Ethics of AI in the Exam Room: Consent, Authorship, and the Draft You Reject

The ethical question is not whether the model is clever

Physicians are being asked to admire AI. Patients are being recorded by it. Those are different events, and only one of them is an ethics problem you have to solve before Monday. Cleverness is a technical claim. The ethical claims are simpler: did this person know they were being captured, could they refuse without losing care, will a human read what was generated before it changes their treatment, and can you stand behind the signed result. If you cannot answer those four, do not turn the tool on. You can answer them without a philosophy department.

There is also an ethics of omission. Leaving clinicians to finish charts at 9 p.m. harms attention, mood, and the next morning's patients. Refusing every tool because it is new is not automatically the moral high ground. The obligation is proportionate. Use what returns you to the patient. Reject what fabricates the record, hides the capture, or pushes action out under your name while you are elsewhere. That standard is old. The software is what is new.

This piece is for the physician who will be in the room, not for a panel. It is about consent you can say out loud, authorship that matches the signature, the duty to read, bias in whose voice is heard, and the visits where the right act is to turn the tool off. HIPAA is the floor. Ethics is what you do when the floor would still allow something you would not want done to your own parent.

Consent a patient can actually use

A poster in the waiting room is notice. It is not consent. Consent is a short conversation with a real no. 'We use software that listens to help me write the note. You can say no, and your visit will not change. I can also pause it if we get to something you want kept to this room.' Then stop talking and let them answer. If your workflow cannot honor the no, do not ask. Asking without a path to honor the answer is a performance.

Consent decays. The patient who agreed in January may not remember in June, and the patient who agreed to a note may not have agreed to a recording retained by a vendor, used to train a model, or visible to every clinician who opens the chart. Say what you know. If you do not know whether the audio is kept, find out before you claim it is deleted. Uncertainty spoken as certainty is its own kind of disrespect.

Some patients cannot consent in the usual way: emergencies, altered mental status, some pediatric visits, some patients with guardians. Have a default that is conservative. In an emergency, treat the patient and document without a gadget if the gadget would delay care. For a child, do not assume a parent's yes covers a private adolescent concern that arises mid-visit. Pause. The tool should make pauses easy. If pausing is a submenu, the product is telling you what it values, and it is not the patient's no.

Recording, retention, and who else can hear

There is a difference between a note and a recording. A note is a clinician's account. A recording is the visit itself, available to anyone the system allows, for as long as the system keeps it. Patients may accept a note and refuse a recording, and they are allowed to make that distinction even if your vendor's architecture blurs it. Ask the vendor, in writing: is audio stored, for how long, who can play it, is it used to train models, and can a patient request deletion. Put the answers where your clinicians can find them.

State law on recording conversations is not uniform, and a clinic with sites in more than one state cannot assume the strictest banner ad from the vendor covers them. All-party consent jurisdictions are not a trivia question if a family member on speakerphone is part of the visit. Your compliance officer's job is the statute. Your job is not to be the person who was surprised. If you do not know the rule where you practice, you are not ready to leave a microphone on.

Open notes changed the audience. The patient will read what the tool drafted, including the clumsy paraphrase and the diagnosis you mentioned only to discard. Ethics here is not only privacy from outsiders. It is fairness to the person in the note. Do not let a model call someone noncompliant, anxious, or poor historian unless you would say that to their face and you believe it is clinically necessary. The draft does not have a face to say it to. You do.

Authorship is the signature, not the keystroke

When you sign, you are saying the document is your account of the visit and your plan. It does not matter that a model proposed a sentence, any more than it used to matter that a transcriptionist typed one. What matters is that you had a fair chance to correct it and that you took the chance on the parts that can hurt someone. 'I did not have time to read it' is an explanation. It is not an ethical one, and it will not feel like one later.

Ghost authorship runs in both directions. A model should not be a silent coauthor of your assessment. You should not be a silent coauthor of a staff member's AI-written message sent under your name. If your name is on it, your review has to be real. Groups that treat the physician's name as a rubber stamp are making an ethical choice, not an efficiency choice. Name it in policy: what may go out under a physician's name, and what review occurred.

Students and residents need a clearer rule than attendings, because their notes are already a place where authorship is shared. An AI draft plus an attending addendum plus a resident who did not see the patient is a document with no author. Decide, in the program, whether trainees may use the tool, whether they must still produce their own assessment, and how the attending can tell what was heard versus inferred. If you cannot tell, you cannot supervise.

Hallucination, omission, and the duty to read

The failures that matter are boring. A dose changed. A negation dropped. A family history imported from the wrong patient or from a similar sentence in the model's memory of other notes. A complication listed because it is common after the procedure you did, not because it happened. You will not catch these by feeling whether the note 'flows.' You catch them by reading the lines that move drugs, knives, and dispositions, every time, while you still remember the room.

Omission is as serious as invention. A model that shortens the history may drop the symptom that brought the patient in, because it was mentioned once and the rest of the visit was about something else. The note looks focused. The next clinician misses the reason for the visit. Your duty to read includes a duty to notice what is missing that you know was said. That is harder than noticing a strange phrase. It requires that you remember the visit, which is the whole argument for signing the same day.

Do not outsource the duty to a confidence score you cannot interpret. If the product highlights low-confidence words, use the highlights as a hint, not as a boundary. High confidence and wrong is the ordinary failure of fluent systems. A human who was not in the room can still help — a transcription editor with the audio, a nurse who was present — but the signer remains responsible. Share the work. Do not share the blame away.

