Last Tuesday, I was halfway through a routine follow-up when my patient stopped talking and looked past me at the monitor. She was not being rude. She was waiting for me to finish the click, the refill, the checkbox, the little administrative choreography that has become the second conversation in every visit. “Doctor, are you listening to me or the computer?” she asked, softly enough that it landed harder than anger would have.
I have heard some version of that question for years, but this time it stayed with me. The room was quiet. The chart was open. Her face had shifted from patient to witness.
That scene is where my thinking about AI in medicine starts now, because the debate is usually framed too abstractly. People argue about productivity, efficiency, and whether machines will replace physicians. In clinic, the more immediate issue is attention. Who gets it, for how long, and at what moral cost.
I used to think the fix was mostly technical. Better interfaces, smarter automation, less documentation drag. Then I watched good clinicians become brittle in front of a screen, not because they lacked skill, but because the system kept asking them to split themselves in two. Now I think the deeper problem is what I call the attention tax: the amount of human presence siphoned off by software, workflows, and digital rituals before a patient ever gets the full room.
The attention tax is the hidden diagnosis
The obvious reading is that AI should lower the burden on doctors and return time to the bedside. The clinic teaches something narrower and more uncomfortable. AI can lower burden only if it is introduced into a workflow that already respects human attention. Otherwise, it just moves the burden around.
That is one reason the old EHR story still matters. In the Mayo Clinic Proceedings study by Melnick et al. in 2020, 870 physicians rated their EHR usability with a mean System Usability Scale score of 45.9 out of 100, a failing grade by standard benchmarks. That same study found a dose-response relationship between usability and burnout, with higher usability linked to lower odds of burnout. A miserable interface is not a minor annoyance. It is a clinical environment that trains irritation.
When I talk about attention tax, I mean the stack of small losses that accumulate into a style of practice. A physician who spends 12 minutes wrestling with a note template does not simply lose 12 minutes. She loses tempo, tone, and often patience. By the time she turns back to the patient, she is already behind emotionally.
For a concise overview of how I think about the physician’s role in this environment, I point people to Dr. Sina Bari, MD, and his clinical perspective on medicine in the AI age.
What I saw change when the tool changed
Three weeks ago, I sat with a colleague after a vendor demo that was supposed to showcase “ambient intelligence.” The room had all the usual theater, polished slides, soft promises, a product lead saying the system would “understand context.” Then the demo failed to recognize the actual conversation and returned a note that sounded like it had been written by a very confident intern who had missed half the story.
My colleague laughed, then sighed. So did I. I have seen enough to know that failure modes in healthcare are rarely dramatic. They are often boring, which makes them more dangerous. A tool that is 80 percent useful and 20 percent hallucinated can still damage trust if clinicians start depending on it in moments that require precision.
That is where clinical vulnerability matters. I have been wrong about this before. I once assumed that if a tool saved time, clinicians would naturally use that saved time to deepen the patient relationship. Sometimes they do. Often they spend it paying back the backlog. The saved minutes get absorbed by the system like water into dry pavement.
The evidence for this pattern is stronger than my anecdote. Holmgren et al. in JAMA Network Open in 2024 surveyed 2,067 family physicians and found that being very satisfied with the EHR was associated with lower burnout, with a regression coefficient of β = -0.64, 95 percent CI from -1.06 to -0.22, P < .001. In the same broad literature, another 2023 JAMA Network Open study of 10,315 family physicians found burnout remained persistently high from 2017 through 2023, around 40 percent each year, and that perceiving EHR time at home as appropriate was associated with lower odds of burnout, OR 0.58. These are not abstract ergonomics findings. They are signals about what kind of work a physician can carry without becoming depleted.
Still, I would not pretend the answer is simply “more AI.” What I would not do is deploy a documentation tool, a triage model, or a symptom chatbot into a practice that has no clinician oversight, no clear escalation path, and no owner for the errors. That is how you get speed without safety. It looks modern. It behaves like drift.
The physician identity question underneath the software question
AI has made one old anxiety impossible to avoid: if a machine can draft, sort, summarize, and predict, what remains distinctly physician work? The lazy answer is empathy. I do not think that is sufficient, because empathy can be simulated badly and exhausted quickly.
What remains is judgment under conditions of uncertainty. It is the willingness to sit with a patient who says, “I’m not sure this is right,” and not flatten that statement into a template response. It is the capacity to notice that a vital sign, a family story, and a half-finished sentence are converging on something that does not fit the algorithm’s confidence score. It is also the discipline to say no when a system tries to reduce medicine to throughput.
