Decisionalizing the problem of reliance on expert and machine evidence

Abstract

Abstract This article analyzes and discusses the problem of reliance on expert and machine evidence, including Artificial Intelligence output, from a decision-analytic point of view. Machine evidence is broadly understood here as the result of computational approaches, with or without a human-in-the-loop, applied to the analysis and the assessment of the probative value of forensic traces such as fingermarks. We treat reliance as a personal decision for the factfinder; specifically, we define it as a function of the congruence between expert output in a given case and ground truth, combined with the decision-maker’s preferences among accurate and inaccurate decision outcomes. The originality of this analysis lies in its divergence from mainstream approaches that rely on standard, aggregate performance metrics for expert and AI systems, such as aggregate accuracy rates, as the defining criteria for reliance. Using fingermark analysis as an example, we show that our decision-theoretic criterion for the reliance on expert and machine output has a dual advantage. On the one hand, it focuses on what is really at stake in reliance on such output and, on the other hand, it has the ability to assist the decision-maker with the fundamentally personal problem of deciding to rely. In essence, our account represents a model- and coherence-based analysis of the practical questions and justificatory burden encountered by anyone required to deal with computational output in forensic science contexts. Our account provides a normative decision structure that is a reference point against which intuitive viewpoints regarding reliance can be compared, which complements standard and essentially data-centered assessment criteria. We argue that these considerations, although primarily a theoretical contribution, are fundamental to the discourses on how to use algorithmic output in areas such as fingerprint analysis.