Is AI turning you from physician to proofreader?

An artificial intelligence (AI) scribe sat in on the visit, listened and wrote the note. It was clean, well-organized and finished before the patient reached the parking lot. It also stated the patient had diabetes. She didn't.

That error is real, one of several flagged in a 2024 University of California, Davis Health pilot with results published this year in JMIR Medical Informatics, alongside an AI note that told a patient to start aspirin never prescribed, and another that got a metformin dose wrong. The study's overall verdict was reassuring: 94.7% of notes were free of serious errors. Which is precisely the problem. When a tool is right almost all the time, the temptation is to stop checking, and in that same pilot, nearly 15% of notes were signed off with no edits.

This is the quiet trade the AI scribe asks of you. It lifts the documentation burden, yes. In exchange, it hands you a new and arguably harder job: catching the rare, serious error hidden in a stack of notes that all look fine. The typing is gone. The vigilance can't be.

The risk of automation bias

There's a name for the trap, and a national regulator has already flagged it. In its 2025 safety guidance on ambient scribes, the Australian Commission on Safety and Quality in Health Care warns that clinicians are "susceptible to automation bias when they accept ambient AI scribe outputs as being accurate and complete." The phrase is worth sitting with. Automation bias isn't carelessness or fatigue. It's a predictable feature of how human attention works when a machine does the first draft well. The better the tool performs, the more it earns your trust, and the less scrutiny each note gets. A scribe that was wrong half the time would keep you sharp. One that's right 95% of the time invites you to glide.

That glide is exactly what the editing data capture. When physicians change a median of 9% of an AI note's words, and sign nearly 1 in 7 without touching a thing, the AI hasn't eliminated the work of documentation — it has relocated it, from composing to catching. And catching is the harder cognitive task. Writing a note forces you to think through the encounter; rubber-stamping a fluent, plausible draft does not. The error that tells a patient to start aspirin doesn't announce itself. It reads as confidently as every correct line around it.

Which is why the regulators keep returning to a single, unglamorous point: The responsibility never transfers. "Responsibility for the quality and accuracy of ambient AI scribe information and summaries included in health care records remains with the clinician," the Australian guidance states, flatly, without exception for how busy the day was or how good the tool usually is. The AI is not a colleague who shares the medico-legal risk; it is a drafting instrument, and you are the author of record. The scribe writes. You sign. And your signature means what it has always meant.

An argument against AI?

None of this is an argument against the technology. It would be a poor one, because the upside is real and well documented. The same body of research raising the alarm also shows AI scribes giving physicians something they have been starved of: time, and attention for the person in front of them, instead of the screen beside them. Studies have tied ambient documentation to lower self-reported burnout and less time spent typing after hours, the so-called "pajama time" that drives so many physicians out of the profession. For a clinician who has spent a decade finishing notes at 10 p.m., a tool that hands back even part of the evening is not a gimmick. It is a genuine improvement in a working life. Any honest account has to start there.

But "saves time" and "needs watching" are not competing claims. They are the same claim, seen from two ends. The time is saved precisely because the AI does the drafting; the watching is required precisely because the AI does the drafting. You cannot keep the first without accepting the second, and the danger lies in pretending otherwise: in treating the scribe as a finished product rather than a first draft, in booking the reclaimed minutes as pure profit while quietly skipping the review that the minutes were supposed to fund. The UC Davis team is blunt about this in its conclusion: Careful clinician review "remains imperative," not optional, not eventually, but now. The efficiency is real. It is also conditional. It holds only as long as someone is still reading.

The question for a practice isn't whether to adopt an AI scribe; that decision is largely made, and for good reasons. The question is whether the workflow around it is honest about the second half of the bargain. Does the schedule leave room to actually review notes, or does it quietly assume the AI got it right and pack the day accordingly? Is "percentage of notes edited" something the practice even tracks, the way UC Davis did, watching for the physician who stops editing altogether as a warning sign rather than a productivity win? The tool will keep its promise. Whether the practice does depends on choices that have nothing to do with the AI and everything to do with how the humans decide to use it.

Physician or proofreader?

So: physician or proofreader? The framing is a provocation, and like most provocations, it hides a false choice. The doctor reviewing an AI-drafted note is not being demoted to proofreader. She is doing something that proofreading only resembles from a distance, bringing clinical judgment to bear on whether a plausible-sounding document matches the patient she actually saw. A proofreader checks whether the words are correct. A physician checks whether they are true, and only a physician can. The aspirin that shouldn't be there, the diabetes the patient doesn't have, the metformin dose that's subtly wrong — none of these is a typo. Catching them is not clerical work. It is medicine, performed on a draft instead of a blank page.

The danger was never that AI would turn doctors into proofreaders. It is that doctors, handed clean and confident drafts day after day, might start proofreading when the moment still calls for a physician, skimming where they should be scrutinizing, trusting where they should be verifying, signing where they should be reading. The tool doesn't make that choice. The schedule that assumes the AI got it right makes it. The productivity target that books the saved minutes twice makes it. The quiet erosion of vigilance, one unedited note at a time, makes it.

Generative AI is going to write a great deal of medicine's paperwork, and on balance that is probably good. But the signature at the bottom will keep meaning what it has always meant: that a physician read this, weighed it and stands behind it. Hold onto that, resource it, build the day around it, and the AI scribe is exactly the gift it promises to be. Let it slide, and you haven't saved time at all. You've just moved the risk somewhere harder to see.

Four rules for using AI scribes safely

1. Treat every AI note as a first draft, never a final one.
2. Check medications and diagnoses before anything else.
3. Track how many notes go out unedited.
4. Protect the time the AI gives back; don't spend it twice.

Purnendu Bala is a health care revenue cycle analyst focused on AI-assisted workflow design for independent physician practices. He studies how operational inefficiencies impact financial performance across U.S. and global health care systems and contributes analysis through OutsourceRCM, a U.S.-focused medical billing and RCM services company.