. So when I sat down last month to think about the right pair of shoes for today’s talk, the thing that I, that my mind always jumps to when I think about the Bay Area is obviously the Black Panther Party, not just because of their politics and style, but also because I lived in Chicago for a period of my life, and when I lived there, I began learning about people like Fred Hampton and just learning more about the history of that organization. So I wanted a shoe for today that reflected the Bay Area via the Black Panther Party and paid a little homage to my time that I spent in Chicago.

And for that purpose, the choice was obvious. I went with the Triple Black Jordan 3s. The Jordan sneaker obviously pulls in Michael Jordan and those, for those hip in the sneaker lingo, you would know that this Triple Black colorway on a pair of Jordans is referred to the Black Cat colorway. So I’m wearing these Black Cat Jordan 3s today, and I’ve been putting all of my experiences and conversations I’ve had with folks into these shoes.

That way if things go really well, I can pull these out and be reminded of all the wonderful conversations that I had. Or if things go poorly, I can just throw these in the back of my closet so that I never have to think about this moment again. And for the graduate students in the room, that’s how you compartmentalize in these academic streets. So what is it that I actually wanna talk about today?

So I’m gonna be talking about this paper that is clunkily titled, Legitimation by Misidentification: Credit, Discrimination, and the Racial Epistemology of Algorithmic Expansion. Now, beyond being a mouthful, this piece is me trying to do my best to make sense of a recent dynamic in consumer credit markets that I don’t think existing accounts have a good or satisfying explanation for. And this dynamic is best exemplified by a 2019 congressional hearing.

This hearing was about evaluating whether increasing the types of data that are used in consumer credit scores could expand access to credit. Now, these new types of data, or so-called alternative data go beyond typical things like amount of debt borrowed, types of loans, and history of repayment to incorporate newer data sources like rental and utility payment history checking account information, and perhaps more concerningly, things like shopping habits, college major and one’s occupation. Now, that we are seeing new kinds of credit scores or new firms that are leveraging newer types of data in their underwriting, right?

Like, that is happening is probably not that surprising, right? That’s kind of what we would expect in this AI, big data, yada yada moment. But I think this is surprising for at least two reasons. So first, it represents a noted shift within Congress.

So from 1968 to 1989, Congress organized hearings, public discussions, and introduced dozens of bills that sought to limit the information used to determine credit access and by extension, used in credit scores. And at the time, Congress scoffed at notions offered by credit scorers of the period like William Fair, co-founder of Fair, Isaac & Company, FICO, when William Fair suggested that all legally available information must be used.

They scoffed at this idea. But today, companies parrot effectively the same claim, where they note that all data is credit data. And instead of experiencing pushback as they did in the 1970s, they’re starting to find openness. Across the aisle, Democrats and Republicans express a growing comfort with expanding the data used in consumer credit scores.

And beyond policymakers, state and local governments stay in tanks, and even some consumer advocacy organizations are increasingly leaning into this idea that we want to expand the data used to create consumer credit scores. Second, I think this is surprising for a second reason, and in fact the most puzzling aspect of this, is that this push for alternative data in credit scoring unfolds largely unaffected by concerns about reinforcing racial inequities. We have a vast literature over at least the last 10 years that has accumulated information that the use of data in scoring functions as a growing problem of racial exclusion.

And in contemporary policy discussions, many critics raise these points. However, those critiques have failed to stymie the march towards more data. Now, that wouldn’t be all that surprising if it weren’t for the fact that proponents of expanding data articulate the expansion in the language of inclusion and racial inclusion.

Right? So the deepening of credit scoring via expanded data that goes into them is not only combating critiques of exclusion or simply ignoring those critiques, but this expansion is able to be positively framed in the language of inclusion and it is resonating as such. So this paper is me trying to make sense of this.

What explains why legislators are comfortable with expanding the data in consumer credit scores today, but were not in the mid-20th century? And how is it possible that alternative data resonates as racially inclusionary in the face of claims and widespread critiques to the contrary? So at the heart of this, this is an instance where we see an effort at deepening and expanding the reach of algorithmic credit scores. So as I’ve been trying to understand the answer to these questions I’ve been going back to and revisiting the prior, what prior research has to say about the mechanisms that drive or enable the expansion of algorithms and data collection across various domains.

And this isn’t the most super cohesive body of work but there are, generally speaking, two broad flavors of explanation. The first of which locates the drivers of algorithmic expansion in internal organizational dynamics. This manifests in more bureaucratic or extractive variations, but the idea is that organizations leverage algorithms and data in their pursuit of efficiency.

