Last month I sat in on yet another demo for a new ambient AI scribe tool, the kind that’s supposed to turn a recorded visit into a clean note by the time you walk out of the room. It was good. Really good. Good enough that I left thinking about all the charting I wouldn’t have to do that night, and also thinking about something else, which is that the room the AI heard and the room I was sitting in were not quite the same room.
I still don’t think AI is going to replace what I actually do for a living anytime soon. That’s not a defense of my job. It’s closer to an observation about what gets lost between a conversation and a transcript of that conversation, and why that gap matters more in medicine than almost anywhere else.
Patients don’t present in an order that helps you build a differential diagnosis. Their symptoms don’t arrive pre-sorted into medical concepts either. There’s usually a real, logical explanation waiting underneath all of it, but the road to that explanation runs through a conversation that wanders, backtracks, and sometimes contradicts itself. An AI system can only work with what actually gets said. When the signal isn’t in the words, there’s nothing for it to find.
This past year, colleagues at Stanford and I started wondering whether patient-reported outcome measures (PROMs) could be pulled from ambient scribe transcripts, since they record every word of a visit. PROMs are sometimes the only way to track outcomes, and they’re usually the hardest thing to collect without handing a patient yet another form. We built an AI tool to test this, using PROMs we rely on in otolaryngology. What we found was a lot of missing information. The AI hadn’t hallucinated or mixed anything up. The outcome we wanted just hadn’t been spoken out loud.
Real physician voices, twice a week
Free, and one click to unsubscribe.
This is just one study. I don’t want to lean on it harder than it can hold. It doesn’t mean AI can never pick up on what a patient leaves unsaid; voice models can already flag roughness or strain from audio alone, and I could see a future system noticing a slow decline across years of visits the way a doctor eventually does. What it still won’t have is the why.
Here’s what I mean. I had a patient tell me his hoarseness was “not that bad, probably just a lingering cold.” He’d been dealing with it for months and was scared of what an honest answer might turn out to be. His transcript says exactly that: not that bad. What it doesn’t say is that he wouldn’t look at me while he said it, that he’d waited four months to book the appointment, and that his voice sounded far rougher than any cold produces. That same week, I had another patient describe his voice loss as the worst thing that had happened to him all year, delivered in a voice that sounded, if anything, stronger than mine. Two patients, two transcripts, both technically accurate, and both nearly useless on their own for telling me who I actually needed to worry about.
I don’t think that’s a limitation AI eventually engineers its way out of. Figuring out when someone is underselling a symptom, or overselling one, or just doesn’t have the words for what they’re feeling, isn’t really a data problem. It’s something closer to pattern recognition built out of years of sitting across from people who didn’t behave the way any training set would predict.
Physicians have something AI doesn’t have yet, and maybe never will: not a bigger memory, but intuition, that gut sense that something’s off before you can say why, the one built from every patient who taught you something that data alone never would have. That’s the part of this job that’s hardest to explain to someone who hasn’t done it, and it’s also the reason I’m not particularly worried about being replaced. Let the AI take the typing. I’ll keep the room.
Nancy Jiang is an otolaryngologist.
