AI scribes fail to note patient experiences, U.K. review finds

Researchers at the University of Edinburgh searched three databases on March 20 for work on how ambient artificial intelligence (AI) scribes are being implemented. They screened 1,333 results and were left with 27 articles. Thirteen were empirical studies.

The narrative review, published Sept. 3 in BMJ Digital Health & AI, mapped those articles against the Non-Adoption, Abandonment, Scale-Up, Spread and Sustainability (NASSS) framework, a model for evaluating whether a health technology survives contact with the organization that buys it.

Research on ambient scribes has concentrated on accuracy, documentation burden and time saved, the authors write. How the tools change clinical work, patient disclosure and the function of the note itself has drawn far less attention.

"Many clinicians are excited about ambient AI scribes, because they promise to cut down on paperwork," Lucas Seuren, a research fellow at the university's Centre for Biomedicine, Self and Society and the review's lead author, said in a statement. "But the experiences of patients are poorly considered, and there are real risks that the patients' stories are lost. This can further disadvantage people who already face marginalisation in health and social care services."

Seuren wrote the review with Robin Williams and Kathrin Cresswell, both also at the University of Edinburgh. Their evidence base is largely British.

In a survey of 598 U.K. general practitioners published this year in npj Digital Medicine, 40% said they were using AI scribes and another 23% said they had used them before. Among users, the tools sat in on a mean of 60% of consultations.

The rollout the Edinburgh authors describe is driven in part by political and economic pressure on health systems, including long waiting lists and growing costs, for which they cite England's 10-year health plan.

At Kaiser Permanente, 7,260 physicians used AI scribes across more than 2.5 million patient encounters in the 14 months ending in December 2024. At the University of California, San Francisco (UCSF), roughly 70% of physicians use one, and about 90% of those physicians say they love it.

"We've almost gotten to the point where, if we turned it off, we might lose a fair number of doctors," Robert M. Wachter, M.D., chair of the department of medicine at UCSF and author of "A Giant Leap: How AI Is Transforming Healthcare and What That Means for Our Future," said in a March interview. "It's almost become an expectation of practice now."

Many ambient scribes run on proprietary large language models from large U.S.-based technology companies, the review notes. Training data and biases are opaque, and the models may carry assumptions rooted in the U.S. health care system, including insurance-based documentation practices.

The tools were built around the paperwork American practices already do. What the review says about the note itself — what it captures, and how patients behave when a recording is running — does not turn on who pays for the visit.

A rapid review published in JMIR AI in October 2025 by a separate team screened more than 1,400 studies on ambient scribes and found six that met rigorous criteria for real-world evidence.

A general-purpose tool doing specialty-specific work

Ambient scribes construct the patient note uniformly across health care domains, treating a cardiology note the same as a primary care note. Performance varies widely across clinical, organizational and health system settings.

Deployment into new settings without validation may not deliver the expected benefit and can disrupt workflows, impair decision-making and worsen health inequities, the authors write.

"Ideally, you would want one that documents quite differently for a cardiologist than it does for a primary care doctor," Wachter said. "Right now, they do a little of that, but not perfectly." He expects that to improve as vendors accumulate data and reps.

The 30% of UCSF physicians who have passed on the technology are not all holdouts, he said. In primary care, where the history is long and the intake broad, the tools earn their keep. In specialties working from a set series of questions and already checkbox-driven documentation, less so.

Some physicians tried a scribe, caught errors and decided they could not trust it. Others balked at what it asks of the physical exam, which the technology cannot observe.

"You have to narrate it, which is a little weird if you have never done that before," Wachter said.

From author to editor

Ambient scribes generate longer notes than clinicians do, which can create more follow-up work, the review found. Physicians remain responsible for what the note says, which turns them from the authors of clinical documentation into its editors.

That shift carries implications for professional practice and training, the authors write, and over time may cost junior clinicians opportunities to build reasoning skills. Anecdotal reports in the literature describe clinicians who do not recognize their own notes or remember a patient at the next visit.

Omissions were the most frequent error in a real-world evaluation published this year in JMIR Medical Informatics, turning up in 18% of the notes physicians reviewed. Hallucinations appeared in 11.5% and accidental inclusions in 9.3%. Of 356 notes sampled from the 7,545 that 31 physicians generated with an ambient scribe, 94.7% were free of significant errors, and the authors concluded that clinician review of every note remains essential.

That pilot ran in July and August 2024. Research is struggling to keep up with how quickly the tools are being adopted, the review's authors write, and setting a research agenda for a field moving this fast is difficult for the same reason.

"If I've used the scribe in the last 49 notes and they were perfect, am I really going to be fully attentive as I review note No. 50?" Wachter said. "If I'm human, the answer is no."

At UCSF, no technical mechanism confirms that a physician has read a note before signing it.

"We are not that worried here about deskilling," Wachter said. "It is hard for me to imagine I am going to lose my ability to read over a note and see whether it is accurate."