Instead of relying on subjective or potentially error-prone human decisions, data-driven assessments aid in standardizing and rationalizing the decisions that companies make. So in the context of credit, this enables lenders to identify who to lend to with greater speed and accuracy.

Hopefully accuracy. And of course this bureaucratic, this push is supported by technological advances that make it possible for organizations to collect and extract or analyze new types of information. In a complementary organizational account, besides this, the bureaucratic focus is what I’m labeling here as sort of extractive, which roots algorithmic expansion in this sort of contemporary data imperative which makes accumulating information paramount to firm survival in this sort of era of capitalism.

This suggests that the expansion of data is less about alleviating bureaucratic problems per se and more about extracting larger profits. And this serves as a material profit-driven impulse to be particularly cunning in collecting, obfuscating, and exploiting information from people. All right, so organizational insights primarily emphasize these internal material interests that organizations have in pushing towards the expansion of data.

External, or what I’m labeling here crudely as political accounts, I don’t really have another good word for this, these external accounts take the position that organizations are at times coerced by, or find opportunities created by legislative efforts. In this regard, one particularly important historical thread illustrates how a mechanism sparking the expanded use of algorithms and data was the shifting political landscape of the 1960s and 1970s. In this period, social movement organizations advocated for the passage of anti-discrimination laws, and these laws had the effect of problematizing more subjective approaches to figuring out who to hire or who to lend to, thereby unwittingly providing an external impetus, pushing organizations towards relying more heavily on data and scoring as a kind of more procedurally rigorous or hopefully more accountable alternative.

A second thread of political explanations, which shares an affinity to the extractive variations of organizational accounts, centers the relatively permissive regulatory regime in the United States around collecting data and using it to develop scores. So in this case, organizational ambitions to gather more data and extend scoring are only realizable because there is a legal context that creates few barriers which enables rather than constraints these organizational ambitions.

And we were talking earlier about GDPR in Europe, can’t do much of what I’m gonna talk about today in Europe, but can in the US. So these are kind of the existing arguments that I’ve been able to at least find that get at the drivers of algorithmic expansion. And while certainly these things play a role, these accounts don’t really seem to fit the contours of the contemporary expansion in credit markets.

It appears that what is aiding the expansion of algorithms in this credit context, right today, is an ability to legitimately identify scoring and more data collection as inclusionary, in general and racially inclusionary in particular. And prior explanations can’t really make sense of this.

In part because prior explanations offer contrasting intuitions about the consequences for scoring and expanded data collection on racial exclusion. Bureaucratic explanations suggest that, in part, by standardizing decisions, data reduces pressures for biases that shape decisions, thereby potentially reducing racial inequality. Extractive variations contend that algorithms are inherently racialized and reinforce and increase inequality and thwart attempts at accountability. And political explanations exhibit a sort of similar kind of ambiguity.

On one hand, they demonstrate that the expansion of scoring was a response to mid-20th century concerns about discrimination, so we might think this has a myriad of impacts on racial inequality. But on the other hand, the lack of robust regulatory and legal barriers creates a context for extractive actors to operate. Now, this ambiguity and contested interpretation matters because it is present in policy discussions.

Some argues that algorithms and more scoring and more data increase or reinforce racial exclusion, while others contend the exact opposite. And rather than producing some like regulatory gridlock, there seems to be a growing consensual orientation towards expansion, which is increasingly, as I mentioned, justified in these inclusionary terms. So not only do prior explanations yield contrasting intuition, but they’re also poorly suited to understanding why a particular interpretation of the racial consequences of scoring has gained traction over others.

So it seems to me, that in order to understand the contemporary expansion of data and credit markets, that we need to understand how scoring has gained traction as a legitimate tool to promote racial inclusion in the face of claims to the contrary. And to make sense of this, I argue we must attend to more epistemic questions that weave through the gaps in prior explanations. What does that even mean?

In this regard, I think Eric Schoon’s recent work on legitimacy is useful. Schoon conceptualizes legitimacy as a relational phenomenon that emerges from how a given object conforms to expectations of an audience to garner support. Schoon’s model emphasizes the relations and processes that generate legitimacy avoiding an essentialist ontology of legitimacy being something an object is or has.

Now, why am I saying this? Well, for the question at hand, I think this is super important and indeed even profound. To be a legitimate tool of racial inclusion, as scoring increasingly resonates, scoring and data must necessarily conform to expectations of being not exclusionary. But crucially, scoring need not actually be non-exclusionary, only interpreted and understood as such.

