Efficiency pitfalls of explainable AI in clinical diagnostic and treatment human-AI workflows.

The transparency of AI decision content is heralded as important but its integration into clinical workflow may harbor unintended consequences. This mixed methods study looked at AI-supported diabetic retinopathy diagnosis, both with and without visual explanations (highlighting lesions). The visual explanations increased diagnostic time and decreased efficiency, showing how benefits of AI may actually decrease productivity when needed "human in loop" and transparency are incorporated.