Ethical guidance for reporting and evaluating claims of AI outperforming human doctors

Abstract

An increasing number of academic reports contend that Artificial Intelligence (AI), in particular machine learning systems, surpasses medical practitioners’ performance in various clinical tasks and specialisms 1 , 2 , 3 . These outperformance claims—as we will refer to them in this article—vary in their formulation. Commonly used terms include ‘outperform,’ ‘surpass,’ ‘exceed,’ ‘better than’, and ‘superior to,’ but all claims are based on the fundamental assumption that AI can be directly compared to and exceed the expertise of medical practitioners on some level (for examples, see Table 1 ). These reports have contributed to excitement about AI’s potential value for medical contexts and have raised hopes about automating specific diagnostic tasks 4 , 5 . The claims about outperformance have also been met with skepticism because of methodological flaws in AI studies. For example, it remains uncertain whether AI can outperform medical practitioners in clinical practice because model performance is often evaluated in unrealistic settings 4 . Furthermore, many studies fail to transparently report the circumstances under which AI is compared to medical practitioners, making it hard to verify claims of outperformance 1 . Several scholars have concluded that reports on AI’s performance in medicine are “exaggerated,” 6 , 7 and that it is “time to reality check” these kinds of claims to distinguish genuine potential from hype 4 . Some scholars even warn against using terms like ‘outperform’ because overpromising language risks being misinterpreted by the media and the public 1 , 6 , 7 , 8 and may result in “sidestepping ethical concerns leaving no space for issues and criticism” on AI’s functioning in medical practice 8 .

Topics & Concepts