In other words, the question of whether expanding data and credit scores actually produces genuine inclusion is potentially less important than how institutions come to know the answer. Right?

In the present case, we have a situation where one particular perspective is coming to dominate in the face of opposition, and we need to understand why or how one answer is beating out the competition. And I think Somers and Block offer one influential way of resolving this tension.

They demonstrate that some ideas have an epistemic privilege, meaning they circulate with a sort of built-in plausibility and credibility that make them more compelling than alternatives. In their analysis, they highlight how the so-called perversity thesis, which is this idea that government social supports hurts people in poverty as opposed to helps them, that this idea has its own, in their language, epistemological bootstraps, its own built-in plausibility and credibility that enables it to resonate, and as they argue, gives it power in shaping discourse over hundreds of years. And this use, this insight is useful for me in helping explain why some ideas dominate. Some ideas are just better than others and make more sense than others, regardless of necessarily their sort of underlying truth.

But the focus on an idea that has its own bootstraps means that their analysis leaves to the side the mechanisms through which less intuitive or more contested ideas gain traction. So my sort of general argument is that I argue that ideas without this built-in plausibility require some scaffolding that confers epistemic privilege by guiding interpretations and understanding in certain directions as opposed to others. An example of this is Dan Hirschman’s recent work where he highlights the importance of, ‘Knowledge infrastructures,’ that facilitate the tracking of empirical phenomena. Such infrastructures like the decennial census, or in his case, over a whole host of surveys that left the growth of the top 1% of incomes invisible, guide attention towards tracking and understanding problems in certain ways, but leave invisible significantly other kinds of problems.

Relatedly Beth Popp Berman argues that besides infrastructures to aid empirical identification, there exists more diffuse modes of thinking that provide a conceptual toolkit for how to frame problems ex ante and evaluate and interpret empirical evidence ex post. What does all this mean?

So if we pair these insights, I think we get a sense of there being this sort of underlying epistemological infrastructure composed on one hand of tools of conceptualization, and on the other hand, methods of empirical identification that are important in structuring the cognition of problems and shaping the evaluation of solutions. And with this epistemic focus, we begin to see that missing from prior explanations of algorithmic expansion is attention to the conceptualizations and tool, empirical tools that undergird or support policy debates, thereby making possible the identification of scoring as legitimately racially inclusionary. So the rest of my talk then is gonna be focused on offering a sort of formation story here of how, one, a particular conceptual model of exclusion, namely discrimination, emerged in consumer credit markets. Two, how this model was institutionalized via regulators’ efforts to empirically identify discrimination.

Three, how that pairing of conceptualizations and empirical identification work together to produce an understanding of credit scoring as race neutral, and thereby not, and therefore not racially exclusionary. And finally, how all of that came together to have the consequence of underwriting contemporary efforts at expanding data and credit scores under this sort of banner of racial inclusion. So this is gonna be where we’re headed here.

And given this focus, my empirical approach is not gonna be geared towards deconstructing what could have been over the history of consumer credit scoring, but towards delineating the properties of the path that emerged, and explaining how this path has enabled the contemporary orientation towards expansion. And to do this, I constructed a corpus of documents composed of congressional hearings, regulatory documents, court cases, and assorted reports, and news articles from 1968 to 2019.

And I focus on this period considering 2000 or 1968 was the first year in which Congress passed the first big consumer credit legislation. And 2019 aligns with the creation of a Consumer Financial Protection Bureau policy that was designed to nurture innovations like alternative data and consumer credit scoring. Though, as you’ll see, my analysis here sort of bleeds into a few years before 1968 and a few years after 2019 because of really important context. The process of gathering this data involved focusing on identifying key congressional hearings as the sort of main source of information via targeted queries in the ProQuest Congressional Database, which includes, a comprehensive collection of legislative histories, hearing transcripts, and other documents related to public congressional interactions.

I queried this database for keywords that emphasized credit scores, discrimination, and any reference to consumer credit legislation. From these searches, I produced basically a list of 14 hearings from 1968 to 2019 that were the most relevant for my questions.

By most relevant for my questions, I mean lots of conversations around credit scores has happened in conversations about social media, and a lot of those social media hearings were not necessarily super related to the questions that I’m concerned about. I then validated and supplemented these 14 hearings with information from the prior historiography of consumer credit and credit scores, and then added news articles, court cases, and other documents that sort of provide, again, more context and background to what I was observing. The analysis of over this, these 4,000 pages of documents was composed of four steps, close readings and coatings in NVivo, which were sort of there to sort of help me get organized and track discussions.

