SEX, CAUSATION, AND ALGORITHMS: HOW EQUAL PROTECTION PROHIBITS COMPOUNDING PRIOR INJUSTICE.

• Author: Hellman, Deborah

TABLE OF CONTENTS INTRODUCTION I. LOOMIS TEES UP THE QUESTION II. EQUAL PROTECTION AND SEX-BASED CLASSIFICATIONS A. The Fit Framework B. Prohibition on "Stereotyping" C. The Real Differences Exception 1. Real Differences as Biological Differences 2. The Rationale for Biological Real Differences 3. Internal Confusion III. REINVENTING REAL DIFFERENCES A. Correlation and Causation B. Real as Morally Significant C. The Anti-Compounding Injustice Principle in Constitutional Law IV. SEX-BASED CLASSIFICATIONS IN RISK ASSESSMENT V. CONCLUDING THOUGHTS: SHOULD THE PAST BEPROLOGUE? INTRODUCTION

In a case that garnered significant attention, the Supreme Court of Wisconsin considered whether an algorithm deployed to predict recidivism may use the sex of a criminal defendant to inform the sentencing judge. (1) The Wisconsin court concluded that the use of sex within the algorithm was permissible because it did not violate the U.S. Constitution's guarantee of due process. But the court left open the question whether it might, nonetheless, violate equal protection. (2) Although the Supreme Court denied cert in the defendant's attempted appeal, (3) the growing use of sex in states' risk assessment tools places an increasing onus on the Court to weigh in on the issue. (4) The resolution of this issue matters legally and practically. Use of sex has the potential to help women significantly. Women commit fewer crimes overall, and especially fewer violent crimes. (5) For this reason, sex is used within predictive algorithms because it strongly correlates with the likelihood of committing crime, particularly violent crime. (6) If women are not grouped together with more violent men, women may be sentenced more lightly and released on parole more frequently. (7)

May states use sex-based classifications to determine how men and women are treated in this context? Perhaps surprisingly, U.S. constitutional law is not clear on this point. Traditionally, sex-based classifications are subject to "intermediate scrutiny." Whatever canonical phrases one uses to describe this level of review--"exceedingly persuasive justification," (8) "substantially related to the achievement of an important governmental objective" (9)--intermediate scrutiny is intermediate, neither as demanding as "strict scrutiny" nor as permissive as "rational basis review." This placement suggests that the use of sex-based classifications is sometimes constitutional. Two developments (one older and one more recent) make that inexact precept even more difficult to utilize. First, in the 1996 case United States v. Virginia, (10) the Supreme Court invalidated the male-only admissions policy at the Virginia Military Institute (VMI) in a manner that some judged to be a version of strict, rather than intermediate, scrutiny. (11) Second, there has long been an exception to sex-based equal protection doctrine. When sex-based differential treatment is grounded in so-called "real differences," it is permissible. (12) Yet recent cases demonstrate that this exception is growing increasingly amorphous. (13) Together these two developments suggest that the constitutional law governing sex-based classifications is ripe for reexamination.

At the same time, the use of "big data" and machine learning to develop algorithms for a wide range of contexts including not only criminal justice but also employment, (14) lending, (15) and targeted advertising (16) is becoming more common. May these algorithms use sex? And if so, how? Of course, only some of these contexts involve state actors and so only some will be governed by the constitutional law of sex discrimination. However, the interpretation of statutes governing the use of sex-based classification often is informed by constitutional law. (17)

Now is thus an apt time to revisit the question when, and if, sex-based classifications should be constitutionally permitted. The issue is timely as sex-based classifications are used in the algorithms relied on by many states to predict recidivism, (18) and it seems unlikely that the Supreme Court can avoid addressing the issue for too long.

Current equal protection doctrine does not provide a clear answer. It does not because sex-based equal protection doctrine is both ill-defined and hazy and, at the same time, shifting and unsettled. This Article proposes a new norm around which the doctrine can cohere, one which will provide guidance in answering the many questions--like the one raised in Loomis--that the use of algorithmic decision-making will give rise to.

