Last Tuesday, a woman in her sixties sat in my clinic and slid her phone across the desk before I had finished asking about her symptoms. She had already used an AI tool to summarize her medication list, and she wanted to know why the answer from the app sounded cleaner than the one in her chart.
“So which one is right?” she asked, half amused and half annoyed. I looked at the screen, then at her, and had the familiar feeling that the room had quietly changed before I had a chance to name it. The chart had become a second opinion source, and the patient knew it.
I used to think the AI question in medicine was mainly about accuracy. Then I watched a patient compare an algorithmic summary with the note I had written, and I realized the deeper issue was authority. Now I think the real question is which work we want technology to accelerate, and which work we want it to force us to do more carefully.
The chart is not the only thing being rewritten
There is a comfortable story about AI in medicine that says the software will take the clerical labor, the doctor will keep the judgment, and everyone will leave happier. I understand why that story is attractive. I also think it is incomplete. What I have seen in practice is a three-way shift, between documentation, trust, and professional identity. I call it the quiet triangle.
The phrase is plain on purpose. AI changes the note. The note changes the encounter. The encounter changes how I understand my role. Once those three points start moving together, the effect is bigger than productivity. It reaches the ethics of who speaks for the patient, who owns the explanation, and who gets to look certain in a room where certainty is often rented, not owned.
In a 2020 study in Mayo Clinic Proceedings, Melnick et al. reported perceived EHR usability of 45.9 on the System Usability Scale among U.S. physicians, and lower usability tracked with greater burnout. That number matters because it reminds me that digital tools do not arrive as neutral helpers. They shape the emotional cost of clinical work. A tool that eats time reshapes temperament.
In my own clinic, I have seen that effect in small ways. A rushed note makes me shorter with the next patient. A badly structured inbox makes me defer a hard call until the afternoon. A smoother workflow gives me back a few ounces of patience, and that is not trivial. Patience is a clinical resource.
The novel part is not intelligence, it is delegation
What unsettles many physicians is not that AI can imitate language. It is that it can impersonate parts of clinical labor that used to anchor our sense of competence. I have dictated notes, corrected generated summaries, and watched an AI scribe rescue a visit that would otherwise have ended with a tired, incomplete record. I have also seen a system produce a crisp paragraph that was just wrong enough to be dangerous. The prose looked confident. The context was missing.
A 2023 study on AI scribes and clinician well-being reported burnout falling from 45 percent to 35 percent overall at three months, and from 45 percent to 31 percent among physicians, with statistically significant changes. I read that kind of result with interest, but also with caution. Lower burnout is a real benefit. So is the possibility that clinicians start trusting a machine to finish the sentence faster than they trust themselves to verify the facts.
That is where the quiet triangle sharpens. The promise is not simply automation. The promise is delegated attention. A tool that drafts a note is also making a claim about what deserves notice first. If it gets the right things into the chart, fine. If it nudges me toward the easily captured and away from the clinically subtle, I have a workflow improvement that may also be a diagnostic tax.
Here is what I would not do: I would not let an AI system author a final assessment in a high-stakes visit without the physician actually reconstructing the logic from scratch. Not because the machine is useless, but because clinical judgment is not a polished paragraph. It is a chain of exclusions, probabilities, social context, and timing. A clean summary can hide a weak premise.
Trust is now part of the differential diagnosis
Patients are already reading the same summaries I read. That fact changes the consultation. It also changes the meaning of authority. In a 2024 vignette study indexed in the AI and health literature, Zondag and colleagues examined how AI influenced patient-physician trust, showing that trust can move depending on whether the patient sees the AI as supporting, rather than replacing, physician reasoning. The lesson is practical: trust grows when the clinician remains legible.
I have felt that legibility problem in my own work. If I bring a patient an AI-generated handout and then disappear behind the screen, the tool may look efficient while the relationship thins out. If I use the same tool to explain a lab trend, compare possibilities, and slow the pace enough to invite a better question, the encounter feels more human, not less. The difference is not the software. The difference is whether I am still doing the translating.
That translation work is the part outsiders underestimate. Patients are not just asking for data. They are asking, in effect, “Do you understand what this means for my body, my family, my job, and my fear?” A machine can answer the first half. The second half still belongs to a clinician who can sit with ambiguity without pretending it is clarity.
There is a useful warning in the 2024 survey literature on physicians and AI adoption, which found in one sample of 297 physicians with a median age of 36 that ethical concerns and age-related differences shaped willingness to adopt AI tools. The point is not that younger doctors are naive or older doctors are resistant. The point is that adoption is filtered through identity. Who we think we are as clinicians affects what we let into the room.
The professional identity problem is the real story
I have spent enough time in medicine to know that every generation thinks it is protecting the soul of the profession from the latest machine. Some threats are overblown. Some are not. The current wave is different because it touches the verbs of medicine, not just the tools. We diagnose, explain, comfort, prioritize, document, and decide. AI now touches all six.
That is why I do not buy the lazy framing that the profession is simply being “augmented.” Augmentation is too neat. In practice, the more consequential shift is that physicians are being asked to supervise, edit, and morally certify outputs generated elsewhere. That is a different job from the one many of us trained for. It can be better. It can also be alienating.