Notes on the hearings helping me identify important moments. A third step, which was notes on my notes which involved my own commentary and just trying to sort of narrow down the focus of what’s really important for across these hearings, and then finally, a set of analytic bullets and memos, which were these kind of meta documents that produced from the initial three steps and were then enhanced with relevant insights from secondary sources where appropriate. And with this analytic approach, I proceeded basically chronologically through my corpus of documents moving between notes, memos, and prior literature, and moving through the data in this way proved useful as what emerged were three key dynamics.

The first details the emergence of a particular conceptualization of discrimination in credit markets. The second demonstrates how that conceptualization was institutionalized via particular empirical methods used by regulators and courts to identify discrimination, and the final section illustrates how this underlying infrastructure had the consequence of enabling a wide cadre of actors to legitimately identify credit scores as non-exclusionary, thereby underwriting the expansion of data collection in credit markets. And I refer to these sections using the shorthand of emergence institutionalization, and consequence, and because these dynamics are broadly delineated in time, I used these dynamics to sort of structure the results, so I’m gonna be marching us through this timeline shown here. Now, after all of that wind up let’s get into this.

So this period of emergence highlights the importance of establishing a particular conceptual definition of discrimination as there was no legislative model of discrimination in the 1960s. Indeed, there was not even a sense that racial discrimination existed in consumer credit markets. When Congress passed the Consumer Credit Protection Act in 1968, they created the National Commission on Consumer Finance, the NCCF, to study issues in consumer credit markets for more or less the first time at the congressional level. The NCCF issued their report in 1972, and on the question of racial discrimination, they concluded that they did not find sufficient evidence to prove the hypothesis that there is racial discrimination in the granting of consumer credit.

Right now, this conclusion is somewhat odd when we think about 1970s US. Indeed, it flew in the face of the Kerner Commission Report, which five years prior to the release of the NCCF report, had noted that discriminatory credit practices were among the key reasons animating the racial unrest in the 1960s. All right, so we have this sort of weird dynamic here, but ultimately because of this sense that there was not racial discrimination in credit markets, the model of discrimination that emerged and the consequences of what happened was not so much about race, but around issues of gender and the actions of gender activists in the 1970s.

So groups like the National Organization of Women, the Center for Women’s Policy Studies, amongst many others, brought forth such compelling evidence of gender of gender discrimination that from 1972 to 1974, over three dozen bills were introduced discrimination on the basis of gender and marital status. Only one bill over this period even referenced race.

This evidence was made particularly clear during hearings in June of 1974, where during these hearings, women’s organizations revealed how pervasive the overt consideration of gender and marital status was in credit markets. An example of which is shown here, which comes from a 1960s credit application that I actually gifted to my dissertation advisor on my graduation. This application asks for a lot of information.

But in the yellow box, if married, all information must apply to the husband. All right, so clear examples like this, very powerful and compelling for legislators in 1974. In the face of overwhelming testimony, the issue was not if equal credit would be passed, but whether the law could be written without destroying a creditor’s right to make business judgments on the basis of an individual applicant’s creditworthiness.

And here is where efforts by women’s organizations faced significant friction. As Greta Krippner notes, “When the position of any individual woman was considered, individual creditworthiness seemed a reasonable standard.

But when the structural features of markets that produced group disadvantage were taken into account, individual merit stood against the murky shadows of history.” And in congressional debates, feminist organizations chose to sidestep that larger issue of structural inequality to focus on a reduced form of discrimination.

In an article submitted for the Congressional Record, co-director here of the Center for Women and Policy Studies Margaret Gates indicated that “Women as a class are economically disadvantaged. And this unfortunate fact explains why women, why fewer women than men obtain loans and credit. But it does not justify the denial of credit to a woman who is, by all objective criteria, as qualified as a man.” During these 1974 hearings, activists took on a logic that all else equal, women should not be denied because they are women.

This view focused on rooting out overt considerations of gender or marital status in decisions. Limiting the definition of discrimination in this way foreclosed upon an understanding of how structural forces constituted women’s position in society as disadvantaged. Indeed, concern for these structural manifestations of inequality was all but absent during these 1974 hearings. Indeed, there were many of these groups organized against such a reckoning.

So when the ink dried on the Equal Credit Opportunity Act in October of 1974, Equal Credit Opportunity Act, you will hear me refer to that as ECOA, Congress banned the use of gender and marital status as over considerations in credit decisions. Race and other protected categories would eventually be added to ECOA in 1976, but this original 1974 legislation christened a particular model that identified discrimination as simply the use of the categories of gender or race.

