Can Ambient AI Handle Emergency Medicine?

Ambient AI scribes have arrived in the emergency department (ED), and they are not leaving. The technology, which passively records clinician-patient conversations and drafts a structured clinical note, has moved from pilot project to institutional mandate at a pace that surprised even its early adopters. The Permanente Medical Group gave 10,000 physicians access in late 2023 and saw nearly 1,000 of them logging more than 100 encounters each within weeks.1 Stanford, Mayo, UCLA, and Mass General Brigham have since followed with their own deployments. But nearly everything we know about these tools comes from outpatient clinics. One patient, 1 problem, a quiet room, a closed door. The emergency department doesn't work like that. Everything, All at Once Consider what an emergency physician does in a single shift. A 23-year-old with a pulmonary embolism that speaks Spanish. An 85-year-old with altered mental status and no reliable historian. A 6-year-old with a forearm fracture and a terrified parent. A 40-year-old in psychiatric crisis. Each encounter demands different medical knowledge, a different communication approach, and entirely different documentation. A cardiologist documents variations on a theme. An emergency physician writes a fundamentally different note for every patient, from scratch, dozens of times per shift. That workload helps explain why emergency medicine consistently leads all specialties in burnout. National surveys estimate the rate at roughly 65%, well above the average of 45% for all physicians, and unlike most fields, those rates remain elevated even after adjusting for hours worked.2,3 When asked what drives it, documentation burden reliably tops the list. A multicenter study in JAMA Network Open found that burnout dropped from 51.9% to 38.8% after just 30 days of ambient AI scribe use.4 Those results came from ambulatory settings, but they raise an obvious question: if documentation relief can move the needle that much in a clinic, what could it do in the specialty where the burden is heaviest? The First ED-Specific Data The early returns are arriving. Our group at Stanford recently published one of the largest ED ambient scribe studies to date, covering 8,740 encounters, of which 976 used the AI tool.5 The adoption pattern was revealing: physicians reached for it most often with lower-acuity patients and avoided it when interpreters were involved. When they used it, the tool was associated with shorter documentation time and less total EHR time. For straightforward visits, it produced a usable first draft of the History of Present Illness (HPI) and made a dent in on-shift charting time. A Mayo Clinic comparison added important nuance. Morey and colleagues put AI scribes head-to-head with human scribes across 710 encounters and found that the AI version was associated with more time in the notes section, more physician-contributed characters per note, and comparable or lower quality scores.6 In other words, reviewing and editing an AI draft takes real work. It is not the effortless handoff the marketing implies. A separate survey across 4 EDs at Brown found that while most physicians preferred AI-assisted documentation over charting alone, confidence in accuracy remained limited, particularly for complex encounters.7 Where the Cracks Show These challenges are not hypothetical. Emergency departments are acoustically chaotic: overlapping conversations, alarms, overhead pages, and constant interruptions. Speech recognition struggles with all of it.8 Stanford data showed physicians avoided the tool for interpreter-mediated encounters, which represent a substantial share of volume in linguistically diverse EDs.5 Current AI scribes are built for English-language, 1-on-1 conversations. If the technology cannot handle multilingual, multispeaker encounters reliably, it risks becoming a tool that works best for the patients who already face the fewest barriers to care. Psychiatric presentations raise their own concerns. A patient in psychosis may produce tangential or disorganized speech that carries real diagnostic value, but distinguishing clinical signal from delusion requires judgement the AI doesn't have. A paranoid patient may become distressed by the presence of a recording device. Whether the scribe faithfully captures that kind of encounter or smooths it into a coherent narrative that strips out the pathology, is an open and consequential question. Patients with dementia present a related problem: their spoken histories are often unreliable, and the information that matters most comes from caregivers or the prior chart. And then there are the sickest patients. A patient on BiPAP for respiratory failure is not narrating symptoms. A trauma resuscitation involves 6 people talking at once. A cardiac arrest generates documentation from vitals, labs, and clinical reasoning. Almost none of that is captured in a spoken conversation. For the highest-acuity encounters in the ED, ambient scribes have very little to work with. The Road Ahead The technology is moving fast, and emergency medicine has an opportunity to shape where it goes. Practical guidance on consent for ambient recording, particularly involving psychiatric patients, minors, and victims of violence, would help standardize practices that currently vary by institution. Multilingual performance standards need to be established before these tools scale further into diverse departments. And the ED needs its own evidence base: research on documentation accuracy during high-acuity encounters, performance in noisy environments, and impact on after-shift charting. Ambient AI scribes address a real problem. The documentation burden in emergency medicine is unsustainable, and the burnout data makes the case for intervention as clearly as anything in the literature. But the ED is going to stress-test these tools in ways that outpatient clinics never will. The noise, the language barriers, the psychiatric presentations, the critically ill are not edge cases. They are the job. The question is whether we're building for the emergency department as it actually is, or for a version of it that looks a lot more like a clinic.