A few months ago, a patient stopped mid-sentence and asked me a question I wasn’t fully prepared for.

We were halfway through the visit. I wasn’t typing. I wasn’t clicking through dropdown menus or half-listening while navigating a template. I was just — present. And she noticed. She said the visit felt more like a real conversation. Then she paused and asked, “Wait, is something recording this?”

She wasn’t wrong to ask.

I’ve been using an ambient AI scribe tool occasionally in my practice. The appeal is real — anyone who’s spent an evening catching up on charts after a full clinic day understands it immediately. The tool listens, interprets the encounter, and generates a note. You walk out of the room with documentation mostly done. That’s not a small thing.

But here’s what doesn’t get talked about enough: The note still has to be read. Carefully. What the AI interprets and what actually happened in that room are not always the same thing. I’ve caught phrasing that was close but not quite right, clinical details that got flattened, or context that simply didn’t translate. The note goes into the permanent medical record. That means the physician still has to own every word of it — the AI doesn’t.

So we’re in this strange middle ground. The technology genuinely helps. And it genuinely requires more vigilance than it might appear to.

This isn’t theoretical anymore

AI in medicine has moved past the pilot phase. It’s embedded in imaging workflows, clinical decision support, scheduling, documentation, and patient communication. Hundreds of AI-enabled medical devices have been authorized by the FDA, most through regulatory pathways that don’t require new randomized trial data. That doesn’t automatically make them unsafe — but it does mean widespread adoption is happening alongside evaluation, not after it.

For physicians, the shift is already here. The question is whether we’re being honest about what it actually involves.

The burden relief is real — so is the responsibility

Documentation fatigue isn’t a complaint. It’s a clinical problem. When a physician is spending more time navigating a chart than listening to a patient, something has already gone wrong. Ambient AI tools, when they work well, restore something that quietly eroded over years of EHR implementation: the sense that you’re actually in the room with someone.

That matters. Patients notice it too, as mine did.

But the same patient who appreciated the conversation asked about the recording. She wanted to know what was being captured, where it went, who had access to it. And the honest answer is that the standards around those questions are still being written. Consent practices vary by institution, by physician preference, sometimes by nothing more than whether someone thought to mention it. That’s not a framework. That’s improvisation.

Medicine runs on trust. And trust, once a patient starts wondering what’s happening in the background of their own appointment, is hard to rebuild in the same visit.

The governance gap is the real problem

The technology is not the issue. The issue is that we haven’t built the infrastructure around it yet.

Liability when an AI-generated note contains an error — who owns that? Informed consent for ambient recording — what does that look like, and who’s responsible for obtaining it? Data privacy for what gets captured in an exam room — where does it go, how long is it stored, and under what circumstances can it be accessed?

These aren’t hypothetical questions. They’re questions patients are already asking, and physicians are answering them individually, inconsistently, without institutional guidance. The AMA has begun addressing some of this, including concerns about AI-generated physician impersonation and synthetic media. That conversation sounds distant until you realize it’s the same underlying issue: AI is now inside the physician-patient relationship itself, and we don’t yet have agreed-upon rules for how it should behave there.

Physicians are carrying this whether we chose to or not

Most patients aren’t reading FDA authorization summaries or validation studies. They’re deciding whether to trust the person in front of them.

That means whatever AI does or doesn’t do correctly in that encounter, the physician is accountable for it. The ambient scribe that mishears a medication name, the note that drops a nuance that mattered — that’s not a software problem to the patient. It’s a physician problem.

That’s not an argument against using these tools. It’s an argument for being clear-eyed about what using them actually means. We are not passive users of AI in the exam room. We are the last line of review, the ones who sign the note, the ones the patient is looking at when they ask what’s being recorded.

That’s a real responsibility. And right now, most of us are carrying it without much institutional support.

Where this goes

AI is not going away, and I’m not suggesting it should. Used thoughtfully, these tools can give physicians something that’s been slowly taken from them — time and attention inside the clinical encounter. That’s worth pursuing.

What’s interesting is that the market itself is starting to signal something. Doximity recently released their AI scribe free of charge to all verified U.S. physicians and other clinicians. It records the encounter, generates a structured note, and then discards the original recording, which is a meaningful design choice when patients are already asking what’s being stored. For physicians in smaller or independent practices who’ve been priced out of the enterprise ambient scribe market, that accessibility matters.

But there’s a harder conversation underneath that. Basic transcription is commoditizing, and the real competition is shifting to what happens after the transcript — how well these tools structure documentation in ways that support coding, complexity capture, and revenue cycle performance. That’s not inherently wrong, but it’s worth being clear-eyed about. The pitch started as burnout relief. It’s becoming something more complicated.

Which brings me back to what I actually care about: The note still has to be read. Whatever the platform, whatever the price point, the physician signs it. That hasn’t changed, and it shouldn’t.

We need honest conversations about what adoption looks like on the ground — not the version where everything works seamlessly and patients are fully informed. The real version, where a patient asks mid-visit if she’s being recorded, and you have to decide in the moment how to answer that in a way that’s both truthful and doesn’t erode the trust you spent the last 10 minutes building.

The future of AI in medicine won’t be defined by algorithms. It’ll be defined by what physicians do with them — and whether the systems around us finally catch up to the responsibility we’re already carrying.

How are you using AI? Share in the comments.