Do AI scribes help equally across patient languages? A 209,009-encounter study found uneven associations
Same-day note closure was 10.2 percentage points higher with an AI scribe in non-English language-concordant visits and 1.7 points higher in English visits after adjustment. Interpreter-mediated estimates were too imprecise for a conclusion, and the study did not test note accuracy or patient outcomes.
The short version
- What happened
- The peer-reviewed study analysed 209,009 in-person encounters involving 117,121 patients and 286 clinicians across 14 adult primary-care clinics; AI scribes were used in 47,743 encounters.
- Who it affects
- Clinicians may finish documentation sooner in some language-concordant visits, but health systems should not assume the same benefit in interpreter-mediated care.
- What remains unproven
- This was a retrospective cross-sectional study at one Boston academic health system using two products, limiting generalisability.
Keep in mind
This was a retrospective cross-sectional study at one Boston academic health system using two products, limiting generalisability.
Clinicians self-selected scribe access and encounter-level use; adjusted associations cannot establish causal effects.
Interpreter-mediated groups were small, producing wide confidence intervals and inconclusive estimates.
The dataset lacked product identity, adoption timing, visit duration, insurance status, detailed comorbidity measures and independent confirmation of language concordance.
Same-day note closure does not measure note accuracy, clinician workload, clinical decisions, patient experience or health outcomes.
Source analysis · 1 linked source. Primary analysed source date: 6 November 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The finding: faster note closure was not distributed evenly across communication settings
A peer-reviewed JAMA Network Open study published on 6 November found that ambient AI scribe use was associated with a higher probability of closing a primary-care note on the day of the visit, but the size and certainty of that association depended on how clinician and patient communicated. The strongest adjusted association appeared in visits where a patient preferred a language other than English and the clinician spoke that language: same-day closure was 10.2 percentage points higher with a scribe, with a 95% confidence interval from 7.4 to 12.9 points.
For English language-concordant visits, the adjusted association was much smaller at 1.7 percentage points, with a 95% confidence interval from 0.9 to 2.6 points. Estimates for visits involving video, telephone or in-person interpreters, or no documented interpreter despite nonconcordant language, were too imprecise for a firm conclusion. The study therefore does not support a claim that scribes improve documentation equally for everyone. It identifies a large association in one bilingual setting and an evidence gap precisely where conversations may involve multiple speakers, switching languages and more complex acoustics.[1]
What the 209,009-encounter denominator contains
The researchers analysed 209,009 in-person encounters from May 2025 through May 2026 at 14 adult primary-care clinics in one Boston academic medical centre. Those encounters involved 117,121 patients and 286 clinicians. AI scribes were used in 47,743 encounters and not used in 161,266. The study began four months after initial deployment so that the technology had time to spread through the system. Two scribe products were available, although the dataset did not identify which product was used in each encounter.
Most visits—193,035, or 92.4%—were English language-concordant. Another 9,638 were non-English language-concordant. The nonconcordant groups were far smaller: 1,076 encounters used a video interpreter, 2,182 a telephone interpreter, 734 an in-person interpreter and 2,344 had no interpreter documented. Those denominators explain why the interpreter estimates are uncertain. A system may contain hundreds of thousands of visits overall while still providing too few observations to estimate effects reliably for specific language-access workflows.[1]
The distinction that matters
What the study measured—and what it did not
The comparison links scribe use to one documentation outcome across six communication settings.
Exposure
Clinician chose whether to use an AI scribe
47,743 of 209,009 encounters
Outcome
Check whether the note was signed that day
Timeliness, not accuracy
Subgroups
Separate language and interpreter settings
Uneven, often imprecise estimates
Inference
Adjust observed differences statistically
Association, not causation
How the comparison was made—and why it cannot establish causation
Clinicians self-enrolled for scribe access on a rolling basis and decided whether to use the tool during each encounter. The exposure was identified from electronic health-record data. The outcome was narrow and operational: whether the clinician signed the note on the same calendar day as the visit. Communication type was recorded in the health record by clinicians, combining patient language preference, clinician language concordance and interpreter modality. It was not independently reassessed by the research team.
The authors used mixed-effects logistic regression with patient- and clinician-level random intercepts. The model adjusted for age, clinician type, clinical complexity, clinic setting and calendar time, and included an interaction between six communication categories and scribe use. Sensitivity checks used clinician fixed effects and within-clinician comparisons among clinicians who had both scribe and non-scribe encounters. Those analyses reduce some observable differences, but they cannot remove unmeasured selection. A clinician may choose a scribe for visits, patients or days that differ in ways the dataset did not capture.[1]
The unadjusted pattern was large in bilingual visits and modest in English visits
Before adjustment, 82.9% of non-English language-concordant encounters using a scribe had same-day closure, compared with 72.1% without one. English language-concordant encounters showed 80.0% with a scribe and 78.4% without. The adjusted estimates closely preserved the contrast between a roughly ten-point association in the bilingual group and a small association in the English group. The overall interaction across communication categories was statistically significant, indicating that one pooled average would hide meaningful variation.
That variation should not be interpreted as proof that the technology works better in every non-English language. The paper groups languages together and does not report product-specific performance, time since adoption or the duration of each visit. Only one of the two tools supported non-English, non-Spanish documentation without an interpreter, which may have influenced who used a scribe and when. Insurance status, detailed comorbidity measures, bilingual-patient classification and exact encounter timestamps were unavailable, leaving plausible residual differences between compared visits.[1]
Same-day closure matters, but it is not a measure of note quality or safer care
Closing notes promptly can reduce after-hours administrative work and make information available sooner for follow-up, referrals and other clinicians. It is therefore a useful operational outcome. But a note can be signed quickly and still contain omissions, transcription errors, incorrect speaker attribution or unsupported clinical statements. The study did not compare note accuracy, completeness, edits, diagnostic decisions, clinician attention during visits, patient understanding, follow-up care or health outcomes. It also did not measure whether the scribe reduced total documentation time.
Interpreter-mediated visits require particular care. Multiple speakers and language switching can challenge speech recognition and attribution, while a professional interpreter has a defined role that a documentation system should not obscure. The inconclusive results do not show that scribes harm those visits; they show that this dataset could not establish a clear association. Health systems should preserve interpreter services and evaluate scribe performance inside those workflows rather than assume that an efficiency gain measured in English or language-concordant care transfers automatically.[1]
What health systems can do now—and what evidence would change the assessment
A responsible deployment would monitor results by language, interpreter modality, clinic and product rather than publish one system-wide average. Hospitals should audit note accuracy and correction burden, ask patients and interpreters about the experience, and verify that consent and privacy information is understandable in the languages served. Bilingual workforce development and professional interpretation remain separate investments: an ambient scribe records a conversation; it does not replace the human communication needed to conduct the visit safely.
Confidence would rise with randomised or strong quasi-experimental studies that assign scribe access prospectively, identify the product and language, and measure note quality, clinician time, patient trust, clinical decisions and downstream outcomes. Larger multisite samples are needed for each interpreter modality. Evidence would weaken if apparent speed gains disappeared after accounting for adoption timing or if faster closure came with more corrections or clinically important errors. For now, the study shows an uneven association in one health system—not a universal efficiency effect or proof of equitable clinical benefit.[1]