About 71% of respondents to a US family medicine residency survey reported using AI for work in the previous month. About 36% of all respondents reported AI education in the previous six months. Those figures describe reported use and education in October 2025; they do not measure competence, patient benefit or whether training caused adoption. National Resident Survey study
For educators, the interesting gap is between having used a tool and knowing how to assess its output. The survey supplies a starting point for that question, rather than an answer about who uses AI well.
Keep the population and the questions visible
The study, published in Family Medicine in October 2026, analysed AI questions completed by 15,259 family medicine residents after their In-Training Exam. Of those respondents, 10,858 reported recent AI use. Common reported purposes included medical reference, decision support, learning and documentation.
These are self-reported activities within one specialty’s training population. They are not observations of how accurately residents performed clinical tasks. Several uses could occur in the same person, so the categories should not be added together as separate groups of residents. The survey also cannot establish how much of a working day depended on AI. Methods and overall results
The published table contains an inconsistent training subgroup count. This article therefore uses the reported overall findings, rounded to whole percentages, and makes no comparison of training rates within the AI-user subgroup. It also does not subtract the two headline percentages to estimate how many AI users lacked training.
Opening a tool is one skill; judging it involves others
Yale’s account of the study highlights the need to connect AI instruction with critical appraisal. Wendy Barr, the lead author, frames the issue as learning to use information appropriately while clinicians are still developing their reasoning skills. Those concerns are an educator’s interpretation, not harms measured by this survey. Yale’s study discussion
An earlier primary-care education paper proposes a more specific vocabulary for the training task. Its competency domains cover understanding a tool, appraising evidence, deciding when to use it, operating it, communicating with patients and recognising unintended consequences. It is a proposed framework, rather than a trial demonstrating that a particular curriculum works. Primary-care AI competencies
That vocabulary exposes what a yes/no education question leaves open. An introduction to an interface could count as education without showing whether someone can recognise an unsuitable output. Conversely, a survey answer cannot reveal every skill learned through supervision or experience.
What evidence would add to the picture?
A useful next step would connect defined teaching with assessed skills and subsequent practice, keeping the clinical task and the AI system identifiable. Repeating an adoption question alone would still leave those outcomes unresolved.
This is also a different question from what clinicians ask a particular evidence platform. Platform use describes demand within a service; a training survey describes respondents’ activities and reported education. Neither, without further evidence, establishes that the resulting decisions were better.