JMIR Publications

ABSTRACT

Background:

Documentation burden contributes to physician burnout and may shape primary care career trajectories. Ambient artificial intelligence (AI) scribes have spread into routine clinical use faster than the evidence supporting them, and most published evaluations involve attending physicians. Resident studies to date are small cohorts focused on efficiency and well-being endpoints. How residents take up these tools in routine practice, which tasks they use them for, what supervision concerns they hold, and whether scribe access changes how trainees view a primary care career remain open questions.

Objective:

This study aimed to characterize family medicine resident physicians’ self-reported adoption, perceived workflow impact, and AI-related attitudes following a multisite ambient AI scribe implementation, and to identify training and supervision considerations relevant to graduate medical education (GME).

Methods:

We conducted a pre-post survey study with supplementary qualitative analysis at 4 Kaiser Permanente Northern California (KPNC) family medicine residency programs between October 2024 and April 2025. All postgraduate year 1 through 3 (PGY1-3) family medicine residents (N=72) at participating sites were invited via email from program leadership to complete pre- and post-implementation surveys. Surveys captured documentation practices, perceived burden, AI use, and primary care career intentions using investigator-developed Likert items and open-ended questions. Wilcoxon signed-rank tests evaluated paired changes (n=23), with effect sizes (r) reported; open-ended responses underwent inductive thematic content analysis. A sensitivity analysis characterized the minimum detectable effect size.

Results:

Twenty-three residents completed both surveys (paired response rate 32%; 23/72). AI use rose substantially for note generation (mean 1.26 to 3.61 on a 5-point frequency scale; P<.001; r=0.851) and history collection (1.26 to 3.39; P<.001; r=0.825). Documentation efficiency measures showed small-to-medium effect sizes that did not reach statistical significance, consistent with limited power (80% power only for effects of d≈0.6 or larger). Residents’ endorsement of unassisted note writing as an important skill did not decline with adoption (3.61 to 3.83; P=.17). Qualitative themes included increased confidence in managing a patient panel, perceived time savings, greater patient engagement during visits, and renewed interest in primary care, alongside persistent concerns about workload expectations and note verbosity.

Conclusions:

Following ambient AI scribe implementation across 4 programs, family medicine residents reported rapid, substantive growth in AI use for documentation tasks (reflecting scribe adoption alongside concurrent general-purpose AI uptake), while continuing to value unassisted documentation skills and raising concerns about workload expectations and note quality. Self-selection, the low response rate, the small paired sample, and reliance on investigator-developed self-report items limit generalizability. These findings complement recent resident pilots centered on objective efficiency endpoints by characterizing self-reported adoption, perceptions, and career outlook across all 3 training years, and they provide formative data for GME implementation with guardrails and for larger controlled studies using validated instruments and objective electronic health record-derived measures.