And this would have very important consequences. Legislation and subsequent regulatory guidance produced by the Federal Reserve in its Regulation B, which is the regulation governing credit scoring practices, would say nothing about how group-based differences were expressed through non-group characteristics, basically what is shown in the bottom.

This would effectively give lenders carte blanche to include anything they wanted in their credit scoring and underwriting except for the explicitly banned criterias. Indeed, that is explicitly written in Regulation B. However, shortly after the passage of the Equal Credit Opportunity Act to include race in 1976, Congress began wanting to go further, began wanting to exclude the use of information, began to, wanting to exclude the use of more pieces of information from credit decisions. But they would run into challenges as this, the model of discrimination passed in 1970, ’74 created important barriers. Specifically, and this sort of key moment here, is that Congress would grapple directly with the question of whether the correlated nature of a whole host of characteristics and race would be deemed illegal or not when they met in 1979 to discuss this Senate Bill 15, which sought to eliminate the use of ZIP codes in consumer credit scores.

So by 1979, the use of credit, use of ZIP codes in credit decisions and credit scores had become very common. Everybody was doing it. But this raised the specter of racial redlining, and indeed the title of the hearing, Credit Card Redlining, sort of made that intentionality explicit here. So the Senate heard lots of testimony like that shown here of Columbia Business School professor Noel Capon that argued, “The use of ZIP codes effectively discriminates on the basis of race.”

However, the further efforts to limit the information in credit decisions got from the actual categories of race or gender, the more challenging it became to see those features as linked to inequalities according to the ECOA model of discrimination. Member of the Board of Governors of the Federal Reserve, Nancy Teeters, drove this point home to the Senate.

She indicated that, quote, “It may be the case that ZIP codes should be prohibited. Congress should take note of the fact, though, that other characteristics can be called into question.” Teeter contended that if you eliminate ZIP codes because of its potential to racially discriminate, then you must consider homeownership or, quote, “Almost anything related to financial status.” The implications of this proved too much.

Senator Paul Tsongas expressed unease and discomfort with continuing to allow the use of ZIP codes in credits decisions, but noted, quote, “I don’t think this is anything we should legislate.” So Senate Bill 15 failed to move out of committee. From 1979 to 1989, nearly a dozen bills in the House and Senate that sought to eliminate the use of geography, occupation, work title, and course of study because of their potential to racially discriminate met a similar fate. And the failures of these bills is important because it settled a way of conceptualizing discrimination that emerged with the passage of the original Equal Credit Opportunity Act.

This model functioned according to a narrow overt usage model bracketing any inequities that manifested through other, through variables other than those protected categories themselves. So, with congressional desire to legislate what goes into credit scores effectively evaporating in the wake of the failure of Senate Bill 15, any future constraints on credit scoring would have to proceed through the courts and via the efforts of regulators that were tasked with moving from this conceptualization stage to the stage of empirically identifying discrimination. And over this period, we would see regulators and courts do so in particular ways, which is nicely reflected in the case of Cherry versus Amoco Oil. So in 1978 Mrs. Claire Cherry, a white woman living in a predominantly Black area of Atlanta, applied and was declined for a credit card with Amoco Oil.

Amoco sent Mrs. Cherry this letter here that stated they are declining her because of their experience in her geographical area. Claire Cherry then was like, “All right, well, that’s kind of racist.”

(audience laughing)

She sued them based on the notion that using geography had a racially disparate impact or had the effect of discriminating on the basis of race. And while Congress chose to not legislate this as specifically illegal in 1979, the courts had some authority here. Her lawyer enlisted the help of a sociologist, go sociologists at Georgia State University, Robert Snow.

And Robert Snow put together this analysis here. Snow put together an analysis that showed a reliably significant relationship between a significant relationship where approval rates by Amoco Oil decreased significantly as the percent of Black in Atlanta zip codes increased. Although Snow presented this robust evidence, the court agreed with Amoco Oil in 1980 with their decision that said this system does not illegally discriminate on the basis of race.

So this case reveals a very important key challenge of empirically identifying discrimination in scoring systems. Even the most sophisticated analyses showing stark racial disparities may fail to become institutionally legible.

And this challenge created significant frustrations. By the late 1980s there were over a half dozen or so regulators that had some responsibility covering fair lending. And in 1989, each of these regulators presented a slew of statistical evidence on the state of fair lending to Congress.

