There is a broad consensus that human supervision holds the key to sound automated decision-making: if a decision-making policy uses the predictive outputs of a statistical algorithm, but those outputs form only part of a decision that is made ultimately by a human actor, use of those outputs will not (per se) fall foul of the requirements for due process in public and private decision-making. Thus, the focus in academic and judicial spheres has been on making sure that humans are equipped and willing to wield this ultimate decision-making power. Yet, proprietary software obscures the reasons for any given prediction; this is true both for machine learning and deterministic algorithms. And without these reasons, the decision-maker cannot accord appropriate weight to that prediction in their reasoning process. Thus, a policy of using opaque statistical software to make decisions about how to treat others is unjustified, however involved humans are along the way.

Technological tools currently being developed are capable of substantially assisting judges in their daily work. In particular, data analysis of widely available court decisions will evaluate, in an unprecedented way, the activity of the courts and the quality of justice. In doing so, it will allow for more efficient and faster dispute resolution, as well as cost reductions for litigants and society. Technological evolution will probably not cause the disappearance of humans from judicial adjudication but a new, progressive and subtle redistribution of tasks between men and machines.

A little after 2 a.m. on February 11, 2013, Michael Vang sat in a stolen car and fired a shotgun twice into a house in La Crosse, Wisconsin. Shortly afterward, Vang and Eric Loomis crashed the car into a snowbank and fled on foot. They were soon caught, and police recovered spent shell casings, live ammunition, and the shotgun from the stolen and abandoned car. Vang pleaded no contest to operating a motor vehicle without the owner’s consent, attempting to flee or elude a traffic officer, and possession of methamphetamine. He was sentenced to ten years in prison.

When agents insert technological systems into their decision-making processes, they can obscure moral responsibility for the results. This can give rise to a distinct moral wrong, which we call “agency laundering.” At root, agency laundering involves obfuscating one’s moral responsibility by enlisting a technology or process to take some action and letting it forestall others from demanding an account for bad outcomes that result. We argue that the concept of agency laundering helps in understanding important moral problems in a number of recent cases involving automated, or algorithmic, decision-systems. We apply our conception of agency laundering to a series of examples, including Facebook’s automated advertising suggestions, Uber’s driver interfaces, algorithmic evaluation of K-12 teachers, and risk assessment in criminal sentencing. We distinguish agency laundering from several other critiques of information technology, including the so-called “responsibility gap,” “bias laundering,” and masking.

One important criticism of algorithmic systems is that they lack transparency, either because they are complex, protected by intellectual property, or deliberately obscure. There is a debate about whether the EU’s General Data Protection Regulation (GDPR) contains a “right to explanation” This chapter addresses the informational component of algorithmic systems. We argue that information access is integral for respecting autonomy, and transparency policies should be tailored to advance autonomy. We distinguish two facets of agency (i.e., capacity to act). The first is “practical agency,” or the ability to act effectively according to one’s values. The second is “cognitive agency,” which is the ability to exercise what Pamela Hieronymi calls “evaluative control”. We argue that respecting autonomy requires providing persons sufficient information to exercise evaluative control and properly interpret the world and one’s place in it. We draw this distinction out by considering algorithmic systems used in background checks, and we apply the view to key cases involving risk assessment in criminal justice decisions and K-12 teacher evaluation.

Chapter 3 takes the conception of autonomy outlined in chapter 2 and explains how it grounds moral evaluation of algorithmic systems. It begins by offering a view of what it takes to respect autonomy and to respect persons in virtue of their autonomy, drawing on a number of different normative moral theories. The argument starts with a description of a K-12 teacher evaluation program from Washington, DC. It then considers several puzzles about the case. Next, the chapter provides an account of respecting autonomy and what that means for individuals’ moral claims. It explains how that account can help us understand the DC case, and we will offer a general account of the moral requirements of algorithmic systems. Specifically, we offer the Reasonable Endorsement Test, according to which an action is morally permissible only if it would be allowed by principles that each person subject to it could reasonably endorse. The chapter applies that test to the Loomis, Houston Schools, and Wagner cases. Finally, the chapter explains why the book does not focus directly on “fairness.”