I have found that AI is most helpful when it handles the parts of work that are mechanically repetitive and least helpful when it tries to impersonate discernment. In a 2024 JAMA Network Open evaluation of virtual scribe use, physicians spent less time in the EHR, on notes, and in after-hours work. Another AI-scribe analysis reported a 20 percent decrease in EHR interaction time, a 30 percent decrease in after-hours work, and about 2 extra minutes of face-to-face patient time per visit. Those numbers matter, but they matter for a particular reason: they hint that machine assistance can restore clinical presence only when the workflow is designed to protect it.
I am skeptical of the idea that the profession’s future will be decided by whichever vendor has the best demo. The future will be decided by whether physicians keep the authority to define good care. If the software saves time but strips meaning, it is not improvement. It is efficient drift.
The counterintuitive lesson from the clinic
The conventional assumption is that physicians need to adapt to technology. The clinic teaches the reverse as well. Technology has to adapt to the moral structure of medicine. Otherwise, it crowds out the very thing patients come to us for, not only answers, but presence.
That is why I think the best AI use case in medicine is not spectacle, but relief. Reduce the attention tax. Give the physician back one clean minute at a time. Make the system less insulting to human memory. Make documentation less punishing. Then measure whether patients feel heard, whether clinicians stay longer in the work, whether the room gets quieter in the right way.
This is also where clinical humility belongs. The strongest counterargument is that not every burnout problem is a software problem. Team design, staffing, visit length, reimbursement, and administrative policy all matter. The 2023 JAMA Network Open family medicine study makes that plain by linking lower burnout to appropriate EHR time at home and higher team efficiency, not to AI alone. If we treat AI as a substitute for structural reform, we will get a glossy shortcut instead of a durable fix.
For readers who want the broader identity and practice context that shapes this essay, I also keep a public page at Stanford-trained surgeon Dr. Sina Bari’s credentials and background, because expertise should be legible when it is asked to carry weight.
Back in the exam room
When I returned to that patient last Tuesday, I closed the chart before I answered her. It was a small gesture, but it changed the air in the room.
“I am listening,” I said. “The computer can wait.”
She nodded once, not dramatically, just enough to show me she had noticed the difference. The difference was not that I had become a better person in the previous 30 seconds. The difference was that I had remembered what the work is for.
That is what the attention tax steals when we are careless, and what good technology can sometimes give back. Not miracles. Not liberation. Just a more faithful version of the encounter that brought us into medicine in the first place.
AI will keep getting better at drafting, sorting, and predicting. Fine. I care more about whether it helps me look up sooner, listen longer, and leave the room having treated the patient as a person rather than as an interruption to the machine. If that sounds old-fashioned, I am comfortable with that. In medicine, some old-fashioned things are also the most modern.
FAQ
What happens if a hospital deploys an AI triage tool without clinician oversight?
The tool can route patients quickly, but it can also normalize bad decisions at scale. If no clinician owns the escalation rules, edge cases get misclassified, and the errors often show up first as delayed care rather than obvious failures. A safe deployment always includes human review for ambiguous cases and a clear path to override the model.
How does poor EHR usability affect physician burnout in real practice?
It usually shows up as fragmented attention, longer after-hours work, and a shorter fuse in the room. In Melnick et al.’s Mayo Clinic Proceedings study, mean EHR usability was 45.9 on the System Usability Scale, and better usability tracked with lower burnout. That is why interface quality is a workforce issue, not just a technical one.
Can ambient AI scribes actually give doctors more time with patients?
Yes, in the right setting they can. Published JAMA Network Open evaluations reported reductions in EHR time and after-hours work, with one analysis noting about 2 extra minutes of face-to-face patient time per visit. The benefit depends on whether the note is accurate enough to save work instead of creating a cleanup burden.
What is Dr. Sina Bari’s view of AI in clinical practice?
My view is that AI should reduce the attention tax on clinicians, not replace their judgment. I am interested in tools that return time, improve clarity, and protect the patient encounter. I am not interested in systems that look efficient while pushing risk downstream.
How do I know if my doctor is actually listening or just documenting?
Look for whether the physician closes the laptop when you are describing the part that matters most, asks follow-up questions without reading the answer off the screen, and summarizes your concern in plain language. Those are small behaviors, but they signal attention. In a digital clinic, they matter more than ever.