Across a two-and-a-half-hour meeting or hearing, it’s not really a meeting, two-and-a-half-hour hearing, each regulator introduced evidence of stark racial disparity, with Black loan rejection rates being roughly twice as high as white applicants. Yet no regulator concluded there was any racial discrimination. The Office of the Comptroller of the Currency, in particular, noted that they had received 37,000 complaints from 1987 to the middle of 1989. And after investigating each one, quote, “No violations were cited.”

Moreover, they did nearly 3,500 compliance examinations across 239 banks, also finding no indication of illegal discrimination. Senator Alan Dixon was in disbelief when he noted, “I think it to be somewhat incredible that the substance of the testimony is mostly that you haven’t found any violations when the evidence is pretty clear. I find it pretty close to remarkable that we never find any violations.” And in the same hearing community organizations also presented information of being disheartened and sort of discomforted by the fact that no one in Congress was, or no one, no regulator was able to articulate any evidence of discrimination.

And this kind of exchange was common through this sort of late 20th century period and conveys a clear dissonance between the enforcement and the realities of people navigating credit markets on the ground. This disconnect emerges because of the tools regulators used to identify racial discrimination. This was put very clearly by the New York Federal Reserve vice president, who explained to Congress at a field hearing in Brooklyn, that regulators focused on, regulators focused on using statistical analyses to identify statistically significant race variables in regression frameworks. If present, then they added controls for loan size, income, credit score.

If race was still significant at that point, then they added even more variables. If it was still significant, they continued to add more variables to try and explain these racialized differences in lending that regulators were observing. In effect, what counts as racial discrimination is identified through the existence of a statistically significant race variable in regression.

As regulators tried to bridge the racial differences in White loan rejection rates, here shown with an example data on the left and Black rejection rates on the right, their approach treated differences in income and credit scores as functionally unrelated to race, and most importantly, unrelated to the enterprise of rooting out racial discrimination in credit markets. As a result, stark disparities became reframed as legitimate outcomes rather than raising regulatory alarms about violations of fair lending law.

This regulatory approach was reinforced in this, in this period, by the way also that anti-discrimination legal doctrine of disparate impact would involve and be interpreted by courts. So, as commonly understood, disparate impact or the effects test is intended to pick up on the ways that discrimination can operate through non-group characteristics, right?

The disparate impact was the grounds and basis along with Mis, along which Mrs. Claire Cherry sued Amoco Oil. And in that case, we saw that disparate impact was not an effective tool. And this is in part by how this legal doctrine has been interpreted. So, disparate impact as a legal concept in the US originates from this case in 1970 when the Suprem, this 1971 Supreme Court case, Griggs versus Duke Power.

And it has evolved to be a three-step burden-shifting framework that has been applied to the enforcement of Equal Credit Legislation. In Step 1, a plaintiff must pres, make a claim, typically using statistical evidence that some racial disparity exists because of a lender’s policy.

Step 2 shifts the burden to the lender to indicate that disparity is linked to a legitimate business necessity. If this is possible, then it shifts back to, the burden shifts back to the plaintiff to identify a less discriminatory alternative that will be just effective as the lender’s initial policy. In the context of credit and credit scoring, getting past the second step is functionally impossible, right? Attempts to meet hurdles of this burden-shifting framework were dead on arrival.

As countless industry representatives have explained to Congress, credit scores are created by evaluating which variables are associated with loan repayment. Only those variables that are demonstrably tied to debt repayment are included in the creation of credit scores and credit decisions. This provides a baseline business necessity defense for any claim of disparate impact. Indeed, the Office of the Comptroller of the Currency issued a very clear bulletin in 1997.

They noted that the OCC, the Office of the Comptroller of the Currency, will conclude a variable is justified by business necessity and does not warrant further scrutiny if the variable is statistically related to loan performance and has an understandable relationship to an individual applicant’s creditworthiness. As a result of this, all litigation challenging credit scoring on disparate impact grounds have failed. In 2004, in a 2004 case, courts summarily rejected claims that the use of credit histories and credit scores create a disparate impact on Black consumers. The court notes that ECOA regulations explicitly permit lenders to use, quote, “race-blind factors” such as credit histories and income when determining whether to extend credit.