Sex-based classifications are sometimes permissible and sometimes impermissible under current doctrine. The doctrine utilizes two conceptual schemas to identify impermissible uses of sex-based classifications: the fit framework and stereotyping. When sex is too loose a proxy for some target trait or when it relies on a stereotype, it is impermissible. By contrast, when the sex-based classification is grounded in a "real difference" between men and women, it is permissible. While seemingly promising, these schemas are unhelpful because the degree of fit is unspecified, the concept of stereotyping is undefined and the designation of a "real difference" is normatively unmoored.

Yet, imminent within the doctrine are the seeds of an alternative approach. The real differences doctrine correctly focuses on whether sex is causally related to the trait for which it is a proxy, rather than whether it is merely correlated with the target trait. What is missing, however, is a focus on what causal mechanisms are morally and constitutionally problematic. The heart of the normative contribution of this Article picks up where the doctrine leaves off. I argue that when a history of sex-based injustice provides a plausible causal explanation for the observed correlations between sex and other traits, then sex-based classifications are presumptively problematic. For example, in 1973 when the canonical sex- discrimination case Frontiero v. Richardson (19) was decided, sex was an accurate proxy for having a dependent spouse. In deciding whether the armed services could use sex as a proxy for having a dependent spouse in the context of providing benefits, the Supreme Court focused on how good a proxy sex was for dependency. (20) Instead, the Court should have focused on why significantly more men had dependent spouses than did women. Prior sex-based injustice that kept women from employment opportunities likely caused the correlation. As sex was an accurate proxy for dependency because of prior sex-based injustice, its use should be presumptively impermissible.

By contrast, when no such plausible causal hypothesis is present, then the sex-based classification may be permissible. For example, different fitness requirements for women than for men are generally seen as permissible. (21) While sex is an accurate proxy for the ability to do significant numbers of push-ups or chin-ups, just as sex was an accurate proxy for dependency, the correlation between sex and chin-up ability is not likely caused by sex-based injustice. As a result, the use of an explicit sex-based classification in differential training requirements may be permissible. I say "may" here rather than "is" because there is more than one reason that sex-based classifications are morally and constitutionally problematic. (22) If another reason is relevant, the sex-based classification may still be impermissible for that reason.

Current equal protection doctrine is unclear and under-theorized regarding whether and when sex-based classifications are permissible. This Article offers a more coherent and normatively appealing account, which, like current doctrine, permits some sex-based classifications and prohibits others. Flowever, unlike current doctrine, this account draws the line between permissible and impermissible uses of sex-based classifications in a manner that is reasoned and defensible.

The constitutional principle I offer rests on the moral claim that governmental actors have obligations to avoid compounding prior injustice, including injustice for which they are not responsible. (23) In my view, this Anti-Compounding Injustice (ACI) principle is already inchoate in equal protection doctrine. It provides a normatively appealing account of canonical sex-based equal protection cases, like Frontiero. In addition, the ACI principle offers a way to sympathetically reconstruct the "real differences" exception. Once we recognize the ACI principle at work, it is easy to see how it can be extended to help determine whether sex can be used in risk assessment tools and other predictive algorithms.

The advent of big data together with machine learning is likely to substantially increase the influence of the past on the future. (24) Data-driven analysis is inherently based on the past. What is data, after all, but information about past events? Yet, equal protection law has often functioned, at least in part, to disrupt the grip of the past on the future. (25) If technological advances create a situation in which the past will control the future to a significantly greater degree, it will become necessary to revisit the current legal settlements. This Article presents a reconstruction of current doctrine that provides the tools needed to handle these new developments.

This Article proceeds as follows. Part I describes State v. Loomis, focusing particularly on its treatment of sex classifications in the context of recidivism risk predictions. Part II turns to current equal protection law and its treatment of sex-based classifications. In this part, I describe the organizing principles that define when sex-based classifications are presumptively impermissible and the central exception outlined in the jurisprudence that delineates when sex-based classifications are permissible. The goal of this part is to show that the doctrine is confused and normatively ungrounded. Part III offers an...