Last Tuesday, Mrs. R sat on the exam table with her coat still on, one hand folded over the other, and asked me, "Doc, are you actually reading me, or are you just feeding the computer?" The room was quiet except for the soft fan from the workstation and the small, embarrassing pause that followed. I have heard versions of that question more often this year than I care to admit.
Three weeks earlier, in a different clinic session, a resident had turned from the screen and said, "I don't want to become a note writer with a stethoscope." That line stayed with me. It landed because it was funny, and because it was true.
I used to think the main danger of medical AI was technical error, the obvious kind, the wrong recommendation or the hallucinated answer. Then I started noticing something quieter in my own practice, and in the people around me. The bigger pressure was identity drift, the slow feeling that the machine was not just assisting the visit but rearranging the visit's moral center. Now I think the central question is not whether AI can help medicine produce more information. It can. The question is what kind of physician remains after information becomes cheap.
I call this the bedside residue. It is the part of the doctoring that remains after the chart is done, after the algorithm has spoken, after the templated note has been closed. It includes hesitation, judgment, silence, and the patient sense that someone is still there. The residue is small. It is also the whole profession.
The seduction of helpfulness
The easiest story about AI in medicine is that it saves time. Sometimes it does. In a 2026 JAMA study of ambient AI scribes, burnout among ambulatory clinicians fell from 51.9 percent to 38.8 percent after 30 days of use. Another 2026 analysis in JAMIA found reduced mental demand, from 6.71 to 6.11, and reduced note-writing effort, from 6.75 to 6.02. Those are real gains. I would be dishonest if I pretended otherwise.
I have felt that relief myself. A strong ambient tool can return a few minutes to an otherwise crowded day. It can reduce the cognitive sting of documentation. It can make the last patient of the session feel less like a race against the cursor. That matters.
But helpfulness has its own gravity. Once a machine becomes reliable at the easy part, humans start to defer the hard part to it as well. That is where the trouble begins. A system that drafts notes can slowly become a system that drafts judgment, and then a system that edits the doctor into a reviewer of its own output.
That is the wrong bargain. I would not hand a model autonomous authority over diagnosis, consent language, or goals-of-care framing in a real clinic without direct clinician ownership at every step. I would also not let a vendor define success only by minutes saved. Minutes are useful. Trust is the currency that actually determines whether patients tell the truth in room 7.
What patients are really asking
The 2024 patient preference study indexed in PubMed found that patients preferred a human doctor most, a human doctor with AI second, and AI alone last. The sample used a 3 x 4 experimental design across disciplines, and the human-doctor condition scored better on trust, disclosure, adherence, and satisfaction. The pattern varied by specialty, with psychiatry behaving differently from the others. That variation matters, because it tells us patients are not objecting to tools. They are objecting to disappearance.
Mrs. R was not asking whether I had a chatbot in the chart. She was asking whether my attention still belonged to her. Patients know when the room has been hollowed out. They can feel when the doctor is present but not available.
I have seen this in ordinary ways. A colleague once presented a case beautifully, then admitted he had not looked up from the note template until the patient was halfway out the door. Another time, a family member interrupted me mid-visit and said, "Can you just talk to her like a person?" It was not a rebuke so much as a rescue.
The lesson is uncomfortable. The machine can widen access to information, but it cannot make a patient feel witnessed. That work remains stubbornly human.
Physician identity under pressure
The profession has always absorbed new instruments by renegotiating its self-image. The stethoscope changed bedside listening. The EHR changed the shape of clinic. AI changes something more intimate, because it interferes with the one thing physicians quietly use to measure themselves, the sense that our mind, trained over years, is the instrument.
That is why the literature on physician identity matters. A 2025 qualitative study, AI in Healthcare: Identity Threat or Opportunity? Insights From Medical Specialists, interviewed 19 physicians in Belgium and found a persistent tension between compliance and resistance. Physicians often described AI as a complementary aid, sometimes as a second opinion, but rarely as a replacement for judgment. I recognize that ambivalence. It is not anti-technology. It is a defense of authorship.
