10 ways AI is creeping into medical care

More to read: Inside AI malpractice law: What policies can reduce physicians' AI liability risk?

The federal government is paying to build an AI that manages patients without a physician in the loop

The contracts went out Sept. 9.

The Advanced Research Projects Agency for Health (ARPA-H) committed $62.7 million over four years, $33.7 million of it in the first, to build an artificial intelligence (AI) agent that monitors patients with heart failure between office visits and acts on what it finds. UpDoc and Tempus AI are among the performers.

Haider Warraich, M.D., the cardiologist running the program, said the work centers on being able to "change or renew or refill existing prescriptions," and on writing new ones.

The Food and Drug Administration (FDA) and the Centers for Medicare & Medicaid Services (CMS) are in the program from day one, writing the regulatory and payment pathway while the software is still a prototype. Johns Hopkins University's Applied Physics Laboratory oversees the program.

Over 39 months, the agent gets benchmarked against cardiologists and tested in trials under an FDA Investigational Device Exemption. No federal health research program has tried to build an autonomous clinical AI for real-world deployment before.

Medicare has let more than 200 companies into pilots that can involve AI, The New York Times reported Sept. 14. Through a program aligned with that effort, the FDA turned down many applicants and cleared four otherwise-unapproved models. Officials inside the administration have discussed paying tech-company AI 60% to 80% of what a physician earns for the same service, the paper reported, citing a person involved. Some officials told the Times the pace has outrun the evidence. Others pointed to Silicon Valley investors who had no real foothold in federal health policy until this administration.

Models built to treat patients on their own "are not ready for prime time," John Whyte, M.D., M.P.H., CEO of the American Medical Association (AMA), told the Times.

Who’s using AI, and what are they using it for?

Eighty-one percent of physicians reported using AI professionally in the AMA's 2026 survey, released in March. In 2023, the number was 38%. Nearly 1,700 physicians responded, averaging 2.3 use cases each.

The most commonly cited use case is summarizing research and checking standards of care, at 39%. Paperwork fills out the rest: 30% draft discharge instructions, care plans or progress notes; 28% handle work billing codes and visit notes, 28% generate chart summaries, 19% let it take the first pass at a patient portal reply. Assistive diagnosis sits at 17% — the bottom of the list.

Among 2,784 U.S. hospitals running Epic, 62.6% had an ambient AI documentation tool by June 2025, according to a study in The American Journal of Managed Care. Three products took more than 80% of those installs. Adjusted adoption was 70.2% at nonprofit hospitals versus 28.8% at for-profits, and 67.6% in the top operating-margin quartile against 58.0% in the bottom.

Freddie Yang and Ilana Graetz, Ph.D., wrote that a rollout tilted that way widens the gap between hospitals that can pay for the tools and hospitals that cannot.

What happened to the endoscopists

Nineteen experienced endoscopists at four centers in Poland used AI polyp detection in routine practice for three months. Researchers then went back and measured what those same endoscopists found on their own, software off. Across 1,443 unassisted colonoscopies, the adenoma detection rate had fallen from roughly 28% to 22%, a 6-percentage-point drop, according to results published in The Lancet Gastroenterology & Hepatology.

The study was observational, the sample was small and trainees were not included. Still, its authors called it the first documented case of de-skilling from clinical AI.

It likely doesn’t help that physicians already assume it’s happening.

Eighty-eight percent told the AMA they have at least some concern about AI-related skill loss, and 70% are worried specifically about the students and residents training now. Asked what would earn their trust, they ranked clear liability rules above everything else.

The economist's case

Dhruv Khullar, M.D., M.P.P., an associate professor of population health sciences at Weill Cornell Medicine, argued in the New England Journal of Medicine on Sept. 12 that AI agents will grow the clinical workforce rather than shrink it.

Radiology is his first exhibit. There are more radiologists in the United States today than a decade ago, with higher average incomes and larger workloads, across exactly the stretch when imaging AI arrived.

