This working paper introduces and formalizes the concept of the outcome fallacy in artificial intelligence: the systematic tendency to evaluate AI systems primarily through outcome-based performance metrics while neglecting the quality of the underlying decision processes. The paper argues that outcome metrics—such as accuracy, efficiency, or downstream impact—are structurally insufficient proxies for decision quality in complex, uncertain, and dynamic environments. Decisions are not outputs but processual, informational, and temporal commitments made under uncertainty. Outcomes, by contrast, are retrospective, context-dependent, and often epistemically shallow indicators that arrive too late to support learning, governance, or accountability. Building on decision theory, cognitive science, organizational studies, and AI governance literature, the paper reframes decision quality as a system-level property that cannot be inferred from results alone. It analyzes the structural limitations of outcome-based evaluation, including proxy optimization, Goodhart effects, temporal mismatch, counterfactual blindness, and silent decision degradation. The work further examines the organizational, societal, and governance consequences of outcome-centric AI evaluation, including distorted incentives, erosion of human judgment, moral outsourcing, and regulatory blind spots. The paper proposes a decision-centric perspective on AI evaluation, emphasizing decision processes as first-class objects of analysis. It outlines principles for decision-quality assessment, highlights the role of process-based and leading indicators, and argues for decision-aware governance and oversight mechanisms. Conceptual case illustrations from finance, healthcare, and organizational strategy demonstrate how systems can appear successful by outcome measures while failing at the level of decision quality. This work is intended as an anchor paper for decision-centric approaches to artificial intelligence, contributing to ongoing debates on AI evaluation, robustness, governance, and human–AI collaboration. It is suitable for researchers, practitioners, and policymakers concerned with the long-term reliability, accountability, and integrity of AI-enabled decision systems.
Aleksandra Pinar (Mon,) studied this question.