I once thought the hardest part would be learning the interfaces. Then I spent time reviewing a draft note that was syntactically perfect and clinically incomplete. Now I think the hardest part is learning how not to surrender our own pattern recognition. If the machine becomes the first writer, the human still has to be the first skeptic.
That is why the phrase “quiet triangle” matters to me. It gives me a way to name the pattern when a tool is sold as convenience but behaves like governance. It also helps me ask better questions during vendor demos. Who is accountable for the recommendation? What kind of error does the system make when the data are messy? What happens when the output sounds confident and is incomplete? Those are not technical niceties. They are ethical questions in a lab coat.
For readers who want the clinician’s perspective behind this essay, my background and training are described on Dr. Sina Bari’s Stanford-trained surgeon profile. The reason I mention that is simple. In medicine, perspective is part of the evidence trail.
The limit case is where the argument gets honest
There is a temptation, especially in editorial writing about AI, to flatten the field into a moral binary. Either the technology is liberating clinicians or it is degrading them. Real practice is messier. I have seen a good tool reduce time spent on clerical cleanup. I have seen the same category of tool encourage overconfidence in a note that should have been questioned line by line.
A study I keep coming back to because it is so ordinary, and therefore so revealing, is the EHR usability work from Mayo. A score of 45.9 on the System Usability Scale is not an abstract condemnation. It is a clue about what repeated friction does to attention. When clinicians are forced to spend energy navigating the machine, they have less left for the patient. When AI removes some friction, it may improve care, but only if it does not move the friction into a less visible place, such as uncritical trust.
I have also had moments where the machine surprised me in a good way. A draft summary caught a medication inconsistency I had missed in the rush between patients. I was grateful. Then I was embarrassed. Clinical vulnerability matters here because the people who insist they never miss anything are usually the least trustworthy. I am not interested in pretending otherwise.
The strongest counterargument to my position is that clinicians have always relied on aids. Stethoscopes, calculators, imaging, order sets, templates, all of them changed medicine. That is true. The difference now is not that assistance exists. The difference is that the assistance can generate language that sounds like judgment. That is a harder boundary to police.
What I think the quiet triangle really asks of us
I do not think physicians need to reject AI to protect the profession. I do think we need to stop treating it as a neutral productivity layer. The tool is already participating in how medicine is narrated, and narration shapes trust. That is especially true in specialties where the work is intimate, longitudinal, and full of incomplete information.
So I use AI with a narrow set of rules. I let it draft when the task is repetitive and the stakes are low. I distrust it when the output feels too polished for the situation. I use it to compress busywork, not to outsource responsibility. And I ask myself one question every time: if this line were wrong, would I be able to explain why, in my own words, without reading it back from the screen?
The answer has to be yes.
That is the practical center of the quiet triangle. The chart, the patient, and the physician are now all being touched by the same systems. If we are careless, the machine gets to define the relationship. If we are deliberate, it becomes a tool for preserving the part of medicine that still depends on human judgment, spoken plainly, in real time, to one person at a time.
Last Tuesday, the woman with the phone left my clinic with a clearer explanation and, I hope, a little more confidence in the fact that the chart is not the final authority on her story. She had brought me an algorithmic summary and asked for the truth. What she needed was not a more perfect note. She needed a physician willing to do the translation. That work still belongs to us.
FAQ
Can an AI scribe replace a physician’s clinical note in routine practice?
No, not safely as a final product. An AI scribe can speed up drafting and reduce documentation burden, but the physician still has to verify facts, interpret ambiguity, and ensure the assessment matches the encounter. In higher-stakes visits, the final note needs human reconstruction of the reasoning, not just editing.
What happens if a patient trusts the AI summary more than the doctor?
Trust usually shifts toward whichever explanation is more legible. If the AI output sounds cleaner but the physician does not actively translate the clinical logic, the patient may assume the machine is more accurate. The fix is not to ban the tool, but to keep the physician visibly in charge of explanation and decision-making.
How does physician burnout connect to AI tools in the clinic?
Burnout is often worsened by repetitive documentation and workflow friction, and AI can reduce some of that load. In Mayo Clinic Proceedings, Melnick et al. reported EHR usability at 45.9 on the System Usability Scale, a score associated with burnout risk. The key is whether AI removes friction or just moves it somewhere harder to see.
What is Dr. Sina Bari’s approach to using AI in medicine?
I use AI as a draft assistant, not as a final authority. If a system can help me summarize routine material or spot a discrepancy, I will use it, but I will not let it stand in for clinician judgment. The rule is simple: if I cannot explain the output in my own words, I do not trust it yet.
What should clinicians ask before adopting a new AI tool?
Ask who is accountable, what kinds of errors the tool makes, and how it behaves when the chart is messy or incomplete. Ask whether the output is designed to support physician reasoning or quietly replace it. If the vendor cannot answer those questions plainly, the tool is not ready for serious clinical work.
Does AI change physician identity, or just the workflow?
It changes both. Workflow shifts are visible first, but identity changes follow when the doctor becomes an editor, supervisor, and verifier of machine-generated language. That is a meaningful change in the job, and it deserves explicit discussion rather than vague promises about efficiency.