AI scribes and informed consent
AI scribes now sit in on consults at over 60% of US Epic hospitals, with consent obtained via a murmured 'yes' after the patient has sat down. The technology is good. The consent protocol and its execution vary by clinic and by clinician. A patient's-perspective argument for what real consent would look like, why the cookie banner regime tells us we already know how, and where the next category-defining vendor will land.
When I was a medical student, I couldn't join a patient consult without proper consent first. That meant consenting in advance, by someone other than me, while I was kept out of the patient's sightline so they wouldn't feel pressured to say yes to a stranger joining their consult. In my own experience as a patient, the AI scribes now sitting in on consults and writing the notes get a fraction of that preamble. I haven't seen this questioned at the volume it warrants.
The culture I trained in
Clinical observation is essential to training the next generation of doctors, but as a medical student I wasn't essential to the consult itself. Being granted entry was a privilege – and fair enough, the bar should be high. That's the culture I trained in. Like a medical student, these ambient AI scribes are also non-essential observers with variable note writing capabilities. Yet, with the AI, I've been 'consented' after I've sat down, stared down by this faceless observer, and pressured by the clinician's productivity stake in my cooperation.
Without informed consent, this technology runs on resignation, at best, and unwitting acceptance, at worst. Real consent would be a plain pros-and-cons sheet, plus some data use and retention policies, given well in advance, that a patient could decline without confrontation.
Even the in-room ask could be done better. "I have a piece of software that records our conversation and provides a summary. Would you like any more information about that?" And on a later visit, after the patient has had time to consider, "Is it alright if I use it today?" That's a different question from "do you mind if I use the AI scribe?", which is the version I've seen. Even as a doctor myself, I feel uncomfortable declining that ask at the risk of tanking the rapport of the consult. I suspect many patients feel a sharper version of that discomfort, afraid of being labelled insubordinate, difficult, or paranoid, when they should be free to focus on the deeply personal complaint they came in for.
Not the technology, the consent
Some of these tools are great at the job they're sold for. The administrative load on clinicians is crushing, and an ambient scribe gives a meaningful chunk of that time back to the patient in the room. Adoption is already at scale – AI scribes are in use at over 60% of US hospitals using Epic EHR as of mid-2025. Regulators have started flagging concerns; the practice at the patient level hasn't followed. AI software in healthcare is the future, scribes will be part of it, and my issue isn't with the technology. It's with the consent protocol around adoption (on the clinic side) and the in-room consent practice (on the individual clinician side). Both have been treated as an afterthought.
There's a related conversation about whether the notes themselves are accurate. One anecdote on LinkedIn recently described a scribe rewriting his femoral acetabular impingement diagnosis as osteoarthritis, with the incorrect diagnosis then reaching the specialist he was referred to, and the error caught by the patient, not the system. I can't verify the case, but the failure mode cited is what's turning up in published reviews. That sounds like a meaningful lapse, but it isn't what has unsettled me.
What gets recorded
The recording (or transcription, summary, whatever) doesn't include only my medical condition, which I'd happily hand over de-identified. We talk about intimate details of my life, my plans and my feelings, the things that populate the edges of why I came in. That data now sits with a company I can't name, in a jurisdiction nobody mentioned, kept for a period nobody disclosed. I build software. I'll do GDPR and CCPA compliance to the ends of the earth for an app barely anyone uses. My consult gets uploaded to a cloud server on the strength of consent I wasn't informed enough to give.
Two consults
I recently took an elderly relative to the doctor. She was asked the same quiet question on the way in and said yes, because you say yes to the doctor. When he stepped out she believed she was alone and began to speak freely. I had to tell her the session hadn't ended and the room was still listening. She had no concept of what she'd agreed to, and what she'd already said had gone somewhere she couldn't ask for it back. I'm not certain every clinician obtaining that consent fully understands where the data goes either.
This is anecdotal. I don't know if it affects 99% of AI scribe consults or just the two I'm writing about today. Having spent the last six years studying and working in hospitals and clinics, I suspect it's closer to the former.
Security theatre, wrong door
I spend my life clicking through cookie banners. Every site I open stops me at the door to ask, in writing, whether it may store a preference about which articles I read. It's security theatre for choices that barely matter. Where my browsing history gets two paragraphs and a preference pane, the audio of me describing my health and my life to my doctor gets a murmured blanket yes after I've sat down. We built an entire apparatus for informed consent, for one human being sitting in the room, and skipped all of it the moment the observer was a microphone.
Right now consent is clinician-led, which leaves it susceptible to user-level variation and failure. The vendor who solves integrated consent UX with one click will define the category.
We know how to ask permission. Somewhere between the medical student and the microphone, we stopped.