A peculiar way of making decisions is characteristic of modern society. Institutions of all kinds, from firms to governments, represent populations as data tables. Rows reference individuals. Columns contain measurements about them. Statistical machinery applied to these tables empowers their owners to mine patterns that fit the aggregate.

Then comes a leap of faith. We have to imagine that unknown outcomes, future or unobserved, in the life trajectory of an individual follow the patterns they have found. We must accept decisions made as if all individuals were going to follow the rule of the aggregate. We must pretend to ourselves that to look into the future is to look into the past. It’s a leap of faith that has been the basis of consequential decisions for centuries. Fueled by early successes in insurance pricing and financial risk assessment, statistical decision making of this kind has found its way into nearly all aspects of our lives. What accelerated its adoption in recent years has been the explosive growth of machine learning, often under the name of artificial intelligence.

Machine learning shares long established decision-theoretic foundations with large parts of statistics, economics, and computer science. What machine learning adds is a rapidly growing repertoire of heuristics that find decision rules from sufficiently large datasets. These techniques for fitting huge statistical models on large datasets have led to several impressive technological achievements. Image classification, speech recognition, and natural language processing have all made leaps forward. Although these advances often don’t directly relate to specific decision making settings, they shape narratives about the new capabilities of machine learning.

As useful as machine learning is for some positive applications, it is also used to great effect for tracking, surveillance, and warfare. Commercially its most successful use cases to date are targeted advertising and digital content recommendation, both of questionable value to society. From its roots in World War II era cybernetics and control theory, machine learning has always been political. Advances in artificial intelligence feed into a global industrial military complex, and are funded by it. The success stories told about machine learning also support those who would like to adopt algorithms in domains outside those studied by computer scientists. An opaque marketplace of software vendors renders algorithmic decision making tools for use in law enforcement, criminal justice, education, and social services. In many cases what is marketed and sold as artificial intelligence are statistical methods that virtually haven’t changed in decades.

Many take the leap of faith behind statistical decision making for granted to an extent that it’s become difficult to question. Entire disciplines have embraced mathematical models of optimal decision making in their theoretical foundations. Much of economic theory takes optimal decisions as an assumption and an ideal of human behavior. In turn, other disciplines label deviations from mathematical optimality as “bias” that invites elimination. Volumes of academic papers speak to the evident biases of human decision makers.

In this book, we take machine learning as a reason to revisit this leap of faith and to interrogate how institutions make decisions about individuals. Institutional decision making has long been formalized via bureaucratic procedures and machine learning shares much in common with it. In many cases, machine learning is adopted to improve and sometimes automate the high-stakes decisions routinely made by institutions. Thus, we do not compare machine learning models to the subjective judgments of individual humans, but instead to institutional decision-making. Interrogating machine learning is a way of interrogating institutional decision making in society today and for the foreseeable future.

If machine learning is our way into studying institutional decision making, fairness is the moral lens through which we examine those decisions. Much of our discussion applies to concrete screening, selection, and allocation scenarios. A typical example is that of an employer accepting or rejecting job applicants. One way to construe fairness in such decision making scenarios is as the absence of discrimination. This perspective is micro insofar as individuals are the unit of analysis. We study how measured characteristics of an individual lead to different outcomes. Individuals are the sociological building block. A population is a collection of individuals. Groups are subsets of the population. A decision maker has the power to accept or reject individuals for an opportunity they seek. Discrimination in this view is about wrongful consideration on the basis of group membership. The problem is as much about what wrongful means as what is on the basis of. Discrimination is also not a general concept. It’s domain specific as it relates to opportunities that affect people’s lives. It’s concerned with socially salient categories that have served as the basis for unjustified and systematically adverse treatment.

The first chapter after the introduction explores the properties that make automated decision making a matter of significant and unique normative concern. In particular, we situate our exploration of machine learning in a longer history of critical reflection on the perils of bureaucratic decision making and its mechanical application of formalized rules. Before we even turn to questions of discrimination, we first ask what makes automated decision-making legitimate in the first place. In so doing, we isolate the specific properties of machine learning that distinguish it from other forms of automation along a range of normative dimensions.

Since the 1950s, scholars have developed formal models of discrimination that describe the unequal treatment of multiple different groups in the population by a decision maker. In Chapter 3, we dive into statistical decision theory, allowing us to formalize a number of fairness criteria. Statistical fairness criteria express different notions of equality between groups. We boil down the vast space of formal definitions to essentially three different mutually exclusive definitions. Each definition resonates with a different moral intuition. None is sufficient to support conclusive claims of fairness. Nor are these definitions suitable targets to optimize for. Satisfying one of these criteria permits blatantly unfair solutions. Despite their significant limitations, these definitions have been influential in the debate around fairness.

Chapter 4 explores the normative underpinnings of objections to systematic differences in the treatment of different groups and inequalities in the outcomes experienced by these groups. We review the many accounts of the wrongfulness of discrimination and show how these relate to various views of what it would mean to provide equality of opportunity. In doing so, we highlight some tensions between competing visions of equality of opportunity—some quite narrow and others quite sweeping—and the various arguments that have been advanced to help settle these conflicts. With this in place, we then explore how common moral intuitions and established moral theories can help us make sense of the formalisms introduced in Chapter 3, with the goal of giving these definitions greater normative substance.

