Ambient AI Scribes Show Promise in Reducing Physician Burnout, but Experts Urge Caution on Privacy, Accuracy, and Equity
核心洞察
Ambient AI scribes are increasingly adopted in U.S. healthcare, with 62.6% of health systems using Epic (搜索) EHR now employing the technology, though concerns about privacy, accuracy, and regulation persist.
A JAMA Network Open study found burnout dropped from 51.9% to 38.8% among physicians using an AI scribe, with significant reductions in cognitive task load and after-hours documentation time.
Experts warn that AI scribes may introduce errors into medical records, raise unresolved patient consent and data security questions, and could be used by employers to increase productivity demands rather than improve care quality.
As ambient artificial intelligence scribes increasingly find their way into examination rooms across the United States, a growing body of evidence suggests these tools can meaningfully reduce physician burnout (搜索) and reclaim hours lost to electronic health record documentation. Yet experts caution that the technology arrives with unresolved questions about patient privacy, record accuracy, regulatory oversight, and the potential for employers to exploit efficiency gains.
The prevalence of ambient AI scribes has expanded rapidly. A study published in The American Journal of Managed Care in January found that among 2,784 health systems using the Epic (搜索) electronic health record system, 62.6% also used ambient AI technology. Adoption was more common in systems with higher workloads, in metropolitan versus non-metropolitan hospitals, and at nonprofit rather than for-profit institutions. “These patterns suggest potential for uneven diffusion across hospitals and underscore the need for research on impacts on clinician outcomes, care quality, and equity,” the researchers wrote.
Burnout reductions and time savings
The most compelling evidence for AI scribes comes from studies measuring their effect on physician well-being. In a 2025 paper published in JAMA Network Open, Olson and colleagues surveyed 451 physicians who used an AI scribe in the clinic for 30 days. Among the 272 physicians from multiple specialties who completed both pre- and post-intervention surveys, the proportion reporting burnout decreased significantly from 51.9% to 38.8% (OR = 0.26; 95% CI, 0.13-0.54).
The study also found significant improvements on a 10-point scale in burnout (mean difference, 0.47 points; standard error, 0.12), note-related cognitive task load (mean difference, 2.64 points; standard error, 0.13), ability to provide undivided attention (mean difference, 2.05 points; standard error, 0.18), and time spent documenting after hours (mean difference, 0.90 hours; standard error, 0.19).
A separate study by Lisa Rotenstein, MD, MBA, and colleagues, also published in JAMA Network Open, surveyed 1,430 clinicians at Mass General Brigham and Emory Healthcare participating in an 84-day AI scribe pilot. The proportion of clinicians reporting burnout fell from 52.6% to 30.7% (P < .001), and 32.3% of providers said the new documentation practice positively impacted their well-being, compared with just 1.6% at baseline (P < .001).
“This is an important benefit we are seeing, which has so far been hard to achieve at this magnitude via other types of interventions,” Rotenstein told Healio.
A 2025 study in NEJM AI by Lukac and colleagues assigned 238 outpatient physicians to use one of two clinical AI scribe products — Dragon Ambient eXperience (DAX) Copilot (搜索) or Nabla (搜索) — or usual care. Nabla yielded a decrease in time-in-note of nearly 10% versus controls, while DAX failed to reduce this time compared with controls. Both AI tools were associated with improvements in burnout, task load, and work exhaustion. “These secondary end point findings need confirmation in larger, multicenter trials,” the researchers concluded, adding that occasional inaccuracies of the scribe require vigilance on the part of the clinician.
On a large scale, the Permanente Medical Group (搜索) — the largest medical group in the United States, which rolled out AI scribes across its network in 2023 — reported that implementation gave physicians back more than 15,700 hours previously spent on documentation, equivalent to 1,794 working days.
The promise of restoring the doctor-patient relationship
Beyond time savings, proponents argue that AI scribes can fundamentally reshape the clinical encounter. “Ambient AI has the potential to bring about significant changes in what has become the elephant in the room for many physicians: the electronic health record,” said Leonard H. Calabrese, DO, chief medical editor of Healio Rheumatology and professor of medicine at the Cleveland Clinic Lerner College of Medicine.
Kenneth G. Saag, MD, MSc, professor of medicine at the University of Alabama, Birmingham, noted that the technology could restore the interpersonal nature of medical visits. “It can allow us to look the patient in the eye and return to the interpersonal nature of a doctor visit,” he said. “The hope is that it is not creating more work by requiring a lot of editing.”
Anthony A. Romeo, MD, chief medical editor of Healio | Orthopedics Today, described a similar dynamic: “When you use ambient AI, you are more efficient because it is much better than typing, but, even more important, you have this eye-to-eye contact with your patient throughout the entire evaluation.”
Calabrese suggested AI scribes may offer a “newfound opportunity to exploit all aspects of nonverbal communication,” including being physically closer to patients, more fully observant of their emotions, and freer to use respectful touch. However, he added that he has “seen little organized initiative” to teach or re-teach these skills in the new environment.
The ‘pajama time’ problem
Experts emphasized that the burden of electronic health record work extends well beyond clinic hours. “The bigger problem is the back end, when the portal messages start to flood in at the end of your day,” Saag said, referring to the hours many physicians spend answering patient emails at home as “pajama time.” He called this “the real cause of burnout,” noting that physicians engage in “protracted email correspondences with our patients in an uncompensated way.”
