Human-Understandable Representations for Complex AI Decisions

Abstract

Complex AI decision systems -- encompassing deep neural networks, large language models, and ensemble architectures -- produce outputs of considerable practical consequence yet generate internal representations that are fundamentally inaccessible to human cognition in their raw form. Human- understandable representations (HURs) are structured reformulations of model-internal information designed to align with human cognitive schemas, domain knowledge, and explanation needs, enabling meaningful human oversight and accountability for AI-assisted decisions. Despite extensive research on explainability methods, the systematic design of HURs -- as distinct from the application of generic attribution techniques -- has received limited attention. This paper proposes a HUR design framework comprising four representation classes (contrastive, causal, conceptual, and narrative) and evaluates their cognitive effectiveness through a user study involving 186 participants across three professional domains: clinical medicine (n = 62), financial advising (n = 62), and legal practice (n = 62). Participants assessed AI decisions with and without HURs across four cognitive effectiveness dimensions: comprehension accuracy, decision confidence calibration, appropriate reliance, and perceived fairness. Results demonstrate that HUR-supported conditions achieve a 41.8% improvement in comprehension accuracy (SD = 6.2%), a 33.5% improvement in decision confidence calibration (SD = 5.4%), and a 27.6% improvement in appropriate reliance (SD = 4.9%) relative to no-HUR controls. Representation class effectiveness is domain-dependent: contrastive HURs are most effective in legal contexts, causal HURs in clinical contexts, and conceptual HURs in financial contexts. The study contributes the HUR design framework, domain-specific design guidelines, and a validated HUR evaluation instrument.