Accountability in use of clinical AI

The integration of AI in Taiwan's healthcare system is moving beyond administrative functions into diagnostic assistance and treatment pathways. Algorithms are now prioritizing urgent CT scans, flagging suspicious lesions, and suggesting chemotherapy regimens. This indicates a shift from mere paperwork to direct influence on patient care.

This expansion raises critical questions about liability when AI tools err. The article points out that responsibility could be diffused among the attending physician, the hospital, or the software vendor. Previous incidents, such as the Therac-25 accidents and the failure of IBM Watson for Oncology, demonstrate the dangers of overreliance on software without adequate human oversight and contextual understanding.

Legal academic Frank Pasquale's principle, "attribution must precede autonomy," suggests that responsibility should remain with recognizable institutions and clinicians. The article draws an analogy to aviation's cautious deployment of automated air traffic tools, where human controllers maintained final authority. This illustrates a path for clinical medicine to adopt a similar "start, crawl, walk, run" approach.

Taiwan's Artificial Intelligence Basic Act provides broad principles for AI governance, including human oversight and accountability. However, translating these into practical clinical settings requires concrete shifts: tiered, risk-based oversight with rigorous validation, unbundled lifecycle liability assigning responsibility to developers, hospitals, and clinicians, and ongoing post-market surveillance to address data drift as clinical protocols evolve.