Clinical AI is typically evaluated as if either the patient or the model remains stable over time. In practice, neither does. Biological systems evolve, and AI models drift as data, context, and clinical environments change. This paper introduces the concept of dual drift — the progressive misalignment that occurs when both the patient and the model evolve simultaneously. The resulting risk is not primarily discrete failure, but a gradual loss of alignment between model and reality that may remain undetected within current governance frameworks. Positioned within the Universal Resonance Model (URM), this work reframes clinical AI as a coupled dynamic system and argues for a shift from static performance evaluation toward continuous monitoring of system-level behaviour and alignment over time.
Anita Domargård (Wed,) studied this question.