Lily Hu

In this article, I set forth an inquiry into discrimination on the basis of race via two broad questions: What is a consideration of race a consideration of? And what is it for action to be on the basis of that consideration? Beginning with the matter of what constitutes race-based action brings out two conceptual hinges in analyses of racial discrimination, one in race and one in reason, that comprise what I call the race as a reason framework. Not only are there different notions of practical reason that may be implicated in the stricture not to act on the basis of race (as a reason), there are also different conceptions of race that are relevant to these different notions of reasons for acting. Drawing out these two sites of ambiguity generates an illuminating taxonomy of acting on the basis of race that has significant theoretical and normative upshots for discrimination theory.

This article challenges the widespread view that causal relations are wholly non-normative natural relations. In the case of salient social categories such as race and sex, I argue for a dependency of the causal order of things on the moral and political. At the center of my argument is a puzzle about how audit studies, social scientific experiments that approximate the idealized controlled experiment, work to probe the causal significance of race and sex. Essential to these studies is the distinction between differences that constitute rather than confound an effect of their variable of interest—a distinction, which, I claim, embodies judgments about the sorts of differences that constitute the wrongful causal influence of these categories. If our best scientific practices latch onto genuine features of causation, then normative theorizing’s place in scientific causal theorizing entails that normative facts figure in the causal structure.

This article explores the issue of normative distinctiveness in machine learning decision systems alongside Solon Barocas, Moritz Hardt, and Arvind Narayanan’s landmark book Fairness and Machine Learning. What, if anything, is different this time, with the rise of machine learning-based aids to bureaucratic decision-making? I show how a focus on normative distinctiveness can obscure from view a much more significant upshot of machine learning: that the mere existence of feasible alternatives presses new justificatory demands not just on the design of new technical systems but on the prevailing human-centered decision regime. I argue that depoliticizing conventional bureaucratic structures of decision-making leads to a missed opportunity for a broader normative reevaluation of what we owe to each other in a world of expanded practical possibility

Most leading philosophical accounts of discrimination theorize discrimination as a causal notion: roughly, an action discriminates on the basis of X (e.g., race, sex) if X causes (in the right way) the adverse outcome. This article explores the prospects for this causal account. Focusing my attention on the case of sex discrimination, I argue that struggles to settle the causal question regress to question-begging. This argument calls into question not just whether the causal account can handle sophisticated cases, which have recently been conceived of (by some) as discriminatory on the basis of sex, but even classic cases of sex discrimination. The negative argument in turn brings to light an alternative normative-causal approach to discrimination, which I sketch and suggest is promising indeed.

Discussions of statistical criteria for fairness commonly convey the normative significance of calibration within groups by invoking what risk scores “mean.” On the Same Meaning picture, group-calibrated scores “mean the same thing” (on average) across individuals from different groups and accordingly, guard against disparate treatment of individuals based on group membership. My contention is that calibration guarantees no such thing. Since concrete actual people belong to many groups, calibration cannot ensure the kind of consistent score interpretation that the Same Meaning picture implies matters for fairness, unless calibration is met within every group to which an individual belongs. Alas only perfect predictors may meet this bar. The Same Meaning picture thus commits a reference class fallacy by inferring from calibration within some group to the “meaning” or evidential value of an individual’s score, because they are a member of that group. The reference class answer it presumes does not only lack justification; it is very likely wrong. I then show that the reference class problem besets not just calibration but other group statistical criteria that claim a close connection to fairness. Reflecting on the origins of this oversight opens a wider lens onto the predominant methodology in algorithmic fairness based on stylized cases.

Theories of causation within the causal modeling framework are among the most promising and well‐developed approaches to analyzing actual causation today. But since the advent of model‐based theories of causation, authors have struggled with the fact that virtually all such theories issue different causal conclusions when applied to different causal models (of the same case). In recent years, these concerns about model relativity have sharpened into a search for a general theory of model aptness, which lays out principled grounds on which models may be ruled out as inapt for causal analysis. This article presents a new way of understanding what drives model relativity that shows why it is such an enduring problem. I show why extant proposals fail and furthermore, why they are not of the right form to be a solution. Along the way, I draw broader lessons about what an adequate answer to these problems must be like. I then take a step toward such an account by outlining a substantive constraint on when variables and paths in a causal model must not be drawn and when they may.

Machine learning algorithms bring out an under-appreciated puzzle of discrimination, namely figuring out when a decision made on the basis of a factor correlated with race is a decision made on the basis of race. I argue that prevailing approaches, which are based on identifying and then distinguishing among causal effects of race, in their metaphysical timidity, fail to get off the ground. I suggest, instead, that adopting a constructivist theory of race answers this puzzle in a principled manner. On what I call a “thick constructivist” account of race, to be raced is to be socially positioned in the way indicated by a certain set of statistical regularities on the basis of particular phenotypical traits. A thick constructivist sees that acting on the basis of correlations that constitute race qua social position just is acting on the basis of race, because races just are social positions that subject their member individuals to a particular matrix of social relations that define the raced position. This conclusion has considerable ramifications for our understanding of discrimination, algorithms and beyond.

The particular design of any technology may have profound social implications. Computing technologies are deeply intermeshed with the activities of daily life, playing an ever more central role in how we work, learn, communicate, socialize, and participate in government. Despite the many ways they have improved life, they cannot be regarded as unambiguously beneficial or even value-neutral. Recent experience shows they can lead to unintended but harmful consequences. Some technologies are thought to threaten democracy through the spread of propaganda on online social networks, or to threaten privacy through the aggregation of datasets that include increasingly personal information, or to threaten justice when machine learning is used in such high-stakes, decision-making contexts as loan application reviews, employment procedures, or parole hearings. It is insufficient to ethically assess technology after it has produced negative social impacts, as has happened, for example, with facial recognition software that discriminates against people of color and with self-driving cars that are unable to cope with pedestrians who jay-walk. Developers of new technologies should aim to identify potential harmful consequences early in the design process and take steps to eliminate or mitigate them. This task is not easy. Designers will often have to negotiate among competing values—for instance, between efficiency and accessibility for a diverse user population, or between maximizing benefits and avoiding harm. There is no simple recipe for identifying and solving ethical problems. Computer science education can help meet these challenges by making ethical reasoning about computing technologies a central element in the curriculum. Students can learn to think not only about what technology they could create, but also whether they should create that technology. Learning to reason this way requires courses unlike those currently standard in computer science curricula. A range of university courses on topics in areas of computing, ethics, society and public policy are emerging to meet this need. Some cover computer science broadly, while others focus on specific problems like privacy and security; typically, these classes exist as stand-alone courses in the computer science curriculum. Others have integrated ethics into the teaching of introductory courses on programming, artificial intelligence, and human-computer interaction.