The Quiet Work of Staying Human Around AI
Last Tuesday, I finished a skin check five minutes behind schedule and found my computer still waiting for me, cursor blinking in the same note I had started before lunch. The patient, a retired teacher with a new diagnosis and a stack of printed lab results, looked at the screen, then at me, and said, “Are you writing to me or to the machine?”
I laughed because it was the kind of question that lands softly and stays with you for days. I told her I was writing to both, which was true, and also incomplete. The note was for continuity, billing, legal memory, and the next clinician. It was for the machine because the machine now sits in the middle of ordinary care, collecting our words, shaping our tempo, and quietly asking for more of our attention than it returns.
That exchange stayed in my head longer than the visit itself. I have had enough of these moments now to distrust easy slogans about artificial intelligence in medicine. I used to think the central problem was accuracy. Then I spent enough time watching tools enter real clinic flow to see that the deeper issue is authorship, judgment, and the small daily acts through which physicians decide whether they are still practicing medicine or merely validating output. Now I think the real question is simpler and harder: what do we preserve when the machine becomes competent enough to sound like us?
The first thing AI touches is tempo
When people talk about AI in medicine, they often start with diagnosis. I start with time. Time is where the profession reveals itself. A tool that saves thirty seconds on paper can cost five minutes in correction, verification, or moral hesitation. A tool that drafts a note can also make the encounter feel thinner, because the physician begins performing for two audiences at once.
The most honest data we have point in that direction. Melnick et al. in Mayo Clinic Proceedings (2020) reported a mean Electronic Health Record System Usability Scale score of 45.9 among 870 physicians, and each 1-point increase was associated with about 3% lower odds of burnout. That number matters because it shows how little friction the profession can tolerate before morale bends. It also reminds me that AI is being layered onto a workflow many clinicians already experience as structurally hostile.
A 2024 JAMA Network Open study of an AI documentation tool reported that 43.5% of intervention clinicians said documentation time decreased, compared with 18.2% in controls, and 44.7% reported less EHR frustration versus 14.5% in controls. Useful. Not universal. That distinction matters. I have seen colleagues adore a new ambient note system and then quietly stop using it when the drafts became verbose, generic, or just wrong in the places that matter most.
The profession keeps pretending the tool is the story. It never is. The workflow is the story.
The counterintuitive part is that better tools can make identity more fragile
The obvious reading is that AI will free physicians from clerical work so we can return to the human part of medicine. The clinic teaches otherwise. Once a system can compose a plausible assessment, organize a differential, or draft a patient summary, the physician has to decide what kind of expertise still belongs to the person wearing the white coat. That decision is not abstract. It shows up when I am scanning an AI-generated note and asking whether the machine captured the one sentence that changed the whole case.
I have learned to respect the speed of machine-generated prose and distrust its confidence. In one case, a note confidently summarized a patient’s symptoms, but it flattened the timeline in a way that would have obscured a medication effect. That kind of error is not dramatic. It is worse. It is the sort that survives a quick skim and enters the chart wearing a clean collar.
There is a named concept I keep coming back to: judgment slack. It is the margin between what the system says and what the clinician is still required to notice. When that slack gets too wide, people become lazy. When it gets too narrow, people become defensive and overcorrect. The useful zone is in between, where AI does the first pass and the physician still owns the last mile.
What I would not do
I would not let a vendor sell a hospital the fantasy that automation equals supervision. I would not let an AI scribe write unsupervised assessments for a clinician who does not review the text line by line. I would not use a model as a silent coauthor in any note that might later be used to justify a diagnosis, a denial, or a handoff unless the workflow has explicit human verification.
That refusal is not technophobia. It is ethics with a clipboard.
Physicians are understandably tempted by anything that reduces after-hours charting. I understand that temptation personally. I have stayed late enough to know how quickly fatigue can turn a good doctor into a spreadsheet operator. But the answer to overload cannot be blind delegation. If the tool saves time by compressing reflection, the profession pays later in errors that look small on paper and large in practice.
Bias does not disappear when the interface gets smoother
One reason I am wary of the AI enthusiasm cycle is that usability often masks fragility. A polished interface can hide uneven performance across patient groups. In dermatology, where image-based tools are often held up as the poster child for AI success, that gap has already been documented. A 2024 review reported pooled diagnostic accuracy of 0.89 AUROC in lighter skin tones versus 0.82 in darker skin tones, a difference of 0.07. That is not a rounding error. It is a warning label.
When I see those numbers, I think about the patient who walks in with a complaint that has been misread before. I think about how easy it is for a model to appear broadly competent in a controlled demo and then lose altitude when the patient is older, darker-skinned, medically complex, multilingual, or simply outside the training distribution.
The lesson extends beyond dermatology. AI often enters the clinic as a promise of standardization, then reveals the old inequality with better typography.