So here, we see the melding of this particular model of discrimination in the Equal Credit Opportunity Act, the tools regulators use to identify creditworthiness or to identify racial exclusion, and the legal arguments found to be compelling in courtrooms effectively work to dispatch critiques of racial exclusion. Even ZIP codes, which are the go-to example any researcher uses to argue and demonstrate racially disparate impact has failed to be seen as discriminatory in the history of consumer credit scoring and is still today not technically illegal, right? It is precisely because of the infrastructure of Equal Credit Legislation and its enforcement that legitimates an identification of scoring, in this case, as not racially exclusionary according to this sort of model of discrimination.

This ultimately establishes the bounds of debate and the epistemological terrain upon which discussions of credit scoring and the expansion of data used in credit scores would basically operate on in the 21st century. So as we turn to the 21st century, there has been this wave of new technology companies seeking to expand the types of data used in consumer credit scores. And unlike prior period, when Congress begins to discuss the growing use of alternative data, efforts to reduce data fed into these algorithms is basically all but absent. Because of the epistemological bounds set by the original Equal Credit Opportunity Act and its enforcement, a growing group of actors began to see credit scoring as inclusionary by its very nature.

The issue that began to take over congressional and regulatory discussions was not the inequality of credit scoring itself, but the racially uneven problem of not having a credit score, that is, of being unscored. After George Bush signs the Fair and Accurate Credit Transactions Act in 2003, Congress learned that many Americans, many consumers may not have sufficient information in their credit report to create a credit score. Estimates in 2005 put this number around 35 million people. Besides being a large population, credit invisibility or in this problem of, quote unquote, “unscored populations” disproportionately affects some groups more than others.

For example, Black, Hispanic, and populations living in lower-income neighborhoods are all much more likely to not have a credit score. So when folks from the credit scoring industry sort of discussed this problem of the unscored, they noted that, well, if we just expand the data that goes into scores, it would effectively deal with these prob– this problem of the so-called unscored. And this led many legislators to begin seeing that expanded data would allow millions to climb out of the shadows to build a credit history, which would also then be a boon for racial inequity. And this move towards more data reflected, again, the sort of growing comfort that credit scoring itself did not reinforce problematic inequality.

Pursuant to the Fair and Accurate Credit Transactions Act, the Federal Reserve 2007 completed and delivered this report to Congress. And Sandra Braunstein of the Fed summarized the key finding’s in front of the House and brought home that there is no disparate impact or discriminatory impact of credit scores across racial groups.

A Fed report in 2012 would basically make the same point. Even when members of Congress push back on credit scores by noting, “Yeah, but they’re distributed unequally across racial groups,” credit scoring executives are able to effectively and easily dispatch those claims.

Stuart Pratt, the president and CEO of a consumer data trade association noted, “A credit score is blind. It does not know my race.

All it does is look at empirical data. Scores, because they are blind to these triggers, these ECOA triggers, they have removed the risks of those triggers.”

So not only are industry representatives able to push back, but they’re able to do so within the guardrails in language explicitly established by the Equal Credit Opportunity Act. Or as Lisa Nelson, Vice President of Operations at FICO put it, “We sit perfectly inside the regulatory framework that exists today.” So Congress, members of Congress are getting increasingly on board with this. While that’s happening, consumer advocacy groups representing those on the fringes mobilize to bring forth some concerns.

At every step of discussions around alternative data, the National Consumer Law Center, especially with the efforts of Chi Chi Wu, expressed reticence about the use of new data. While they generally found some potential value in giving and providing some consumer choice about including certain kinds of data, they expressed particular concern around growing efforts to include information about educational background and occupation beginning, because, again, of their potential to produce a racially disparate impact. And the House met in 2019, this is the hearing that I sort of began earlier with, to discuss some of these more disconcerting kinds of alternative data. At the center of the hearing was Upstart, a financial technology firm that leveraged new kinds of information.

And during the hearing, many expressed concern about Upstart’s credit scoring tool because it incorporated data about individuals’ employment history and educational background, the exact same characteristics members of Congress sought to ban in the 1980s. CEO of Upstart, Dave Girouard began his testimony.

He said, “In the early days of Upstart, we conducted a retroactive study with a large credit bureau, and we uncovered a jarring pair of statistics.” “Just 45% of Americans have access to bank-quality credit, yet 83% of Americans have never actually defaulted on a loan.

This is not what we would call fair lending.” He continued and said, “We decided to use modern technology and data science to find more ways to prove that consumers are indeed creditworthy.” And this language is very convincing to many members of the House, but others that were present to give testimony raised significant concerns.

They said these new types of data may actually just be proxies for protected traits and result in decisions that have a disparate impact. Relatedly Chi Chi Wu from the National Consumer Law Center noted, “We know there are tremendous racial disparities with respect to traditional credit scores, and alternative financial data is also likely to have these same racial disparities.”

