The AI Scribe Promise Has Two Audiences

Ambient AI scribes have one of the cleanest promises in health technology: let the clinician look at the patient instead of the screen while software drafts the note.

The category is often described in exactly those terms. Microsoft says Dragon Copilot enables clinicians to focus on the patient. Physicians who have used ambient scribes report more eye contact, less cognitive strain and less work spilling into their evenings. In a 2026 survey published in npj Digital Medicine, 40% of 598 UK general practitioners said they were currently using an AI scribe, and another 23% had used one in the past.

Clinical documentation is a real burden, and a tool that gives physicians more attention for the person in front of them could improve both the work and the encounter.

But that value proposition has two audiences.

The clinician or health system adopts the tool to reduce documentation work. The patient is asked to accept it because it may allow for a more attentive conversation. The first benefit is usually measured in time, workload and clinician satisfaction. The second is often treated as an intuitive consequence of the first. In communications terms, the category has borrowed a patient benefit from a clinician outcome.

In my work with independent physicians, I spend a lot of time looking at the point where a product promise becomes a patient conversation. Ambient scribes make that translation unusually visible. A tool purchased to solve a documentation problem enters the exam room as a recording and summarization system, then helps determine how the visit will be represented in the medical record.

The product enters the exam room as workflow software. Patients experience it as part of the conversation.

What the new review actually found

On September 3, researchers from the University of Edinburgh published a review of ambient AI scribes in BMJ Digital Health & AI. Lucas Seuren, Robin Williams and Kathrin Cresswell searched Embase, PubMed and Scopus, included 27 articles and examined the findings through a framework designed to study how health technologies are adopted, scaled and sustained. Thirteen of the included articles were empirical studies.

Their central finding was not that ambient scribes have been proved harmful. It was that the evidence is concentrated in a much narrower area than the category’s broader promises.

Existing research has focused largely on whether the tools reduce administrative burden, save time and produce usable notes. The authors found few empirical studies examining potential risks and even fewer studying implementation and adoption in practice. They argue that ambient scribes do more than automate note-taking: they can shape clinical workflows, patient-clinician interactions and the nature and function of the clinical note.

A review like this can identify gaps, plausible risks and questions that deserve study. It cannot tell us that patients routinely disclose less, that their experiences are systematically removed from notes or that AI-supported visits are worse. The evidence is not yet strong enough to support those conclusions.

Adoption is moving faster than our ability to answer some of the questions used to justify it.

Patient acceptance is not the same as patient benefit

Many patients appear open to the technology.

Before UC Davis Health introduced an AI scribe program, it surveyed 1,893 primary care patients about the technology. In the study published in JMIR Medical Informatics, 48% expected a positive effect on the visit, 33% were neutral and 19% expected a negative one. Patients could see the appeal: among comments describing potential benefits, improved human interaction was the most common theme.

Their concerns were also specific. Accuracy was the most common theme among comments about potential problems, followed by privacy and discomfort with being recorded. Patients placed importance on being informed and asked for permission before the tool was used.

That survey asked people to imagine an implementation. A newer Stanford study in JAMIA Open asked 2,202 patients after outpatient visits in which an ambient scribe had actually been used. Seventy percent found it helpful, and nearly 74% wanted it used in a future visit.

Patient acceptance is not evidence that communication improved.

A separate 2026 study in the Journal of the American Medical Informatics Association compared patient survey results from 31 physicians using an AI scribe with results from 125 control physicians. Across 31,917 surveys, the researchers found no statistically significant difference in the pre-to-post change in communication or satisfaction-with-physician ratings between the two groups. They did not find evidence that the scribes improved the patient experience, but they did not find evidence that they worsened it either.

Acceptance, preference and demonstrated benefit are three different outcomes. A patient can be comfortable with an AI scribe, want the clinician to keep using it and still report no measurable change in the visit.

The room and the record are different communication problems

The most persuasive case for ambient scribes is what happens in the room. If a clinician is less distracted by typing, the patient may receive more eye contact, more visible attention and a conversation that feels less interrupted.

The note has a different job. It carries a selective account of the visit forward to other clinicians, staff, the patient and, depending on the setting, payers and auditors. It is not a transcript. It organizes, compresses and translates what happened into a form the healthcare system can use.

