Trust by design: Crossing the chasm between clinical AI/ML innovation and practice

Clinical artificial intelligence (AI) and machine learning (ML) have advanced rapidly, yet few systems achieve sustained use in routine clinical care. This innovation-practice gap reflects not a lack of model performance, but persistent failures in validation, governance, and trust when AI is deployed in real-world clinical environments. Models that perform well retrospectively often degrade prospectively, misalign with local workflows, or are ultimately abandoned, with current responses relying on people- and process-centric mechanisms such as expert review, bespoke validation, and discretionary governance. These approaches are slow, fragile, and difficult to scale, making trust the primary bottleneck to adoption. We argue for a shift toward trust by design, in which trust is treated as a system property rather than a judgment conferred by individuals or committees. We describe how trusted systems, including privacy-enhancing compute, trusted execution environments, and trusted research environments, embed enforceable guarantees around data use, auditability, and reproducibility. By operationalizing trust through infrastructure, these systems enable scalable local validation, reduce duplicated effort, and support more equitable and durable deployment of clinical AI/ML.