Cataract surgery and joint replacement tell a version of the same story: Making the procedures more efficient allows more patients to have them. English economist William Stanley Jevons described this type of efficiency in the 1860s, when steam engines that burned less coal drove coal consumption up.

Khullar's second point is that no fixed quantity of medical work is sitting there waiting to be divided, which is hard to argue in a country where about a quarter of people live in a primary care shortage area.

His third borrows a distinction labor economists make between tasks and skills. AI automates tasks, from visit documentation to electrocardiogram reads. Medicine, on the other hand, runs on the connective work between them — reading a result against a history, sitting with uncertainty, discussing a plan with a patient.

Automating tasks, Khullar wrote, "won't necessarily translate into the automation of jobs." He did not oversell it. Clinical roles will change, he wrote, and some health care jobs will go.

The bioethicist’s case

The opposite case was published in JAMA on Aug. 17.

Ezekiel Emanuel, M.D., Ph.D.; research assistant Abe Baker-Butler; Curai Health CEO Neal Khosla; and Khosla’s father, venture capitalist Vinod Khosla, wrote that autonomous AI will soon beat both unaided physicians and physician-AI teams at "five core cognitive medical tasks," ready for real-world use in some workflows by 2030.

Two of the four authors stand to make money if they are right. Emanuel and Baker-Butler took the argument to STAT on Sept. 9. Whyte responded the same day in the same publication, writing that the goal "should not be to reproduce today's health care system with fewer humans."

Payment moved before the rules did

The FDA's AI-Enabled Medical Device List held 1,524 authorization records as of Aug. 20, about three-fourths of them radiology, and the agency says the list is not comprehensive. None of those clearances came through a pathway built for a chatbot or an agent.

On Aug. 18, the FDA's Digital Health Center of Excellence opened a discussion paper on generative AI devices, docket FDA-2026-N-7874; comments close on Oct. 19.

Michelle Tarver, M.D., Ph.D., who runs the FDA’s Center for Devices and Radiological Health, said patients and clinicians deserve "a regulatory approach that keeps pace with the rapid innovation of digital health technologies."

CMS, however, did not wait. Its Advancing Chronic Care with Effective, Scalable Solutions (ACCESS) Model opened a first performance period on July 1 and runs 10 years, paying participants to manage chronic conditions and releasing the full amount only when patients hit a measurable goal. Technology companies may get a direct route to Medicare payment through it, CMS says.

The agency is running AI in the other direction, too. Since Jan. 1, the Wasteful and Inappropriate Service Reduction (WISeR) Model has paid technology contractors a performance-adjusted cut of whatever their AI-assisted prior authorization reviews save Medicare in Arizona, New Jersey, Ohio, Oklahoma, Texas and Washington.

States wrote most of the actual rules. At least 14 health care AI laws have been passed by midyear, seven restricting how insurers use AI in authorization decisions and five banning AI chatbot therapy, per the Transparency Coalition.

Delaware's HB 191 bars any nonhuman entity from holding a license as a physician, physician assistant or nurse. Utah went the other direction, piloting, within its regulatory sandbox, the autonomous AI renewal of routine prescriptions for patients with chronic conditions.

Through all of this, liability is unresolved

Follow your clinical judgment and the standard of care, and the law protects you when an outcome goes bad. Act on an AI recommendation that contradicts that judgment, and it "exposes you to liability," Richard Anderson, M.D., FACP, CEO of The Doctors Company and TDC Group, told Medical Economics earlier this year. The legal system, he said, will be completely out of step with how medicine is practiced.

Physicians are buying in anyway. According to Doximity's 2026 Physician Compensation Report, two-thirds said staying current with AI would give them an earnings edge over colleagues who do not, and nearly a quarter expected AI to raise their pay within the year. Both numbers assume two things: that a physician stays in the loop and that the value of the time AI saves accrues to the physician rather than to the payer or the vendor.

Khullar's closing point is one to consider: Beliefs about how many clinicians this country will need shape residency slots, payment policy and where research money goes. So, if you believe a scenario is on the horizon, you start behaving in ways that bring it about.