Grace C. Wright, MD, PhD, of Grace C. Wright MD PC, described the transformation personally: “I cannot describe the elation I felt when I tried the AI tool and was suddenly unshackled from my nighttime computer ritual. I was able to just talk with a patient, listen to and see their language, and then document that in my chart note.”
Accuracy, hallucinations, and the risk of error
Despite the enthusiasm, concerns about accuracy remain prominent. Allan Gibofsky, MD, JD, MACR, FACP, FCLM, professor of medicine at Weill Cornell Medicine and attending rheumatologist at the Hospital for Special Surgery, warned that errors in medical records can have lasting consequences. “Medical records are eternal. If there is an error in a patient’s chart, it gets carried forward and attempts to correct it are, at best, problematic,” he said.
Randy Boldyga, founder and CEO of RXNT (搜索), identified background noise as the most common cause of AI hallucinations. “We had one pediatric physician state that the AI scribe was not as effective when there was a screaming baby in the background but was accurate otherwise,” he said. Hallucination rates have been reported at around 1% to 3%, according to a NPJ Digital Medicine study, but Boldyga stressed that “even small mistakes can have large ramifications on patient safety.”
“There is still a place for physicians in this,” Boldyga said. “We should not rely on AI alone; physicians still play a responsible role in making sure that the data being documented is accurate.”
Evalina L. Burger, MD, BmedSc, MBCHB, chair of orthopedics at University of Colorado Anschutz, raised the concern that AI could “jump to a differential diagnosis” by synthesizing symptoms on its own. “I am afraid to get there because then we are going to have a nightmare,” she said.
Regulatory vacuum and legal exposure
Perhaps the most striking gap is the absence of regulatory oversight. “There is no legal obligation that this technology gets reviewed by the FDA if it does not fall under the rules for review set out by Congress and the FDA’s own interpretation of its jurisdiction,” said I. Glenn Cohen, JD, deputy dean and James A. Attwood and Leslie Williams Professor of Law at Harvard Law School.
Gibofsky echoed this concern: “None of this is currently regulated by any governing body like the FDA.”
Cohen noted that the lack of review means hospital systems must conduct their own evaluations for safety, efficacy, and legal and ethical issues — “not a small undertaking.” While large systems may have resources for such review, free-standing or small practices may not.
On the question of medical liability, Cohen explained that the law places responsibility for accurate records squarely on the clinician. In a malpractice case, an accurate record is the most critical piece of evidence, and clinicians will be held liable if incorrect information in the record led to an adverse event.
Cohen also flagged the issue of “shadow” records — duplicate or unofficial parallel medical records that attorneys have consistently urged clinicians to avoid. Since AI scribes may create both an audio recording and a transcription, Cohen recommends health systems adopt policies stating that the transcribed note is in draft format and not considered a legal part of the patient record until reviewed and signed by the clinician.
Patient consent and privacy
Patient consent presents another legal and ethical challenge. “If I am doing a telehealth visit and using AI capture technology, I have to check the consent laws in the state where the patient is physically located,” Gibofsky said, noting that some states require single-party consent while others require both parties to consent. “This is not currently appreciated by everyone using ambient AI.”
Cohen advised that health systems enact policies ensuring providers always disclose the use of ambient listening, are ready to answer patient questions, and present patients with the opportunity to decline. Clinicians should also be prepared to not rely on the software when patients need to discuss sensitive subject matter.
Romeo added that privacy protections must ensure AI systems do not feed data into public large language models. “Your system has to be protected from going to the large language models into the public sphere,” he said.
Productivity pressures and equity concerns
Several experts warned that the time-saving benefits of AI scribes could be redirected toward increasing patient volume rather than improving care quality. “Time is money, and if AI allows you to see patients more efficiently in less time, the patient load may be increased,” Wright said. “I would prefer to have more issues dealt with in the clinical encounters, such as mental health, life impact and social connectedness, instead of just adding more patients.”
Gibofsky raised a related concern about billing. “What our employers may do is comb through the AI-scriven notes to find the highest level of billing possible,” he said. He cautioned that some patients may be billed for diagnoses that an AI scribe recorded but that the physician does not necessarily need to treat, potentially discouraging patients from seeking care.
Saag highlighted the risk that charging for portal messages could create disparities: “If you charge too much for this, it disincentivizes people with lower economic means to communicate with their provider. This will create more disparities in care.”
A friend or a guarded ally
Despite the concerns, many experts remain optimistic. Wright said she has “come to regard it as a friend in the exam room, especially when it helps me scribe my notes, or take an accurate deep data dive.” Gibofsky, however, struck a more cautious tone: “I share the enthusiasm for having it available to use, but for the time being I will sleep with one eye open knowing that at the end of the day, I am fully responsible for the accuracy of the note and the decisions made for my patients.”
Cohen placed ambient AI scribes in perspective relative to other medical AI applications. “Compared with other AI in medicine implementations we are seeing, ambient listening is on the lower risk for legal, ethical and reputation concerns,” he said. “Nevertheless, clinicians should do all the due diligence on the front end before implementation. There is also an obligation to perform ongoing monitoring and auditing for things like bias and general efficacy.”
As Boldyga summarized: “There is a really good phrase in our industry: ‘Be cutting edge, do not be bleeding edge.’ It is good to have advanced technology but also make sure that it is something physicians can count on.”