Automation Bias, Anchoring, and Confirmation Effects in Physician-AI Diagnostic Interaction: A Systematic Review and Framework for Bias-Neutral CDSS Design under the EU AI Act

Integration of artificial intelligence into clinical decision support systems has created expectations of improved diagnostic accuracy. However, emerging evidence demonstrates that physician-AI interaction introduces systematic cognitive distortions-automation bias, anchoring effects, and confirmation bias-that may compound rather than correct diagnostic errors. Under the EU AI Act, AI-based medical devices are classified as high-risk systems requiring rigorous conformity assessment, yet current regulatory frameworks do not explicitly address cognitive bias in human-AI interaction as a distinct risk factor. This systematic review analyzes empirical evidence from 2020-2026, examining trust calibration as a moderating mechanism. Key findings: incorrect AI suggestions reduced diagnostic accuracy from ~80% to ~20% among less experienced physicians; anchoring effects showed statistically significant positive coefficients under time pressure; confirmation bias produced compounding error effects. Trust calibration interventions demonstrated limited efficacy: explicit reliability prompts did not improve accuracy, whereas accurate calibration was associated with 5.9-fold higher diagnostic odds. The review proposes an evidence-based regulatory framework for bias-neutral CDSS design aligned with EU AI Act and MDR requirements, advocating for mandatory cognitive bias evaluation in conformity assessment.