Whose speech the model hears poorly

Error is not random. Accent, dialect, soft speech, medical jargon from a patient who is also a clinician, and code-switching will produce worse drafts for some patients and some physicians than for others. If you deploy a tool and only track average edit time, you will miss a disparity that sits inside your own clinic. Ethics requires you to look. Sample notes from the speakers and the patients the demo did not include.

When the draft is worse, the burden falls on whoever is tired enough to sign it anyway. That is often the same clinician whose speech the model mishandles, working later than their colleagues to correct it. A fair rollout includes a human path that does not punish them for the model's weakness: editor support, permission to turn the tool off for their sessions, and a vendor obligation to improve on their audio rather than a suggestion that they speak differently.

Patients with limited English, interpreters in the room, and families who speak for a patient create attribution problems that are ethical, not just technical. The note must say who said what when it matters. A paragraph that merges the interpreter, the daughter, and the patient into one 'patient states' is a lie of convenience. Configure the tool to tolerate messier attribution, or write those visits yourself. Convenience is not a reason to erase a speaker.

When the right act is to turn it off

Turn it off when the patient says no. Turn it off when you cannot explain where the audio goes. Turn it off for the portion of a visit about violence, self-harm, immigration status if the patient is afraid, or any fact they ask you not to write in a chart that many people open. You can still document what you must for safety, in your own words, with judgment about detail. The tool does not have that judgment.

Turn it off when you are about to do something irreversible and the draft is in the way: a procedure, a code, a transfer. The record of those moments should be written by people who were there, after the moment, in the format your hospital already trusts. A transcript of overlapping voices is not a procedure note. Do not let a desire to be modern produce a worse record of the most serious minutes of the day.

Turn it off when you notice you are no longer reading. That is a personal stop rule, and it is the one most likely to save you. If you have signed three notes by scrolling, the tool has trained you. Take the rest of the session without it. Tell your partner you did. A clinic that treats that as professionalism, rather than as resistance, will keep its physicians honest. A clinic that mocks it will get signatures that mean nothing.

What you owe your colleagues and your students

You owe them honesty about the tools in the chart they are inheriting. A covering physician should be able to assume that a signed assessment was accepted by the signer, not merely generated beside them. If your group allows unsigned AI text to sit in the note as if it were the plan, you are laying a trap for the night shift. Mark drafts as drafts until they are signed, and do not let draft text flow into handoff tools.

You owe students a model of refusal. They will copy what you click. If you sigh and accept a plan you did not make, they will learn that the record is theater. Narrate the edit once a day. 'I am deleting this because I did not say it.' That sentence teaches more ethics than a module. It also teaches the patient, if they are still there, that you are the author. Let them hear it.

You owe your partners a shared rule so that patients do not get a different standard depending on who is in the room. Write the rule in a page, not a binder: how you ask, how you pause, what the scribe is allowed to draft, what must be confirmed on the order screen, and who can turn the system off for the group when a defect appears. Revisit it when the vendor updates the model. Updates are not neutral. They are a new tool wearing the old name.

Equity, incentives, and the note that gets longer

A tool that makes notes longer will be loved by anyone paid on documentation volume and distrusted by anyone who has to read them at 2 a.m. Length is not a clinical good. Before you celebrate a richer history, ask who benefits if every visit grows a paragraph of boilerplate. If the answer is a quality score, a code, or a vendor's engagement metric, you have an incentive problem sitting inside an ethics problem. Measure whether the extra sentences change a decision. If they do not, they are cost imposed on the next reader, including the patient.

Equity shows up in who gets the careful edit. If the draft is clean for some patients and ragged for others, tired clinicians will correct the ragged ones less often, and those charts will carry more of the model's guesses. That is not a neutral error rate. It is a distribution of risk. Sample by language, by accent of the clinician, and by visit type, not only by average time saved. If one group is signing faster because they are reading less, stop and look. A rollout that improves the mean and worsens the tail has failed the people in the tail.

Be honest with patients about what the tool is for. It is there to help you write, not to judge them, score them, or decide whether they are a difficult historian. If a future feature proposes a risk label from the transcript, that is a new intervention and it needs its own consent, its own evidence, and a way to be wrong in the chart without the label following the patient forever. Do not let a documentation aid become a silent diagnostician because the button was on the same screen.

A Monday-morning ethic

Before clinic, know the answer to four questions. Can a patient refuse, today, without bargaining? Do I know whether audio is stored and whether it trains a model? Is there a human path for notes and messages that must not be drafted? Will I read the assessment and the medications before I sign? If any answer is no, you have an ethics problem that a feature release will not solve. Fix the no, or do not use the tool that day.

During clinic, pause more often than the demo suggested. Say what you are doing when you pause. After clinic, sample yourself: one note where you look only for a sentence you did not mean. If you find one, you do not have a scandal. You have a practice. Delete it, and notice which part of the tool wrote it. Turn that part down. Ethics in this domain is maintenance, not a single decision to be modern.

The physicians who will still like their work in five years will be the ones who used these tools as instruments and refused them as authors. Patients do not need you to be a skeptic about all technology. They need you to be the person who knows what was said, what was decided, and what you are willing to put your name under. That has always been the job. AI changes the temptation, not the job. The draft you reject is not wasted software. It is the proof that the signature still means you.