American physicians went from 10 percent using AI several times a day to 38 percent in twelve months. In the same survey, 74 percent of clinicians said they are worried the tools are eroding the very skills they would need to catch the tools' mistakes. Those two numbers come from the same 355 people, and they are the cleanest summary yet of where medical AI stands in September 2026: adopted faster than it has been validated, and trusted less the more it is used.
The Financial Times reported on Saturday that clinicians are now actively resisting the expansion of AI beyond diagnostics and imaging, citing thin clinical and performance data for the newer categories vendors are selling. Sarah Neville's piece frames the objection narrowly: not a rejection of AI in medicine, but a demand for the evidence that has accompanied radiology tools and has largely not accompanied treatment recommendation engines, documentation assistants, and patient-facing chatbots.
The backdrop is Wolters Kluwer Health's 2026 Future Ready Healthcare survey, fielded by Ipsos between March 11 and 14 among 203 doctors and 152 nurses in the United States, plus 254 patients. Nearly three-quarters of doctors and 70 percent of nurses now use AI at least weekly at work, up from 38 percent and 46 percent a year earlier. Repeated daily use tripled among physicians, from 10 percent to 38 percent, and doubled among nurses, from 16 percent to 32 percent. More than half of doctors use AI to summarise literature or analyse data, and 44 percent use an ambient scribe.
Then the worry list. Deskilling, which the report defines as overreliance on AI that reduces clinicians' ability to independently identify inaccuracies or poor recommendations, was cited by 74 percent. Hallucinations, including fabricated studies, were cited by 74 percent. Advertiser-driven bias from drug or device makers was cited by 72 percent of clinicians and 61 percent of patients, and 75 percent of patients worry about accountability if AI contributes to harm. Awareness of a formal AI governance policy at the respondent's own organisation rose from 21 percent to 27 percent, a six-point gain in a year when usage doubled or tripled.
"The pressure is on healthcare leaders now to close the trust gap with visible, organizational governance and trusted content that tackles these worries, while continuing to drive innovative new clinical solutions," Greg Samios, chief executive of Wolters Kluwer Health, said in the release. Wolters Kluwer sells UpToDate, an AI-backed clinical decision support product, so it has a commercial interest in "trusted content" being the answer.
Peter Bonis, the company's chief medical officer, was blunter with Healthcare Dive about accountability. "I think this is all in flight. People are wrestling with this. It's not clear who is going to be responsible for this profound set of issues that can affect the actual delivery of care and who actually takes the risk related to that," he said. On the usage surge, Bonis cited "increased exposure, increased familiarity," adding: "But really importantly, it's addressing an unmet need."
Why It Matters
The deskilling fear is not hypothetical, and the best evidence for it comes from the one domain clinicians say they trust. A multicentre observational study in The Lancet Gastroenterology and Hepatology in August 2025 tracked 19 experienced endoscopists across four centres. Their unassisted adenoma detection rate fell from 28.4 percent before routine AI exposure to 22.4 percent after it, a six-point absolute decline in the metric most tightly linked to preventing colon cancer. That is the first real-world documentation of clinical deskilling from AI, and it happened in imaging-adjacent work. If the effect shows up where AI is best validated, the burden of proof for treatment recommendations should be higher, not lower.
The evidence base for the fastest-growing category, ambient scribes, is thinner than the marketing suggests. The largest randomised trial to date, published in NEJM AI, assigned 238 outpatient physicians across 14 specialties to Microsoft's DAX Copilot, Nabla, or usual care over two months in late 2024. Both tools showed potential improvements in burnout and task load, but the authors flagged those as secondary endpoints needing larger multicentre confirmation, and physicians turned the scribes on in only about a third of eligible visits. That is a promising pilot, not a mandate, for a product category 44 percent of surveyed doctors already use.
The survey's own internal contradiction deserves attention. Seventy-three percent of clinicians said they are confident they could spot a clinically invalid AI answer without consulting another source. Bonis told Healthcare Dive the remaining quarter is probably an undercount, because an AI can cite one accurate study while omitting the ones that point the other way. Seventy-seven percent said they double-check AI answers against PubMed or UpToDate, which is reassuring until you recall that the same population uses AI multiple times a day. Verification at that frequency either consumes the time savings or quietly stops happening.
Governance is the softest number. Even among the 27 percent who knew of a policy, only 22 percent said their employer had defined the respective responsibilities of clinicians and AI products. Health systems argue that usage runs ahead of policy because the tools fill real gaps, and that demanding trials for every workflow tool would leave clinicians drowning in paperwork. That has merit for scribes. It is weaker for anything that suggests a treatment.
What to Watch
The next inflection points are empirical. The Abridge randomised trial and registered emergency-medicine scribe trials should confirm or deflate the NEJM AI burnout signal. Replications of the Lancet colonoscopy finding in radiology or dermatology would settle whether deskilling is a quirk or a general property of assistive AI. And the 2027 edition of this survey will show whether governance awareness can climb faster than six points a year. If clinicians are still 74 percent worried and 27 percent informed next year, the FT's sources will have been right that the wall is about evidence, and vendors will have chosen not to climb it.
"I think this is all in flight. People are wrestling with this. It's not clear who is going to be responsible for this profound set of issues that can affect the actual delivery of care and who actually takes the risk related to that."— Peter Bonis, Chief Medical Officer, Wolters Kluwer Health