Typing a note used to prompt him mid-sentence to remember what he had not asked. "There's something in the process of writing that is probably more cognitively active than reading over a draft," he said.

One design he pointed to is a confidence signal on the note itself: green when the system is certain, yellow when it could not make out a word or found an internal conflict, so the physician knows which drafts need a closer read.

What the note leaves out

The note as a record strips the consultation of its social dimensions, the review found, leaving out the patient's narrative, values, illness experience and concerns. The chit-chat at the start of a visit that can be highly informative to a clinician is removed.

Gestures, facial expressions, movement and emotional affect go uncaptured, and silence, which can signal hesitation or discomfort, does not appear in the transcript at all. Non-lexical sounds such as "um," "uh-huh" and "hmm" carry information about uncertainty and emotional response, one cited study found, and speech-to-text systems often mistranscribe or omit them.

Those gaps have safety consequences, the authors write. Transcription errors and an incomplete record of the consultation mean red flags such as medications or symptoms can be missed or misrepresented.

When a consultation is being recorded and processed by AI, patients may not feel comfortable raising substance use, domestic violence or mental health, the review found. Clinicians may need to offer patients the option of turning the scribe off, which requires being able to take notes manually.

Consent procedures are largely ad hoc, with no clear guidance on what patients want and need to know or how they can best be informed, the authors write.

Fewer than 10% of patients in a randomized trial at the University of California, Los Angeles declined to have a scribe used during their visit. Roughly two-thirds of surveyed Kaiser Permanente patients said they were comfortable with the technology. UCSF asks for consent, Wachter said, and some patients remain "a tiny bit creeped out by it."

The return is not in the minutes

"We thought it would save 5 or 10 minutes a visit. It turns out, it doesn't really save that much time," Wachter said. Some of what is recovered goes back into the visit rather than the schedule: "Some of that time is repurposed into actual, genuine human contact, which feels better for the doctor and feels better for the patient."

Adopters saved 16 minutes of documentation time and 13.4 minutes of total electronic health record time per eight scheduled patient hours, and added 0.49 visits a week, according to an April 2026 JAMA study of 8,581 ambulatory clinicians at five academic health systems.

Time in the record outside scheduled hours did not change significantly.

The study, led by Lisa S. Rotenstein, M.D., MBA, M.Sc., of UCSF, tracked 1,809 clinicians who adopted a scribe against 6,772 who did not at Mass General Brigham, Emory Healthcare, UCSF, Yale New Haven Health and UC Davis from June 2023 through August 2025, across Ambience, Nuance DAX Copilot and Abridge.

Physicians who adopted a scribe generated 1.81 additional relative value units and saw roughly one more patient per week, with no increase in claim denials, according to a January 2026 JAMA Network Open study led by A Jay Holmgren, Ph.D., M.H.I., of UCSF. The analysis covered more than 1.2 million ambulatory encounters across 1,565 physicians.

That works out to about $3,044 in additional annual revenue per physician under the 2025 Medicare Physician Fee Schedule, roughly the cost of the tool.

Wachter, writing with Rotenstein and David W. Bates, M.D., M.Sc., in a February 2026 JAMA Internal Medicine viewpoint, argued that conventional return-on-investment math misses malpractice risk, patient retention and the cost of replacing physicians who burn out and leave. The American Medical Association puts that replacement cost at $500,000 to upward of $1 million for a single primary care physician.

"The AI scribe probably kind of pays for itself just in pure throughput and time savings," Wachter said. "But the real benefit is the joy in practice, recruitment, retention."

"If I save five minutes, do I turn that into five more minutes on the hamster wheel?" Wachter said. Splitting the recovered time between the schedule and the physician is a management call, not a technical one.

What has not been studied

The review sets out a research agenda that is largely unanswered: what the actual risk profile of ambient scribes is compared with human error and existing documentation workflows, what AI literacy clinicians need to use them safely, which patients are excluded or disadvantaged, what governance fits a clinical tool that keeps changing, and how risks shift as products extend from summarization toward decision support.

Some states have moved without waiting. Texas Senate Bill 1188, in effect since Sept. 1, 2025, permits health care practitioners to use AI for diagnostic purposes only if they review all AI-generated records under Texas Medical Board standards and disclose the use of AI to patients. The statute is written around diagnostic use, though some health care attorneys read it to reach any note produced with AI assistance.

The authors flag their own limits. Their analysis of risks rests largely on editorials and perspectives rather than empirical work, they did not screen citations forward or backward, and they searched neither grey literature nor preprint archives.

Video consultations offer the closest precedent, the authors write. Policymakers initially treated them as a general-purpose substitute for in-person care before it became clear that their value varied sharply by setting and that patients and clinicians had to adapt how they communicated.

Video visits came to be understood as a distinct form of consultation with their own limits rather than a swap for the old one, and ambient scribes should be approached the same way. That includes, the authors suggest, asking whether the patient note itself should be redesigned around what clinicians and patients actually need from it.