Epistemology Reframes Human-AI Complementarity

A new paper on arXiv reframes human-AI complementarity—the idea that humans and AI together outperform either alone—using epistemology. The authors argue that complementarity lacks precise theoretical anchoring, is only formalized as a post hoc accuracy indicator, ignores other interaction desiderata, and abstracts away from cost profiles, making it hard to achieve empirically. Drawing on computational reliabilism, they propose grounding complementarity in justificatory AI discourse, using historical instances to address these challenges.