A Right to a Human Decision

Abstract

Recent advances in computational technologies have spurred anxiety about a shift of power from human to machine decision-makers. From prison sentences to loan approvals to college applications, corporate and state actors increasingly lean on machine learning tools (a subset of artificial intelligence) to allocate goods and to assign coercion. Machine-learning tools are perceived to be eclipsing, even extinguishing, human agency in ways that sacrifice important individual interests. An emerging legal response to such worries is a right to a human decision. European law has already embraced the idea in the General Data Protection Regulation. American law, especially in the criminal justice domain, is already moving in the same direction. But no jurisdiction has defined with precision what that right entails, or furnished a clear justification for its creation.

This Article investigates the legal possibilities of a right to a human decision. I first define the conditions of technological plausibility for that right as applied against state action. To understand its technological predicates, I specify the margins along which machine decisions are distinct from human ones. Such technological contextualization enables a nuanced exploration of why, or indeed whether, the gaps that do separate human and machine decisions might have normative import. Based on this technological accounting, I then analyze the normative stakes of a right to a human decision. I consider three potential normative justifications: (a) an appeal to individual interests to participation and reason-giving; (b) worries about the insufficiently reasoned or individuated quality of state action; and (c) arguments based on negative externalities. A careful analysis of these three grounds suggests that there is no general justification for adopting a right to a human decision by the state. Normative concerns about insufficiently reasoned or accurate decisions, which have a particularly powerful hold on the legal imagination, are best addressed in other ways. Similarly, concerns about the ways that algorithmic tools create asymmetries of social power are not parried by a right to a human decision. Indeed, rather than firmly supporting a right to a human decision, available evidence tentatively points toward a countervailing ‘right to a well-calibrated machine decision’ as ultimately more normatively well- grounded.

Keywords: Machine learning; due process; artificial intelligence; discrimination