Present in both technical and legal scholarship on discrimination is the idea of assigning normative weight to causal relationships. Was group membership the cause of rejection? Would the applicant have been rejected had he been of a different race? Would she have been accepted but for her gender? To understand these kinds of statements and the role that causality plays in discrimination, Chapter 5 of this book is a self-contained introduction to the formal concepts of causality.

Following our formal encounter with fairness definitions, both statistical and causal, we turn to the legal dimensions of discrimination in the United States in Chapter 6. The legal situation neither maps cleanly to the moral foundations nor the formal work, complicating the situation considerably. The two dominant legal doctrines, disparate treatment and disparate impact, appear to create a tension between explicit consideration of group membership and intervening to avoid discrimination.

Extending on both the causal and legal chapters, Chapter 7 goes into detail about the complexities of testing for discrimination in practice through experiments and audits.

Studying discrimination in decision making has been criticized as a narrow perspective on a broader system of injustice for at least two reasons. First, as a notion of discrimination it neglects powerful structural determinants of discrimination, such as laws and policies, infrastructure, and education. Second, it orients the space of intervention towards solutions that reform existing decision making systems, in the case of machine learning typically via updates to an algorithm. As such the perspective can seem to prioritize “tech fixes” over more powerful structural interventions and alternatives to deploying a machine learning system altogether. Rather than predicting failure to appear in court and punishing defendants for it, for example, perhaps the better intervention is to facilitate access to court appointments by providing transportation and child care. Chapter 8 introduces the reader to this broader perspective and its associated space of interventions from an empirical angle.

Recognizing the importance of a broader social and structural perspective, why should we continue to study the notion of discrimination in decision making? One benefit is that it provides a political and legal strategy to put pressure on individual decision makers. We can bring forward claims of discrimination against a specific person, firm, or institution. We can discuss what interventions exist within reasonable proximity to the decision maker that we therefore expect the decision maker to implement. Some such micro interventions may also be more directly feasible than structural interventions.

Taking on a micro perspective decidedly does not mean to ignore context. In fact, allocation rules that avoid explicit consideration of group membership while creating opportunity for a group likely do so by connecting the allocation rule with external social facts. One prominent example is the “Texas ten percent rule” that guarantees Texas students who graduated in the top ten percent of their high school class automatic admission to all state-funded universities. The rule wouldn’t be effective in promoting racial diversity on public university campuses if high school classes weren’t segregated to begin with. This example illustrates that there is no mutual exclusivity between examining specific decision rules in detail and paying attention to broader social context. Rather these go hand in hand.

A consequential point of contact between the broader social world and the machine learning ecosystem are datasets. A full chapter explores the history, significance, and scientific basis of machine learning datasets. Detailed consideration of datasets, the collection and construction of data, as well as the harms associated with data tend to be lacking from machine learning curricula.

Fairness remains an active research area that is far from settled. We wrote this book during a time of explosive research activity. Thousands of related papers have appeared in the last five years of writing. Many of them propose fairness-promoting algorithmic interventions. This text is not a survey of this rapidly evolving area, nor is it a definitive reference. The final chapter, available online, provides an entry point to the emerging research on algorithmic interventions.

The book has some serious, perhaps obvious, limitations.

Large parts of our book are specific to the United States. Written by three authors educated and employed at US institutions, the book is based on Western moral tradition, assumes the laws and legal theory of the United States, and references the industrial and political context of the United States throughout. We made no attempt to address this serious limitation within this book. Indeed, it would require an entirely different book to address this limitation.

A second limitation stems from the fact that our primary goal was to develop the moral, normative, and technical foundations necessary to engage with the topic. Due to its focus on foundations, the book will strike some as a step removed from the important experiences of those individuals and communities most seriously wronged and harmed by the use of algorithms. This shortcoming is exacerbated by the fact that the authors of this book lack first-hand experience of the systems of oppression that algorithms are a part of. Consequently, this book is no substitute for the vital work of those activists, journalists, and scholars that have taught us about the dangers of algorithmic decision making in context. We build on these essential contributions in writing this book. We aimed to highlight them throughout, anticipating that we likely fell short in some significant ways.

The book is neither a wholesale endorsement of algorithmic decision making, nor a broad indictment. In writing this book, we attempt what is likely the least popular position on any topic: a balance. We try to work out where algorithmic decision making has merit, while committing significant attention to its harms and limitations. Some will see our balancing act as a lack of political commitments, a sort of bothsideism.

Despite the urgency of the political situation, our book provides no direct practical guide to fair decisions. As a matter of fact, we wrote this book for the long haul. We’re convinced that the debates around algorithmic decision making will persist. Our goal is to strengthen the intellectual foundations of debates to come, which will play out in thousands of specific instances. Anyone hoping to shape this future of algorithmic decision making in society will likely find some worthwhile material in this book.

A few chapters, specifically Chapter 3 on classification and Chapter 5 on causality, require significant mathematical prerequisites, primarily in undergraduate probability and statistics. However, the other chapters we dedicate to much broader audiences. We hope that students in multiple fields will find this book helpful in preparing for research in related areas. The book does not fit neatly into the disciplinary boundaries of any single department. As a result it gives readers an opportunity to go beyond established curricula in their primary discipline.

Since we’ve started publishing material from this book years ago, instructors have incorporated the material into a variety of courses, both at the undergraduate and graduate level, in different departments. Hundreds of readers have sent us tremendously helpful feedback for which we are deeply grateful.

And to those lamenting our slow progress in writing this book, we respond empathetically:

That’s fair.

Last updated: Thu Jun 23 08:36:58 CEST 2022