When you sit in front of your doctor, you might assume their advice comes from training, experience and maybe a quick look at a medical reference.

Increasingly, it may also come from an AI.

A new study in JAMA offers the first hard count of how widely doctors and other clinicians in the US are using AI in their daily work. The numbers are striking.

Monthly users grew from 355 to 700 000 in two and a half years

The researchers analyzed usage logs from OpenEvidence, an AI tool that answers medical questions with references to research and guidelines. It’s free, but only available to verified US healthcare professionals.

Between January 2024 and June 2026:

  • 952 000 clinicians used it, including about 556 000 physicians, plus nurse practitioners, physician assistants and pharmacists.
  • They asked 271 million questions.
  • Monthly users grew from 355 to more than 703 000, roughly a 2 000-fold increase.

To put the scale in everyday terms (our own calculations from the study’s figures):

  • By June 2026, clinicians were asking about 27.8 million questions a month. That’s roughly 900 000 a day, or about 10 every second.
  • The average physician user asked about 37 questions a month, more than one per working day.
  • And they kept coming back: 92.5% of active users in June 2026 were returning users, not people trying it once.

For context, the US has roughly a million practicing physicians, so more than half a million physician users represents a very large share of the profession.

The real number is probably higher

This study counts just one tool.

Doctors have other options, and surveys suggest many of them use several:

  • In a Doximity survey of more than 3 000 US physicians in January 2026, 63% said they use AI in clinical practice, up from 47% a year earlier. Most of those users said they use it daily.
  • In an international survey of physicians in March 2026, the most commonly reported AI tool was ChatGPT (83%), followed by Gemini (50%), with OpenEvidence at 36% and UpToDate’s AI search at 32%.

These surveys have their own limits: they rely on self-reports and different samples. But together they suggest that the true number of clinicians using AI is likely higher than what any single platform can show.

Questions were mostly about treatment, not diagnosis

The researchers used AI to sort a sample of 38.8 million questions by purpose, and checked its labels against human clinicians. The AI agreed with the clinicians about as often as the clinicians agreed with each other.

The results challenge a popular assumption:

  • About half of questions were about clinical decisions for patients.
  • The biggest category was treatment and management, such as choosing therapies or adjusting medicines: 33% of physician questions.
  • Diagnosis, which much AI research focuses on, made up only about 14%.

That matters because many tests of medical AI focus on whether it can diagnose correctly. But in practice, clinicians mostly use it to help decide what to do next.

Young doctors and nurse practitioners led adoption

Adoption is uneven:

  • Young doctors lead. Physicians within 10 years of graduating make up about a quarter of the workforce but submitted more than 60% of physician questions. Doctors 20 or more years out made up about half the workforce but only about one sixth of questions.
  • Nurse practitioners and physician assistants asked the most questions per person, about 45 a month.
  • Specialties differ. Use was highest in cancer care, internal medicine and family medicine, and lowest in eye care, radiology and orthopedic surgery.
  • Doctors often ask outside their field. The share of questions about topics outside a physician’s own specialty ranged from 35% for cancer specialists to 88% for emergency physicians.

That last finding makes sense. An emergency doctor might see a heart problem, a skin infection and a pregnancy complication in the same shift. AI may act like a colleague down the hall from every specialty at once.

Medicine adopted AI from the bottom up

The bigger story is how this happened.

New medical technologies often arrive top-down: approved, purchased by hospitals and rolled out with training and rules.

This was different. Individual clinicians, especially younger ones, started using a free tool on their own, and its use spread at remarkable speed.

Meanwhile, the authors note, policymakers, professional societies and health systems have been writing guidelines without data on how these tools are actually used.

It’s a striking example of how quickly healthcare workers can adopt a new technology when it solves a real everyday problem: finding reliable answers fast in a field where medical knowledge long ago outgrew what any one person can remember.

What the study doesn’t show

This is a peer-reviewed study in JAMA, and a strong description of how one tool is used. But:

It doesn’t show whether AI improves care. The study can’t tell whether any answer changed a decision, or whether patients did better or worse.

It covers one platform. OpenEvidence is described as the most commonly used AI clinical information platform in the US, but clinicians also use other tools, so this is a partial picture.

It counts questions, not quality. Frequent use doesn’t prove the answers were accurate or that clinicians relied on them appropriately.

Patients can simply ask how their clinician uses AI

There’s nothing patients need to do. But a reasonable step, if you’re curious, is to ask your clinician whether they use AI tools and how they check the answers. Most use them as a reference, much like a textbook or online database.

These tools need testing on the tasks clinicians use them for

AI has already become part of everyday medicine, not in a distant future but now, often through questions doctors type between appointments.

That makes it urgent to test these tools on the tasks clinicians actually use them for, especially treatment decisions, and to understand how their answers shape care.

AI didn’t wait for permission to enter the clinic.

It was invited in, one question at a time.