Do health-AI courses teach where clinical data come from? A US audit found the topic in 8 of 36 programmes
Ethics appeared in 29 programmes, but only eight publicly advertised structural or dataset-level provenance teaching. The audit measured published curricula—not what every classroom actually delivered.
The short version
- What happened
- Researchers screened 43 candidate entries and included 36 programmes at 25 institutional units in 21 US universities, using official websites, syllabuses, catalogues and faculty pages available from April to July 2026.
- Who it affects
- Clinicians, data scientists and public-service professionals trained without provenance skills may treat a clinical record as neutral input rather than the product of access, workflow and documentation choices. Programme leaders can use the public dataset as a checklist, but the audit is not a ranking of teaching quality.
- What remains unproven
- Unreviewed preprint under journal review
Keep in mind
Unreviewed preprint under journal review
36 programmes from 21 US universities
Non-probability and non-exhaustive web search
Publicly advertised curricula only
Search ended July 2026
No preregistered protocol
Silence coded as not reported and retained
Ethics depth not graded
Inter-rater agreement unavailable
Instructor gender unavailable for 41 of 92
No student learning or workplace outcome
Funding and competing interests not separately disclosed
Source analysis · 2 linked sources. Primary analysed source date: 29 October 2026.
Read the full report · 5 min The evidence, methods, implications and limitations
Ethics was common in public curricula; data provenance was not
A cross-sectional audit of 36 US university health-AI programmes found that 29, or 80.6%, publicly described some teaching on ethics, responsible AI, fairness, bias, privacy or regulation. Only eight, or 22.2%, described how data generation or dataset limitations constrain what a model can learn. Six reached the narrowest category: teaching how clinical workflow, documentation, representation or selection shape the record itself.
The distinction matters for clinicians and model builders. A course can discuss fairness in general without teaching students to ask whose symptoms were documented, who was missing from a dataset or whether a billing code is a poor proxy for illness. None of the seven master's programmes met the audit's primary provenance outcome. Three additional programmes taught data management such as storage or interoperability, which raised the broadest count to 11 of 36 but did not answer the same question.[1][2]
Two coders checked official programme material, with silence kept in the denominator
The team searched between April and July 2026 using Google, prior reviews and two AI assistants that generated candidates only. Every included programme and every coded claim was checked against university websites, syllabuses, course catalogues or faculty pages. Of 43 candidate entries, two duplicates were merged and five were excluded, leaving 36 programmes at 25 institutional units in 21 universities. The sample mixed degrees, certificates, courses and professional education; 26 offered at least some online access.
Two investigators independently coded the same archived sources and resolved disagreements by consensus. Public silence counted as not reported and stayed in the denominator, matching the study's question about advertised commitments. The primary definition accepted structural provenance or explicit dataset limitations; ethics received credit for any public mention and its depth was not graded. Wilson intervals put primary provenance coverage at 11.7% to 38.1%, reflecting the small non-random sample.[1][2]
The distinction that matters
What counted as provenance teaching
The audit separated the origins and limits of data from ordinary data management.
Structural
How workflow and representation create the record
6 of 36 programmes
Dataset limits
Bias, missingness, shift or generalisability
2 more programmes; 8 of 36 primary total
Data management
Storage, standards, privacy or database design
3 more; excluded from the primary outcome
Ethics
Any public ethics, fairness, privacy or regulation mention
29 of 36 programmes
The result is a curriculum signal, not an observation of teaching quality
Half the programmes listed team projects, capstones or mentored research, but only five explicitly brought clinical and technical backgrounds together or paired teams with clinical mentors. Among 92 named instructors, 70 had computational backgrounds and three came from social sciences or humanities. No programme publicly addressed the environmental or computational cost of AI. These findings describe what pages promised; they do not show what students learned or how graduates behaved.
The audit was not preregistered, the search was not exhaustive and programmes with sparse web pages may be misclassified. Pre-consensus coding was not retained, so inter-rater agreement could not be calculated. Coverage also ranged from 16.7% under the narrowest provenance definition to 30.6% under the broadest. The authors have released the screening register, coded data and sensitivity analysis, and say they plan annual repeats. Stronger evidence would inspect syllabuses and assessments directly, survey students and educators, and test whether provenance teaching changes decisions on real clinical datasets. Funding and competing interests were not separately disclosed in the reviewed manuscript.[1][2]