For a broader framing of those risks, the World Health Organization guidance on ethics and governance of artificial intelligence for health is still one of the clearest public references. It does not romanticize the technology. It asks who benefits, who is exposed, and who is responsible when the system is wrong.
The physician is still the one who carries the consequences
I used to think AI would mainly change what doctors know. Then I spent more time watching how it changes what doctors feel responsible for. Now I think the more important shift is in moral load. If a model suggests a differential and the clinician accepts it, the judgment still belongs to the clinician. If the model omits a rare diagnosis and nobody checks carefully, the harm still lands in the patient’s body, not in the software update log.
A 2024 qualitative study in PubMed on physicians’ ethical concerns about AI found recurring themes of confidentiality, beneficence, non-maleficence, advantages, risks, restrictions, and precautions. The language is familiar because the moral terrain is familiar. The novelty is the interface, not the obligations. Physicians remain accountable for privacy, explanation, consent, and restraint, even when the model can talk faster than we can type.
I have heard colleagues say they worry patients will think AI is making decisions for us. I think the better fear is subtler. Patients may assume we have outsourced more judgment than we actually have. That is why transparency matters. A patient should know when AI is helping shape the note, the triage queue, or the first draft of an assessment.
For readers who want my broader work and perspective, I keep my professional identity and current writing anchored at Dr. Sina Bari’s editorial and professional site, with credentials and background available on Dr. Sina Bari, MD and Stanford-trained surgeon background.
Clinical vulnerability is part of the job now
There was a patient encounter early in my own adjustment to AI-assisted documentation where I trusted a draft too quickly. Nothing catastrophic happened, but the draft phrased a symptom in a way that was clinically cleaner than the patient’s actual story. Cleaner usually means less true. I caught it before signing. That moment made me uneasy because it was ordinary, and ordinary is where bad habits grow.
I do not think every physician must become an AI specialist. I do think every physician should become fluent enough to recognize when the system is amplifying judgment and when it is replacing attention. Those are different events, and only one of them is safe.
The best version of AI in medicine is humble. It drafts. It sorts. It flags. It saves the physician from repetitive labor so the physician can spend more energy on interpretation, uncertainty, and the patient sitting across from them. The worst version is flattering. It produces polished text that makes us feel more efficient while dulling the edge we need to notice when the story does not fit.
What the profession should keep
The next generation of doctors will learn medicine in a different acoustic environment. There will be fewer clacking keyboards, more generated prose, and more suspicion that the chart is already halfway written before the visit ends. That change can improve care. It can also flatten the apprenticeship of judgment if we are careless.
So I want a slower ethic around faster tools. I want physicians to review every AI-assisted note as if it may one day be read in a courtroom or a conference room. I want hospitals to measure not just time saved, but error types, correction burden, and whose voices disappear in the draft. I want training programs to teach residents how to interrogate a model the way they would interrogate a confusing history. Calmly. Repeatedly. Without being impressed by fluency.
And I want us to admit something uncomfortable: some of the best use cases for AI will be banal. Medication reconciliation. Message triage. Drafting after-visit summaries. Identifying missing labs. That banality is a feature. Medicine is full of small repetitive work that can be delegated safely if the clinician stays in charge of the meaning.
Back in the clinic, the question was already answered
By the time I signed the note from last Tuesday, the patient had already stood to leave. She smiled and said, “Good. I was worried the machine was going to get the last word.”
That line has stayed with me because it names the real boundary. The machine can draft, summarize, sort, and suggest. The physician still has to decide what counts, what matters, and what gets spoken out loud to another human being. That is the job I want to keep.
AI will keep getting better at sounding like medicine. The harder discipline is remaining medicine while it does.
FAQ
What happens if a clinician signs an AI-generated note without reviewing it carefully?
The clinician inherits the content, including any omissions, misstatements, or flattened details. In practice, that can affect billing, handoffs, and downstream clinical decisions. The safer standard is line-by-line review before signature, especially for assessments, medication lists, and timelines.
How does AI note drafting change patient communication in clinic?
It can give the physician more eye contact and less typing, but only if the workflow is built well. If the draft is noisy or inaccurate, the clinician ends up editing instead of listening. Patients notice that quickly, especially when the screen starts competing with the conversation.
Where does AI in medicine create the biggest ethical risk?
The biggest risk is false confidence, because a fluent system can make weak reasoning feel complete. That affects consent, privacy, and clinical judgment at the same time. The ethical burden stays with the physician even when the text came from a model.
What is Dr. Sina Bari’s approach to AI-assisted clinical work?
I use AI only where it clearly reduces clerical burden without taking over judgment. I want it to draft, organize, and flag, then stop there. The final clinical meaning stays with the physician, because the physician carries the consequences.
How do you know if an AI tool is affecting care quality instead of helping?
Watch for correction burden, hallucinated details, loss of patient nuance, and overconfident summaries. If the tool saves time but creates more cleanup, it is not helping. The real test is whether the physician has more capacity for judgment after using it.