AI-Augmented Human Decision Making

Abstract

AI-augmented human decision making -- systems where AI provides recommendations, predictions, or information to support but not replace human judgment -- is one of the most consequential applications of AI, affecting medical diagnosis, legal sentencing, financial investment, policy analysis, and everyday consumer choices. The evidence on whether AI augmentation improves human decision quality is more nuanced than popular narratives suggest: AI assistance can improve accuracy in some domains and contexts while reducing it in others through automation bias (over-reliance on AI recommendations), deskilling (degradation of human expertise through disuse), and algorithmic aversion (systematic rejection of AI advice even when it is correct). This study evaluates eight AI decision support systems across five decision domains: clinical diagnosis, financial portfolio selection, legal risk assessment, scientific hypothesis evaluation, and crisis resource allocation. Experimental evaluation with 284 domain experts assesses decision accuracy, confidence calibration, time efficiency, and AI reliance behaviour. AI augmentation improves mean decision accuracy by 14.2pp across all domains (from 68.4% to 82.6%) but introduces 22.4% automation bias rate. Calibrated uncertainty communication (conformal AI, explanation-augmented AI) reduces automation bias by 38.4% vs. point-recommendation AI. A proposed Collaborative Intelligence Framework (CIF) achieves the best human-AI team performance (accuracy 86.4%, automation bias 12.4%) by combining uncertainty communication with active human engagement prompts. A human-AI collaboration governance framework for high-stakes domains is proposed.