Last Tuesday, a middle-aged man in my clinic paused halfway through describing his headaches, glanced at the screen, and said, “Are you listening, or is the computer?” He was not being rude. He was tired, and I understood why. The room had become a small theater of attention, with me at the keyboard, the chart waiting open, and a vendor demo in my head promising that the machine would now help me “focus on the patient.”
I have heard some version of that moment before. A nurse once asked whether I was “done with the note” before I had finished the exam. A resident, years earlier, told me she felt like she was practicing medicine inside a spreadsheet. A colleague recently said after an AI scribes pilot, “The words are cleaner, but I’m not sure I recognize myself in them.” That sentence stayed with me.
I used to think the core problem was documentation burden, full stop. Then I watched an AI system produce a tidy note that captured the medical facts, the chronology, and even my phrasing, while still missing the thing that made the encounter clinically real: the patient’s fear that the headache was a tumor because his brother had died young. Now I think the deeper issue is something I call the borrowed voice problem. The machine can borrow my style. It cannot borrow my responsibility.
The chart can improve while the doctor recedes
Ambient AI and related documentation tools are often described as a relief valve, and sometimes they are. In a 2024 study of ambient voice technology in primary care, documentation time per patient fell by 28.8%, and time outside working hours fell by 11.8%. A randomized study of AI scribe tools reported a 9.5% decrease in time-in-note for one product, while another showed no significant change. Those numbers matter, because they explain why clinicians keep trying these tools. The burden is real.
But numbers do not settle the question. They only narrow it. In the documentation literature, Melnick et al. in Mayo Clinic Proceedings (2018) found that the average system usability score for electronic health record usability was 45.9, far below the conventional threshold for acceptable usability. That paper helped normalize something many of us already knew in our bones, the EHR was not just inconvenient, it was actively shaping how we practiced. AI scribes enter that same ecosystem, and they do so with an odd promise: they may reduce friction while leaving the architecture intact.
That is where the borrowed voice problem begins. I am not worried only about whether the note sounds polished. I am worried about whether it quietly becomes the authoritative version of a visit, even when it smooths over uncertainty, compression, and human texture. I have seen AI-generated summaries miss an important social cue, flatten a nuanced differential, and make a tentative impression sound more settled than it was. A good note should preserve judgment. A decorative note only preserves prose.
Professional identity is a clinical variable
The obvious reading is that AI in medicine is a workflow issue. The clinic teaches otherwise. Identity changes behavior, and behavior changes care. When physicians begin to feel that their cognitive labor is being outsourced, the effect reaches beyond satisfaction scores. It alters how we teach trainees, how we narrate uncertainty, and how much ownership we take when the machine is wrong.
I have felt that tension in myself. During one outpatient session, I caught an AI draft that described a patient as “reassured” after a normal exam. He was not reassured. He was scared, and he asked three times whether we had truly ruled out cancer. The draft was clinically plausible and emotionally false. I corrected it, then wondered how many such false comforts slip through when the pace is faster and the day is longer. That is a vulnerability I do not mind naming: I am not always certain where efficiency ends and clinical amnesia begins.
The literature is starting to catch up with what clinicians feel. A study of health professional students’ use of generative AI during clinical placements found that 52.6% reported using GenAI in that setting. In another survey, clinicians and trainees described tension between AI adoption and professional formation. I read those findings as a warning. If nearly half of the next generation is already normalizing AI during placements, then the question is not whether they will use it. The question is what kind of doctor they will believe themselves to be while using it.
This is where medical humanities still matter. They give language to the parts of practice that are not reducible to throughput. They remind us that medicine is a moral craft as much as a technical one. I do not mean that sentimentally. I mean it operationally. If we remove the parts of clinical work that train judgment, then we should expect judgment to atrophy. Tools can assist expertise. They can also thin it.
The counterintuitive lesson from AI is restraint
I reject the idea that every improved metric deserves deployment. I would not place an AI system into a clinic simply because it shortens notes, reduces clicks, or impresses a vendor slide deck. I would not allow it to draft final diagnostic language without a physician review pathway. And I would not use it to create the illusion that empathy has been automated. When that happens, the clinician becomes the audience for a machine performance of care.
That sounds harsh, but it is a practical judgment. The most useful AI tools in medicine will probably be the ones that make room for clinical attention, not the ones that mimic it most convincingly. The strongest evidence so far points in that direction. Ambient scribes may help with time, task load, and burnout. They do not remove the need for human interpretation. They do not resolve the ethical problem of authorship. They do not answer whether the note belongs to the person who spoke, the system that assembled it, or the doctor who signed it.
I think the profession is entering a phase of selective subtraction. We should ask which parts of documentation can be delegated without damaging the discipline of seeing, listening, and deciding. A note should be an artifact of thought, not a substitute for it. That distinction matters more now because AI can make a competent surface so quickly that we are tempted to confuse polish with understanding.
There is a reason I keep returning to that Tuesday visit. The patient did not care about usability scores or token counts. He cared whether I could hold onto his story while the machine did its work. By the end of the visit, I had closed the laptop, turned back to him, and repeated the part about his brother. He nodded. “That’s the part I was trying to say,” he told me.
That is the standard I keep in mind now. Not whether AI can make medicine faster, but whether it leaves enough of the physician visible for the patient to recognize.
For readers who want a fuller sense of how I think about clinical identity, documentation, and physician responsibility, I also write from the same standpoint at my editorial home on sinabarimd.com, where the physician’s voice stays central even when the subject wanders beyond medicine.
For my background and training, see Dr. Sina Bari’s physician profile and Stanford-trained surgical background.
FAQ
What happens if a clinic uses an AI scribe but the note sounds more certain than the visit really was?
The main risk is clinical overstatement. If the machine turns uncertainty into certainty, the chart can mislead future decisions, especially when another clinician reads it days later. I treat AI-generated text as a draft that must be checked against the patient’s actual words, exam, and the clinician’s own judgment.
How much documentation time can ambient AI actually save physicians?
Published studies have reported meaningful but variable gains, including 28.8% less documentation time per patient and 11.8% less time outside working hours in one primary care study. Other trials have shown smaller or inconsistent effects depending on the product and setting. The practical answer is that some clinicians gain time, but the size of the gain depends heavily on workflow and implementation.
What is the borrowed voice problem in medical AI?
It is the gap between a note that sounds clinically polished and the physician who still has to own it. AI can borrow a clinician’s phrasing, structure, and style, but it cannot borrow accountability. If the note becomes too smooth, it can obscure uncertainty, emotion, or context that mattered in the room.
How does Dr. Sina Bari think physicians should use AI documentation tools?
I would use them as assistants, not authors. That means reviewing every note for factual accuracy, emotional accuracy, and clinical judgment before signing. It also means refusing tools that encourage passive trust or make the clinician less present during the encounter.
Can AI scribes reduce burnout without changing the clinician-patient relationship?
Sometimes, but not reliably by themselves. Burnout is tied to workload, moral injury, time pressure, and the feeling of being pulled away from patients, so a faster note does not automatically fix the deeper problem. The best-case outcome is when AI returns attention to the patient rather than just accelerating the inbox.