The VERDICT study, led by TUM University Hospital in Munich, investigates how radiologists respond when AI recommendations conflict with their own clinical assessments, focusing on automation bias and human-AI interaction in radiology workflows.
- The VERDICT study recruits radiologists worldwide to examine responses when AI recommendations contradict clinical assessments
- Automation bias is a key concern, where clinicians may overrely on AI recommendations even when incorrect
- Participants complete three text-based clinical scenarios in approximately 10 minutes with no medical images required
- The survey is available in five languages (English, German, French, Italian, Spanish) and recruitment continues until April 15, 2027
Researchers at TUM University Hospital in Munich have launched an international survey investigating how radiologists respond when AI recommendations contradict their own clinical assessments.
The VERDICT study aims to recruit at least 300 radiologists and radiology residents worldwide to investigate how human factors influence AI-assisted reporting decisions.
As AI becomes increasingly integrated into radiology workflows, attention is shifting from algorithm performance alone to how clinicians interpret and respond to AI-generated recommendations.
"A lot of studies in the field of radiology AI have so far focused on developing models and evaluating their performance standalone," said Dr. Su Hwan Kim of the Department of Diagnostic and Interventional Radiology at TUM University Hospital. "But as these AI systems are increasingly adopted, the critical role of human-AI interaction is increasingly recognized."
When radiologists and AI disagree
The VERDICT study focuses on situations in which AI-generated suggestions conflict with radiologists' independent assessments.
"One such phenomenon is automation bias, the tendency to overrely on AI, even when its output is wrong," Kim explained.
The study builds on findings from the research group's previous work on automation bias and seeks to investigate how radiologists approach conflicting information in clinical decision-making.
"We want to understand how radiologists behave when they are confronted with AI outputs that contradict their own assessment," Kim said.
International participation
Participants are asked to complete three short, text-based clinical scenarios and indicate how they would report each case in routine practice. The survey takes approximately 10 minutes, involves no medical images, and has no right or wrong answers.
The scenarios were developed through a two-round Delphi process involving an international expert panel representing radiology, psychology, medical law, and clinical AI implementation.
Implications for clinical AI
According to Kim, the findings could eventually inform AI implementation strategies and policy design.
The survey is available in English, German, French, Italian, and Spanish. Researchers will collect basic information on professional experience, role, country, and practice setting to explore potential differences across countries and experience levels.
Recruitment remains open until April 15, 2027, and the researchers anticipate submitting their findings to a peer-reviewed journal later that year.
Radiologists and radiology residents interested in participating can access the survey at https://radiology-ai-survey.org/.




















