Editor’s note — I’m handing the keys to Diabetic Foot Online this week to my friend and colleague Dr. Lucian Feraru, a gifted podiatric surgeon and Podiatric Residency Research Director at Adventist Health White Memorial, and a co-conspirator in our work on the clinician’s digital twin. Lucian has been thinking hard about what happens to the physician after the machine learns to write. His answer is one of the clearest things I’ve read on the subject: the next clinical skill isn’t generation — it’s verification. Read on. —DGA
The note can write itself now. The hard part is knowing whether it should be trusted.
By Lucian Feraru, DPM, FACFAS, DABPM
I signed a note recently that I did not really write.
The software had pulled the visit together into something that looked pretty good. History, exam, assessment, plan. The grammar was clean. The sequence made sense. The whole thing had the shape of a careful medical note.
That was what bothered me.
If the note had been obviously bad, I would have slowed down. Everyone slows down for the broken thing. The dangerous version is the note that is almost right, especially at 5:15, with another patient waiting, a staff question at the door, and a chart queue that is already too long.
This is the part of medical AI that still feels underbuilt. We talk constantly about whether it can generate. Can it write the note, summarize the chart, read the photo, suggest the code, draft the prior authorization, flag the risk?
Increasingly, yes. At least well enough to be useful.
But once it does that, the physician’s job changes. I am no longer starting with a blank page and building the record from the visit. I am looking at a finished record and asking a harder question: did this thing preserve what actually happened?
That is a different skill.
It sounds easy if you have never done it all day. Just review it. Just verify it. Just keep the human in the loop.
In clinic, “just review it” is not a real workflow. A real review means checking whether the left foot became the right foot somewhere in the note. It means noticing that last month’s exam finding carried forward into today’s visit. It means catching that a wound measurement looks precise even though the image was taken from a bad angle. It means seeing that a plan says vascular follow up was arranged when, in reality, the referral was only placed and no one knows if the loop will close.
Those are not dramatic sci-fi errors. They are ordinary chart errors. That is why they matter. They fit right into the day.
A bad AI output is easy to distrust. A decent one is harder. It earns your confidence line by line, and then asks you to notice the one sentence that should not be there. Or worse, the one sentence that should be there and is missing.
Absence is the worst kind of error to audit. If a risk flag appears, you can decide whether you agree with it. If it never appears, there is nothing to push back on. The chart simply moves on.
We have seen this before. Every clinician already lives inside a warning system. Drug interactions, allergy popups, best-practice advisories, sepsis alerts, quality prompts. Some are useful. Many are noise. Over time you learn to move through them quickly because the work demands it. That is not laziness. It is survival inside a badly tuned system. Medical AI can recreate the same problem at a higher level. If every generated note, code, plan, measurement, and risk flag requires the same kind of vigilance, physicians will adapt the way they always adapt: they will triage attention. The question is whether the system is designed well enough to know where that attention actually belongs.
That is where I think medicine is underestimating the shift. The risk is not only that AI will be wrong. We already know it will be wrong sometimes. The risk is that it will be good enough, often enough, that clinicians stop reading it like a source document and start reading it like confirmation.
Most of us were not trained for that.
We were trained to take a history, examine the patient, interpret studies, make a plan, and document our reasoning. We were not trained to audit fluent machine generated medical language at speed. We were not trained to ask, every few minutes, whether a clean paragraph is quietly laundering an assumption into the medical record.
And yet the signature remains ours.
That part has not changed. The output may come from a model, an ambient scribe, an image system, a coding engine, or some combination of all of them. But the record ends with a clinician’s name. The signature is where the system hands the risk back to the doctor.
There is a phrase I keep coming back to: ownership without authorship.
That is the strange new position many clinicians are being placed in. You own the note, the order, the code, the plan, the measurement. You may not have written it. You may not know exactly how it was assembled. You may not know which sentence came from the visit, which came from old chart context, and which came from a model filling in what sounded likely.
Still, you sign.
I am not arguing that we should go backward. I do not want to manually type every note for the rest of my career. Most doctors do not. The administrative burden in medicine is absurd, and AI can remove some of it.
But if we automate generation and treat verification like a quick clerical step, we are going to build a dangerous bargain. The machine gets the speed. The doctor gets the liability. The patient gets whatever slips through the gap.
Verification has to become a designed part of clinical AI, not an afterthought.
That means the system should show me where a measurement came from. It should mark what was carried forward. It should separate what was observed today from what was inferred from old documentation. It should make disagreement faster than blind acceptance. It should know that a laterality error is not the same as a typo. It should put friction in the places where being wrong is expensive.
More importantly, clinicians need to help define what a real review looks like. Not a checkbox. Not a click before signature. A review that actually matches the risk of the output.
Some parts of a note can be glanced at. Some parts need source evidence. Some parts should never be quietly filled in by a model without making that obvious. A good system would know the difference, or at least be built by people who do.
The next generation of clinical AI will not be judged only by how well it writes. Writing is becoming cheap. The harder question is whether it can support the doctor who has to decide whether the writing is true.
That is where the real clinical skill is moving.
The signature should mean a human agreed. Not that a human was nearby.
Lucian M. Feraru, DPM, FACFAS, DABPM, is a podiatric surgeon and Podiatric Residency Research Director at Adventist Health White Memorial in Los Angeles. His work centers on diabetic limb reconstruction, remission after wound healing, and the role of AI as a cognitive extension — not a replacement — in clinical practice.
Leave a Reply