Research from the University of Cologne demonstrated that reliance on AI assistance can erode diagnostic accuracy. In a 2023 experiment, radiologists reviewed 50 mammograms each with AI support; initial trust in the system led to high performance, but when the AI began providing incorrect answers, the least experienced doctors saw accuracy drop from roughly 80% to under 20%, while senior radiologists fell from 82% to about 46%.
The findings highlight a core weakness of human‑in‑the‑loop models: automation bias, where users accept machine suggestions without verification. Early model success builds confidence, reducing verification effort and potentially degrading professional skills, especially when throughput metrics reward speed over thoroughness.
Source: Corporate Compliance Insights