Doximity asked 3,379 physicians about AI in September. Shrink them to 100 and watch the same room answer seven questions. Most of the dots light up for using it. Far fewer light up for having a written rule.
Primary source: Doximity, “2026 Doctor-Driven AI Report” (3,379 U.S. physicians across 20 specialties, surveyed September 2026 via SurveyMonkey; not population-based) and the Nov. 12 release. All figures are the report’s published percentages.
Light up the room: each dot is 1 of 100 physicians
yesnot this answeroutlined: weekly AI users, for scale
Rounded to whole dots. Each question is shown on its own: the report doesn’t say how the answers overlap, so the lit dots for “caught an error” are not necessarily the same doctors as the lit dots for “uses it daily.” The two setting rows (private practice, government) describe those subgroups, not the whole room.
Use climbed in a year. The rulebook is a single point.
Every usage measure the report tracked in both years went up. Clear policies weren’t asked about in 2025, so there’s no slope to draw, only the red line at 28%.
Clinical decision support and literature search are shares of AI users; the other lines are shares of all respondents. “Reduced after-hours work” is the share saying AI has cut it.
Who should write the rules?
82% of respondents say physicians. 10% say the AI companies. Respondents could pick more than one.
80/20 · the gap
Of 100 physicians, use AI at work weekly and say it’s permitted or supported where they work. say there are clear policies and approvals. That leaves at least in every 100 whose organization allows AI without a clear rule for it. In private practice the room drops to lit dots; in government hospitals and clinics it’s .
Read this before quoting a number
A self-selected sample from a company that sells AI. Doximity says the survey isn’t population-based and its members opted in. Doximity also markets AI tools to these doctors. Read the adoption dots as a ceiling.
“Clear policies” is what the doctor believes. A policy that exists but that nobody on the unit can find shows up here as no policy. That’s arguably the right way to count it.
The subgroups have no denominators. The report gives 53% (government) and 14% (private practice) without saying how many respondents were in each. Academic centers and health-system employees aren’t broken out.
The overlap is unknown. “At least 53 in 100” is the floor from subtracting 28 from 81; the report doesn’t cross-tabulate the two questions.
A survey is the tool here, not a record. Nobody audited whether the AI use was clinical, what tools were used, or what the errors were.