13 November 2026 · Medows
AI Scribes' Blind Spot: The Interpreter
AI scribes sped up notes across 209,009 Boston primary care visits, but the gain vanished when an interpreter was in the room.
The interpreter is on a phone propped against the monitor. The patient answers in her own language, the phone answers in English, you ask the next question, and the phone turns it around again. Every exchange takes three turns instead of one. This is the visit where you most want help with the note: it runs longest, and you are least sure you caught everything.
It is also the visit where the AI scribe, on the best evidence available, does nothing measurable.
What the study measured
A cross-sectional study published in JAMA Network Open on November 6, 2026 looked at 209,009 encounters across 14 adult primary care clinics at an academic medical center in Boston, from May 2025 to May 2026. It covered 117,121 patients and 286 clinicians. An AI scribe was used in 47,743 of them.
The outcome was mundane and well chosen: same-day note closure. Did the note get finished before the doctor went home.
Encounters were split by language. English language-concordant visits made up 193,035 (92.4%). Visits with a patient who prefers a language other than English (NELP), seen by a clinician who speaks that language, numbered 9,638 (4.6%). The remaining 3% were interpreter-mediated or had no interpreter recorded.
The gain landed where the barrier was not
Unadjusted, NELP language-concordant visits closed 82.9% of notes same day with a scribe, versus 72.1% without. English language-concordant visits went from 78.4% to 80.0%.
Adjusted, reported as average marginal effects: NELP language-concordant visits gained 10.2 percentage points (95% CI, 7.4 to 12.9; P < .001). English language-concordant visits gained 1.7 points (95% CI, 0.9 to 2.6; P < .001). The interpreter-mediated categories were inconclusive, confidence intervals spanning zero. The interaction across categories was significant (P < .001).
Read that ordering again. The largest benefit went to the bilingual doctor talking directly to the patient, in the patient's language, with nobody in between: six times the English-visit effect. The moment a third voice enters the room, the measurable gain disappears.
The authors are careful about why. The interpreter groups are small (734 to 2,344 encounters), clinicians chose whether to use the scribe, only one of the two products handled a non-English, non-Spanish visit without an interpreter, and the associations are not causal. Their conclusion still lands: extending the benefit to interpreter-mediated encounters is, in their words, essential to preventing a new form of digital health inequity.
What is inside the notes nobody checked
Closure speed is not accuracy. A separate research letter in JMIR Medical Informatics (July 28, 2026) scripted 20 interpreter errors into five simulated encounters and ran them through two commercial scribes. Vendor A carried 55% (11/20) of those errors into the note. Vendor B carried 60% (12/20). Omissions traveled furthest: 67% (6/9) and 78% (7/9).
The split that matters most is whose mouth the error came from. Errors originating in the patient's speech propagated at 80% (8/10) and 100% (10/10). Errors from the clinician's speech propagated at 30% (3/10) and 20% (2/10). Simulated, small, descriptive, and still the clearest signal anyone has published. The scribe is most faithful to the voice it was built around, least faithful to the patient's.
What a point tool cannot show you
Side by side, the two findings describe a tool that saves time in the visits with the least friction and quietly reshapes the record in the visits with the most.
Neither finding is visible from inside the scribe. It hands over a note, drafted and ready. It does not mention that this encounter belongs to the category where its own benefit has never been demonstrated, or that the omission it inherited came from the patient's half of the conversation. The doctor signing is the only check, and the note arrives looking finished.
That gap is the argument for a workspace rather than a point tool. A tool optimizes its own output. A workspace holds the context around it: who was in the room, which language, who interpreted, and therefore what is less certain in this note than in the last one. Verifiability is not a button on a draft. It is knowing which drafts deserve a second read.
The deployments are not waiting for the evidence. On November 11, 2026, India's National Health Authority named EkaCare its technology partner for an ambient AI voice scribe in public health facilities, with company-reported support for more than 15 Indian languages and mid-conversation switching, benchmarked on 6,550 real consultations. A verified doctor must review and approve every note before it is saved. That review step is the entire safety model, and the published evidence on what to look for in a multilingual note runs to two studies.
Sources
- Artificial Intelligence Scribes and Same-Day Note Closure by Patient Language Preference and Interpreter Modality, JAMA Network Open, November 6, 2026 (full text on PMC)
- Propagation of Interpreter Errors by Ambient AI Scribes: Study Using Simulated Clinical Encounters, JMIR Medical Informatics, July 28, 2026
- NHA Chooses EkaCare for AI Voice Scribe Solution, Analytics India Magazine, November 11, 2026
Medows is a clinical AI workspace for the doctor on rounds. Learn more or write to alapan@medows.ai / alapanx@gmail.com.