I have lived that tension in my own work. There are moments when I welcome a model's speed, and moments when I feel it trying to colonize the clinical pause. I have been wrong about both. I underestimated how much documentation fatigue was eroding me, and I underestimated how quickly convenience could tempt me to outsource attention.
This is the self-correction I trust most: I used to think the physician's identity problem would come from being displaced. Then I noticed that the more immediate threat was being flattened into a coordinator of machine output. Now I think the deeper question is whether we still practice discernment when the software is eager to pre-decide.
The bedside residue
The bedside residue shows up in small refusals. It is what stays after we decline to let an AI-generated phrase enter the note because it sounds polished but vague. It is the extra 30 seconds spent asking a patient what the illness has cost them at home. It is the judgment to ignore an impressive-sounding output when the story in front of me does not fit.
In one clinic, a model kept emphasizing a low-risk interpretation for a patient with a symptom pattern I did not trust. The output was tidy. The patient was not. I ordered the workup I would have ordered before the model existed, and I was right to do so. That case surprised me because the danger was not that AI was malicious. It was that it was plausible enough to make me feel lazy for doubting it.
That is the part we still do not say plainly enough. AI can make average work easier, and easier work can make vigilance feel optional. It is optional only until it is not.
For readers who want to know more about my clinical background, I keep a short professional profile on Dr. Sina Bari's clinician bio and Stanford training. I mention that not as branding but as context, because the point of this essay is grounded in the way medicine feels from inside the room.
What I think the profession owes patients
Medicine should adopt AI where it clearly reduces friction and preserves human review. It should reject systems that optimize throughput by sanding off accountability. That sounds obvious until procurement season arrives and the dashboard starts calling everything efficiency.
I would not deploy an AI tool whose main promise is that it will let a clinician supervise more patients without restoring any meaningful time for reflection. I would not praise a system for making the note look cleaner if the visit itself becomes thinner. I would not let a hospital confuse transcription with care.
The counterargument is real. Some clinicians will use AI well and some will misuse it. Some specialties will benefit more than others. The 2024 patient preference data already suggests that psychiatry and other relational fields may not follow the same pattern as procedural ones. That limitation is useful, not annoying. It reminds us that medicine is not one workflow.
Still, the broad contour is clear. Patients do not want a machine wearing a white coat. They want a doctor who can think with tools and remain answerable without them.
Back to Mrs. R
When I returned to Mrs. R's room, I closed the laptop for a moment and said, "I'm reading you. The computer is helping me keep up, but you are the point." The sentence was imperfect, and still better than pretending the question had not been asked. She nodded once. Then she told me what she had been afraid to say at the start, that her medication changes were making her feel "weird in a way I couldn't explain online."
That was the real visit. Not the scribe. Not the screen. The bedside residue, the thing that survives after the machine has done its useful work, was what made the conversation possible.
The profession will keep changing. I expect that. What I do not expect to surrender is the responsibility to notice when the tool starts shaping the moral weather of the room. If medicine forgets that, it will still produce notes, summaries, and triage decisions. It will produce less doctoring.
FAQ
Do ambient AI scribes actually reduce physician burnout?
Yes, in at least one 2026 JAMA study, burnout in ambulatory clinicians fell from 51.9 percent to 38.8 percent after 30 days of ambient AI scribe use. The practical point is that documentation relief can be real, but only if the tool fits the workflow and does not create new review burden.
What is the biggest risk when a hospital adopts AI documentation tools?
The biggest risk is not only factual error, it is clinical drift. When the note gets easier to generate, teams can start accepting weaker judgment, thinner histories, or vague language that sounds professional but hides uncertainty.
How do patients usually feel about AI in the exam room?
Most patients still prefer a human doctor, with a human doctor plus AI ranked above AI alone in the 2024 preference study. People are generally open to tools when the physician remains visibly responsible and present.
What does Dr. Sina Bari think physicians should refuse with AI?
Dr. Sina Bari would refuse any system that tries to replace clinician accountability with automation theater. In practice, that means declining tools that make care look efficient while stripping away review, context, or the physician's ability to say no.
Where can I read more about Dr. Sina Bari's clinical background?
You can read more on Dr. Sina Bari's Stanford-trained clinical profile. That page gives context for the perspective behind this essay and the kind of practice experience informing it.