Interpretability, Trust, and Data in Medical AI
16. Juli 2026
In this interview, Dr. Johannes Winter and Yuliia Borovynska speak with Prof. Dr. Kerstin Lenhof, assistant professor and head of the Integrative Bioinformatics research group at the University Medical Center Göttingen, as well as with Dr. Rosa Esther Martín Peña, head of the Ethics Group at CAIMed and a postdoc at the Center for Law and Ethics of the Life Sciences at Leibniz University Hannover, about the interpretability of AI systems in everyday clinical practice. How understandable do AI models really need to be so that doctors and patients can trust them without having to question every result?
Why Plausibility Is Not Enough in Medicine
Kerstin Lehnhof emphasizes that a language model can use medical terms correctly and still be clinically incorrect. Medical language is not neutral: The same output from an AI system can be interpreted very differently depending on the reader’s prior knowledge, clinical experience, and context. Interpretability is therefore not purely a technical property of a model, but always depends on the person working with it.
Interpretability for Different Audiences
Another key point of the discussion: What AI developers consider interpretable is by no means necessarily so for clinicians or patients. The two researchers agree that AI systems must be designed to be transparent and understandable for all stakeholders, although different mechanisms are required depending on the target audience.
Trust is not a one-way street
Even though AI systems now deliver impressive results in many fields, users must familiarize themselves with the technology and critically evaluate the AI’s results, especially in a medical context.
Regulation, Data, and the Pace of Development
The researchers also discuss the challenge posed by the fact that technological development often far outpaces regulatory and societal discussions. Issues such as the legal framework for trustworthy AI, the use of sensitive clinical data from the Medical Data Integration Centers at university hospitals, and questions of data quality—which can never be answered in general terms but only within the specific clinical context—play an important role in this regard.
The future is hybrid
At the end of the conversation, both researchers agree: The future of medicine lies in the collaboration between humans and AI. Artificial intelligence can be an enormously helpful tool, but it cannot replace a doctor’s responsibility.
CAIMed is committed to ensuring that interpretability, trust, and ethical responsibility are taken into account from the very beginning in order to develop AI solutions that truly stand the test of time in everyday clinical practice.