A Reasonable Apprehension of AI Bias

Lessons from R.v. R.D.S.

Abstract

        In 1997, the Supreme Court of Canada rendered a split decision in a case in which the central issue was whether an African Nova Scotian judge who had brought her lived experience to bear in a decision involving a confrontation between a black youth and a white police officer had demonstrated a reasonable apprehension of bias. The case, with its multiple opinions across three courts, teaches us that identifying bias in decision-making is a complex and often fraught exercise.

        Automated decision systems are poised to dramatically increase in use across a broad range of contexts. They have already been deployed in immigration and refugee determination, benefits allocation, and in assessing recidivism risk. There is also a growing use of AI-assistance in human decision-making that deserves scrutiny. For example, generative AI systems may introduce dangerous unknowns when it comes to the source and quality of briefing materials that are generated to inform decision-makers. Bias and discrimination have been identified as key issues in automated decision-making, and various solutions have been proposed to prevent, monitor, and correct potential issues of bias. This paper uses R. v. R.D.S. as a starting point to consider the issues of bias and discrimination in automated decision-making processes, and to evaluate whether the measures proposed to address bias and discrimination are likely to be effective. The fact that R. v. R.D.S. does not come from a decisional context in which we currently use AI does not mean that it cannot teach us—not just about bias itself—but perhaps more importantly about how we think about and process issues of bias.