Last May, Dr. Anupam Jena, a Harvard physician and economist who studies how doctors actually behave rather than how they say they behave, gave NBC News a three-word verdict on an AI tool most patients had never heard of. “Everyone is using it,” he told the network. On October 15 he and his colleagues put a number on “everyone.” Two weeks later the Pew Research Center published what the patients think, and mostly they think they should have been told.
Pew’s report, released October 29, drew on 3,488 U.S. adults surveyed between June 22 and 28. Of those, 60 percent said they are uncomfortable with AI in their own care for tasks like diagnosis and treatment, against 30 percent who are comfortable. That isn’t a blanket refusal. Asked about AI making a diagnosis, only 3 percent would let it work alone, 47 percent would accept it if a human did the final review, and 41 percent were not open to it at all. What people want is a person answerable at the end of the chain, and some way of knowing the chain exists.
| Measure | Share |
|---|---|
| of U.S. adults | 60 percent |
| Age group | 2022 | 2026 |
|---|---|---|
| Under 30 | 29% | 45% |
| 65 and older | 33% | 26% |
The overall verdict on whether AI will help patients has barely moved since Pew last asked in 2022, with 37 percent now saying better and 36 percent worse, but the generations have traded places underneath that flat line. Among adults under 30, the share expecting worse outcomes rose from 29 percent in 2022 to 45 percent this year, the most pessimistic of any age group, while among those 65 and older it fell from 33 percent to 26 percent. The patients with the most decades of AI-assisted medicine ahead of them have become its sharpest skeptics.
The study Jena co-authored, published in JAMA on October 15, is a cross-sectional look at OpenEvidence, an AI clinical search platform that clinicians unlock by verifying a National Provider Identifier. Between January 1, 2024, and June 30, 2026, 952,008 American clinicians, 556,200 of them physicians, submitted about 271 million queries, and monthly active users climbed from 355 to 703,265 over the same window. The average active user’s monthly queries more than doubled, from 18.0 to 39.5.
What were they asking? The authors classified 51.9 percent of queries as clinical decision-making, with 33.1 percent about treatment and management and 13.9 percent about diagnosis, which is the same ground on which Pew’s respondents said they were uncomfortable with AI in their own care. TechTarget’s account of the study adds that about 70 percent of surgeons’ queries and 73 percent of anesthesiologists’ fell outside their own specialties, that clinicians early in their careers used the platform hardest, and that oncologists and hematologists led every specialty. First author Vishal Patel, a surgical fellow at Brigham and Women’s Hospital, said AI tools had entered routine clinical practice “faster than almost any technology we’ve seen.”
The paper’s construction deserves as much attention as that quote. OpenEvidence supplied the de-identified data. Three of the authors, Samuel Finlayson, Mondira Ray and Travis Zack, list OpenEvidence as their affiliation, and TechTarget describes them as company employees. A large language model sorted the clinical purpose and topic of a stratified sample of 38.8 million queries. TechTarget reports that the company did not fund the work, but the count still comes from the company’s own logs, was co-written by its staff and was sorted by another AI. It measures use. It says nothing about whether any patient fared better or worse.
Accuracy is a separate question, and the answers so far aren’t comforting. NBC reported on a December medRxiv preprint, not yet peer reviewed, which found the tool “accurately answered more complex medical questions less than 45% of the time.” Dr. Hannah Galvin, a pediatrician at Cambridge Health Alliance, told the network, “we have concerns over how it is being used to deliver clinical care.” MaineHealth asks its doctors to “refrain from entering PHI into OpenEvidence.” In practice the health system keeps its patients’ protected health information out of a tool its clinicians remain free to consult about those same patients.
Who pays for the answer
Patients don’t see the business model. OpenEvidence is free to verified clinicians, and NBC reported that it is funded by ads on its app and website, some of them from pharmaceutical and medical device companies. The market-research firm Sacra estimates the company’s annualized revenue at $300 million as of July 2026, up from $150 million at the end of 2025 and $7.9 million at the end of 2024, and records a $250 million raise in September 2026 at a $15 billion valuation. Those are outside estimates, not audited financials, but they still describe nearly fortyfold growth in about a year and a half.
An independent audit is the usual way to find out whether that advertising touches any answer, and producing one is the job of the company whose product hundreds of thousands of clinicians now consult about treatment. Elsewhere in medicine the question would be routine. When drug and device companies pay a physician, the payment is reported to the federal Open Payments program, whose database anyone can search. When the same companies buy ads on a platform physicians consult, the money goes to the platform, and Open Payments was built to track the physician.
Forty-six percent aren’t sure
Pew’s August companion report, drawn from the same June fieldwork, shows how little the public knows about any of this. Only 16 percent of adults said a provider had used AI in their care, and 46 percent said they weren’t sure. Asked how much it mattered to be told, 72 percent called it extremely or very important, and 81 percent said they should be told when AI helps make a diagnosis. More than half, 53 percent, said they have little or no say over whether AI is used in their care, and 63 percent want more.
The tool knows exactly which licensed clinician is asking. Nearly half of patients don’t know whether anyone asked it about them.
Pew’s own headline is that Americans are open to AI as long as a human makes the final call, so their distrust lands less on the software than on the institutions holding it. Asked whom they had a great deal or quite a bit of confidence in to make sure AI serves patients’ interests, 26 percent named doctors, 16 percent AI companies, and 11 percent each the federal government and health insurers. Half, 51 percent, expect AI to make patients’ health information less secure, and 47 percent expect it to damage the relationship with their doctor. These are people who have watched the same institutions handle earlier technology and see no reason to expect different behavior this time.
| Group | Value (percent) |
|---|---|
| Doctors | 26 |
| AI companies | 16 |
| Federal government | 11 |
| Health insurers | 11 |
OpenEvidence’s terms of service, NBC reported, say its services are “in no way intended to serve as a diagnostic service or platform” and are not meant to “substitute for the clinical judgment of a qualified healthcare professional.” The company’s own data, published in JAMA with three of its employees on the byline, show that about one query in seven concerned diagnosis. The disclaimer was written for clinicians and the study for the journal’s readers. Neither was written for the patient.
Sources
- Pew Research Center: Americans Are Open to AI in Healthcare, but Want Humans Making Final Calls (October 29, 2026)
- Pew Research Center: Americans Want Transparency When AI Is Used in Their Healthcare (August 25, 2026)
- JAMA: Patel, Jena et al., Clinician Use of an Artificial Intelligence–Enabled Clinical Information Platform (2026)
- TechTarget: AI-based clinical intelligence is all the rage. But who is using it? (October 2026)
- NBC News: Most U.S. doctors are quietly using this AI tool. Few patients know about it (May 2026)
- Sacra: OpenEvidence revenue, valuation and funding estimates
- CMS: Open Payments