Paper review · October 31, 2026
Data provenance and ethics education in US medical AI programs
Data Provenance and Ethics in the Curricula of U.S. Health AI Education Programs
Analyzing the public curricula of 36 medical AI programs at US universities, the authors found that most teach ethics but rarely teach how clinical data is generated (provenance).
Problem
Failures of deployed clinical AI systems often stem from how the training data was created. Yet it is unknown whether programs training the medical AI workforce teach students to examine data provenance.
Method
From April to July 2026, the authors identified medical AI programs at US universities through a structured multi-source web search and conducted a cross-sectional analysis of their public curricula. Two researchers coded the level of data provenance instruction (structural / dataset limitations / data management only), inclusion of ethics and sustainability, collaboration structure, technical prerequisites, and instructors' academic backgrounds and gender. The primary outcome was whether provenance was taught at the structural or dataset-limitations level.
Results
- Of 43 candidates, 36 programs offered by 25 units at 21 universities were included in the analysis.
- Ethics was covered by 29 programs (80.6%, 95% CI 65.0–90.2), but data provenance was covered by only 8 (22.2%) and structural provenance by only 6 (16.7%).
- None of the 7 master's programs met the primary outcome measure, and no program covered the environmental cost of computation (sustainability).
- 18 programs included team-based projects, but only 5 formed teams that mixed clinical and technical backgrounds.
- Of the 92 instructors whose names were public, 76.1% came from computational fields and 3.3% from the social sciences or humanities. Of the 51 whose gender could be determined, 66.7% were men.
Limitations
This summary is based on the abstract. The study analyzed only publicly available curricula, which may differ from what is actually taught; it was limited to US universities, and the sample (36 programs) is small. Instructor gender could be determined for only 51 of 92 instructors.
Why it matters
The safety of medical AI often depends less on modeling techniques than on understanding the context in which data is produced. This study quantifies that gap at the workforce training stage, which makes it useful when designing degree programs and in-house training.
Key terms
- Data provenance
- Information about who produced the data, and through what process and in what context. It is key to understanding the biases and limitations of clinical data.
- Cross-sectional analysis
- A study design that surveys the state of a target population at a single point in time to identify its distribution and characteristics.
Original
Authors: Raffaele Giancotti, Francesca Santucci, Amama Anjum, Monique de Villa, Rupsa Mitra, Soham Biren Katlariwala, Vishala Mishra, Leo Anthony Celi · Posted 2026-10-29 · License arXiv 비독점 배포 라이선스 · Based on the abstract
This review is an AI-written summary of the paper. Please treat the original paper as the reference.