Every clinical note already involves selection. Human-written notes can omit context, introduce errors and reduce a complicated story to standard clinical language. The relevant comparison is not an AI-generated note against a perfect human record. It is the AI-assisted documentation process against the process it replaces, in a particular specialty, practice and patient population.

What patterns does the technology introduce or reinforce?

An ambient scribe can capture the words spoken in the room, but a useful note requires the system to decide what is relevant, where it belongs and how much detail to preserve. Facial expression, hesitation, tone and gesture may matter to the clinician without appearing in the generated note. A patient’s account may be reorganized around clinical categories that make the record more efficient while making the experience less recognizable to the person who lived it.

Current evidence does not tell us how often that happens or whether clinician review reliably corrects it. “Accuracy” is too small a word for the communication task.

In a qualitative Stanford study of 22 physicians, clinicians were mostly positive about the effect of ambient scribes on their engagement with patients. At the same time, their views on note accuracy and style were largely negative, and they described problems with completeness, length and editing. Several also saw limited usefulness when working with patients who did not speak English.

A note can contain no invented facts and still fail to represent what mattered most. It can be comprehensive and still bury the point. It can sound clinically polished while leaving the patient unable to recognize their own account when they read it later.

Note quality is a communication question as much as a technical one: good for whom, for what purpose and at which point in the care journey?

Consent is part of the product experience

Patient communication about ambient scribes is often treated as a compliance step before the real benefit begins. From a health communications perspective, it is part of the intervention itself.

The patient needs enough information to understand what is changing in the encounter. Is the tool recording audio or processing it in real time? Is any audio retained? Who receives the data? What does the system create? Will the clinician review the draft before it enters the chart? Can the patient decline or ask for the tool to be turned off?

The answers vary by product, organization and jurisdiction. A vague request such as “Is it okay if I use an AI scribe?” asks for agreement before it creates understanding.

A plain-language explanation could follow this structure:

With your permission, I would like to use an AI note-taking tool during our visit. It [state exactly what the tool captures and whether anything is stored] and creates a draft of my note. I review and edit that draft before it becomes part of your medical record. You can say no or ask me to turn it off at any time, and that will not affect your care. What questions do you have before we begin?

The structure is a communication model, not a universal legal consent script. The wording should be reviewed against the actual product, data practices, workflow and applicable requirements. A practice should not promise immediate deletion, clinician review or an easy opt-out unless its operations reliably support those claims.

Patients should hear about the technology before the recording starts. The UC Davis survey found that they preferred to learn about it before or near the beginning of the visit. A short explanation during scheduling or check-in can give patients time to form a question. The conversation in the exam room can then confirm a real choice rather than introduce an unfamiliar system when the patient is already focused on why they came.

Measure the promise being made

If an organization adopts an ambient scribe to reduce documentation burden, it should measure documentation time, after-hours work, editing burden and clinician experience.

If it tells patients the same tool will improve their visit, it should measure that promise directly.

A patient-centered evaluation should ask more than whether patients liked the technology. It should distinguish at least four questions:

• The encounter: Did patients feel heard, understand the conversation and feel able to say what they needed to say?
• The explanation: Did they understand what the tool was doing, what would happen to their information and that they had a meaningful choice?
• The record: Did the final note accurately and recognizably reflect the concerns, context and plan that mattered to the visit?
• The distribution: Did those results differ by language, age, race, disability, visit type or other factors that could make a single overall satisfaction score misleading?

Clinician impressions remain valuable, but they are not a substitute for asking patients. Opt-out rates are useful, but a low refusal rate does not prove informed comfort. Note accuracy matters, but a technically accurate summary does not answer whether the patient’s perspective survived the translation.

Ambient scribes may prove to be one of health technology’s genuinely useful applications. They address a real problem, and early evidence suggests that many clinicians and patients welcome them. Early acceptance makes the broader promise worth studying; it does not establish that the promise has already been delivered.

Giving a clinician more attention for the patient and creating a record that represents the patient well are separate outcomes. A responsible implementation should be able to explain and measure both.

The question is not only whether an AI scribe can hear the visit. It is what the organization adopting it can say, with evidence, about what the patient experiences and what the record remembers.

---