AI-assisted Penal Sentencing: The Epistemic Free-Riding Objection

Several authors (Laquer & Copus 2017; Leibovitch 2017; Chiao 2018) argue that machine-learning algorithms can and ought to be used by judges at the sentencing stage to predict the statistically typical (average or modal) sentencing decision taken by other (actual or counterfactual) judges in relevantly similar cases, and adjust their individual sentences to cohere with these latter ones. Unlike currently deployed AI-informed tools, the proposal here is to use algorithms to predict judicial, not offender behavior. Furthermore, unlike existing static actuarial tables or sentencing guidelines and grids, such algorithms proceed dynamically – viz., by updating sentence predictions based on the decisions taken by individual judges. The contention is that these algorithmic tools can secure more consistency among sentencing decisions while preserving judges’ substantive commitment to reasonably defensible penal principles. The argument of this paper is twofold. First, it argues that the decision-making situation that such proposals would instantiate can be described as one that prima facie satisfies the conditions of the Condorcet Jury Theorem (CJT) – viz., one where the average competence of decision-makers is better than random, where decision-makers share the same goal, and where their judgments are independent. Because of this, the proposal seems epistemically desirable. Second, I draw on List & Pettit (2004) and Dunn (2018) to further argue that, insofar as they believe that sentencing algorithms create situations that satisfy CJT, judges are individually justified to adjust their sentencing decisions with statistically typical ones. Insofar as this happens, and because sentencing is a temporally deployed process, judges’ beliefs that CJT is satisfied would also rationally motivate them to epistemically free-ride on other judges’ decisions, and thereby eventually prompt a situation that violates the independent judgment condition posited by CJT. Thus, envisaged diachronically, the proposed algorithms are an epistemic liability.