AI and XAI second opinion: the danger of false confirmation in human-AI collaboration.

Abstract

Can AI substitute a human physician's second opinion? Recently the Journal of Medical Ethics published two contrasting views: Kempt and Nagel advocate for using artificial intelligence (AI) for a second opinion except when its conclusions significantly diverge from the initial physician's while Jongsma and Sand argue for a second human opinion irrespective of AI's concurrence or dissent. The crux of this debate hinges on the prevalence and impact of 'false confirmation'-a scenario where AI erroneously validates an incorrect human decision. These errors seem exceedingly difficult to detect, reminiscent of heuristics akin to confirmation bias. However, this debate has yet to engage with the emergence of explainable AI (XAI), which elaborates on why the AI tool reaches its diagnosis. To progress this debate, we outline a framework for conceptualising decision-making errors in physician-AI collaborations. We then review emerging evidence on the magnitude of false confirmation errors. Our simulations show that they are likely to be pervasive in clinical practice, decreasing diagnostic accuracy to between 5% and 30%. We conclude with a pragmatic approach to employing AI as a second opinion, emphasising the need for physicians to make clinical decisions before consulting AI; employing nudges to increase awareness of false confirmations and critically engaging with XAI explanations. This approach underscores the necessity for a cautious, evidence-based methodology when integrating AI into clinical decision-making.