Shortly after this congressional hearing, shortly after Upstart discusses its credit scoring system in front of Congress, the Student Borrower Protection Center issued a report titled ‘Educational Redlining,’ demonstrating how Upstart’s use of educational data could exacerbate racial inequality. And this report sparked then the formation of an agreement between Upstart, the Student Borrower Protection Center, and the NAACP Legal Defense Fund to employ an independent third party to evaluate Upstart’s lending model for racially unequal outcomes. Like regulators attempting to empirically identify racial discrimination, the report focused on identifying statistically significant race variables in a regression framework. And when the independent third party released their report in 2021, they, like the National Commission on Consumer Finance in 1972, and regulators all throughout the late 20th century, found no evidence of any fair lending violations.

Insofar as there were meaningful disparities between Black applicants and non-Hispanic White applicants. These disparities were, quote, “Measured on an unadjusted basis without attempting to control for legitimate creditworthiness criteria and therefore does not, standing alone, demonstrate a fair lending violation.”

So with the combination of conceptual tools embedded in the Equal Credit Opportunity Act and the evolution of enforcement through statistical analyses, efforts to contest scoring as racially unjust by the organizations like the NAACP and the National Consumer Law Center smash into this epistemological wall. This clears the way for algorithmic expansion as those with an intuition that credit scores are racialized in objectionable ways are unable to make those critiques legible to political and legal institutions. Without an inability to cognize deeper inequality in credit scores, the orientation towards financial inclusion has effectively become clear. As lawyer and former director for credit practices at the Federal Trade Commission, Anne Fortney, put it to the House, “The obvious solution is to increase the amount of information available to credit score developers.”

So to summarize the argument here, contemporary efforts to expand data and credit scores were seeded with the passage of the 1974 Equal Credit Opportunity Act. ECOA melded a particular conceptualization of discrimination with regulatory enforcement that relied heavily on statistically significant race variables in regression.

Such approaches work to limit the scope of exclusion legible to regulators. Despite stark disparities across racial groups in credit outcomes, the infrastructure permeating regulation and law fostered an understanding of those disparities as not related to race. Consequently, notions of exclusion and inclusion at the turn of the 21st century were anchored within this epistemological infrastructure.

Critics deploying more expansive frameworks failed to generate resonant critiques. Creators of credit scores could easily just fall back on the style of reasoning embedded in Equal Credit legislation and enforcement to convincingly demonstrate an absence of any discrimination traceable to race.

This evidence of absence then legitimated credit scoring as not wrongfully exclusionary according to the law. The only racial exclusion that was then legible to policymakers became the existence of millions of Black and racially marginalized consumers without a credit score. As a result, expanding the data in credit scores became a straightforward, even morally necessary solution, that given the racial epistemology and anti-discrimination enforcement could be legitimately identified as racially inclusionary.

So what? So I think this account is useful as it helps reconcile gaps in prior explanations. While prior research tends to emphasize algorithmic expansion in organizational, technical, or even political terms, I demonstrate how these accounts underappreciate algorithms and their expansion as problems of knowledge, where specific epistemological infrastructures shape both the cognition of problems and define the landscape of solutions.

And I think this is useful in explaining the contemporary expansion in credit markets, as it highlights how a reason why legislators saw expanding data as data in credit scores today as a solution to racial exclusion, unlike legislators in the mid-20th century, is because legislators today literally do not have the tools available to them in legislation and regulation that equip them to see this as a problem. And a key contribution, and I think theoretical payoff to this insight is in beginning to flip the debate at the nexus of economic sociology and the sociology of race on its head a little bit.

So much of this literature is oriented around evaluating whether the use of some algorithm or piece of data in some context increases or decreases racial inequality. And I think what I’ve helped to clarify in this analysis is that maybe the question we should be asking, indeed the more challenging and politically consequential problem question, I think, may have to do with trying to figure out which ways of coming to know the answer to those questions are deemed legitimate. And I could continue going on and on about the relevant contributions to the historiography of consumer credit, what this means for how social scientists should study race and racism, or how we should think about the possibility of ameliorating racial inequality in this sort of big data moment.

But I really want to get into the conversation with you, right? So this is the first time that I’ve actually presented this work. And I have no idea what questions people have.

And I’m eager to hear which parts of this you think are dope. And I’m also most eager to hear which parts of this you think suck or need clarification. So with that, I will conclude. Thank you.

(audience applauding)

(upbeat instrumental music)