Your AI Scribe Lists 100+ Languages. It Still Isn’t Your Interpreter.

Every major AI scribe now advertises multilingual support. Abridge lists 28 languages. Nabla lists more than 35. Suki says more than 80, Freed more than 90, and Heidi more than 110. For a practice with many Spanish-, Vietnamese-, or Haitian Creole-speaking patients, those numbers suggest the language barrier in documentation has been solved.

It hasn't. The research published in 2025 and 2026 shows three problems. Clinicians quietly stop using the scribe when an interpreter is in the room. When they do use it, the scribe often writes the interpreter's mistakes into the chart. And the patient-language instructions that some scribes now produce count as machine translation under federal civil rights rules, which require a qualified human translator to review critical content.

This guide covers what the evidence says, what the federal language-access rules require in 2026, and a visit workflow that lets you use a scribe with limited-English-proficiency (LEP) patients without creating a new kind of error.

The quiet equity gap: scribes get switched off for interpreted visits

The first problem is the visits where the scribe never gets used.

• Emergency department, 2026. Preiksaitis and colleagues tracked an ambient scribe rollout at a high-volume US academic ED (Annals of Emergency Medicine, May 2026;87(5):569–574). The scribe was used in only 976 of 8,740 eligible encounters (11.2%). Encounters that needed an interpreter, along with ambulance arrivals and higher-acuity patients, were significantly less likely to get the scribe. Physicians chose which visits to use it on, and they skipped the interpreted ones.
• Physician interviews, 2025. In a JAMA Network Open qualitative study of ambient scribe users (Shah et al., 2025;8(3):e251904), every one of the 21 comments coded under "accessibility" was neutral or negative (81% negative, none positive). One physician said: "English as a second language is definitely a potential barrier, so I'm not using it if I have a translator that I'm working with for a patient."

This matters because the scribe's benefits are well documented: less after-hours charting, shorter time in the EHR, and more eye contact (see our rollout guide). If those benefits go only to English-speaking visits, the visits that already take longest to document get no help. Interpreted visits also tend to be the ones clinicians most want to finish quickly. So the first thing to measure is how often your scribe is used, broken out by patient language.

When the scribe is on, it trusts the interpreter

The second problem is more serious. With the scribe running and an interpreter present, the scribe hears two versions of what the patient said: the patient's own words, and the interpreter's English rendering. The evidence suggests scribes mostly document the English rendering, including its mistakes.

Rabotin et al., JMIR Medical Informatics, July 28, 2026 (2026;14:e88734). A UCLA-affiliated team wrote five Spanish–English clinical scenarios containing 20 deliberate interpreter errors: omissions, substitutions, and additions. They recorded the scenarios with native Spanish speakers and ran them through two ambient scribes. One was a commercial scribe with explicit Spanish support; the other used GPT-4o. Two reviewers checked every note and agreed 97.5% of the time.

• Overall, interpreter errors ended up in the note 55% (11/20) and 60% (12/20) of the time.
• Omissions were the worst: 67% and 78% propagated.
• The direction of interpretation was the biggest factor. When the interpreter made the error while rendering the patient's Spanish into English, it reached the note 80% (8/10) and 100% (10/10) of the time. When the error was in the clinician's English rendered into Spanish, only 30% and 20% propagated.

In other words, the part of the visit the scribe was least likely to get right is the patient's history, which is the part the note depends on most.

The examples show how this happens. A patient said "I do not hear that well." The interpreter left it out, and one scribe documented "no other hearing problems", a negative finding that nobody stated. In another scenario the interpreter dropped a child's wheezing, and the note recorded only a nighttime cough. The scribes did sometimes get it right: in a few scenarios they recovered the original Spanish and bypassed the faulty interpretation, and one scribe kept a correct "every 6 hours" dosing instruction even though the interpreter said "every 4 hours." But this happened inconsistently, so you can't rely on it.

The study is small and simulated, and its authors call for larger real-world studies. The direction of the finding is still clear: in an interpreted visit, the scribe's note is only as accurate as the interpretation, and sometimes less accurate.

Mixed-language visits break speech recognition in specific ways

Many visits aren't "Spanish" or "English." They're both: a bilingual patient switches languages mid-sentence, a family member answers in Mandarin, the clinician says drug names in English. A UC Irvine team tested exactly this with Abridge, using 24 reenacted primary care conversations: 12 Mandarin–English and 12 Spanish–English (Hu et al., medRxiv preprint, June 2026; not yet peer reviewed).

• Spanish–English went fairly well: a median mixed error rate around 4%, and language switches were detected reliably.
• Mandarin–English was worse: a median around 9%, with much more variation (one excerpt reached 67%). Large chunks of text were deleted at or near the points where speakers switched languages.
• The specific errors are the kind a quick skim misses: "mi presión" (my blood pressure) transcribed as "depresión" (depression); "cystatin C" as "cistatin sí" (the "C" heard as Spanish sí); "losartan" split into "los zartan"; amlodipine as "enlodipine"; and in Mandarin, the words for "fall" and "sleep" confused.

Some of these, like blood pressure becoming depression, put a wrong diagnosis into the chart. That makes them more dangerous than ordinary typos. They are also the same failure type described in our accuracy and hallucination guide: fluent, plausible text that is wrong.

What the vendors actually say (October 2026)

Below is each vendor's documented language support. All of these come from the vendors' own help and product pages; none of them publishes accuracy figures for individual languages.

• Abridge: 28 languages, detected automatically. Every note is generated in English no matter what language was spoken. Abridge's help center says it "works well when an interpreter is facilitating the visit" and that "everything spoken during the recorded conversation contributes to the generated clinical note." The original audio is linked to the note so you can check what was actually said, and that link is the most useful review tool for interpreted visits.
• Nabla: more than 35 languages, including Haitian Creole, Khmer, Punjabi and four regional Arabic variants. You set a main and a secondary encounter language in settings. Notes are generated in English or French. A toggle lets you copy patient instructions in the patient's native language.
• Suki: more than 80 languages for ambient sessions, according to its developer documentation, with the clinical note generated in English.
• Freed: more than 90 languages, with an English SOAP note. Freed says it "can keep patient instructions in the language spoken during the visit."
• Heidi Health: transcription in more than 110 languages, according to its product page.

Two things to notice. First, a long language list is not evidence that each language works well. Abridge, for example, separates out a set of "high-performing" languages that its team has specifically benchmarked and optimized from the rest of its list. Ask any vendor which languages it has actually measured, and what the error rates were. Second, several tools now produce patient-facing text in the patient's language. That is useful, and it is also where the federal rules apply most directly.

The federal rules: Section 1557 still applies in 2026

There has been real confusion about language access since 2025, so here is what changed and what didn't.

What changed: Executive Order 14224 (March 1, 2025) designated English as the official language and revoked the executive order that had required federal agencies to maintain language-access programs. Federal agency language-access guidance has been rolled back since then.

What didn't change: the Section 1557 nondiscrimination regulations at 45 CFR part 92. The 2024 final rule's language-access sections are in force. In June 2026, HHS published a notice that vacated parts of the 2024 rule after Tennessee v. Kennedy (S.D. Miss., Oct. 22, 2025). That decision covered only the gender-identity provisions. The language-access section, § 92.201, was not affected.

What § 92.201 requires of a covered entity:

• Meaningful access for each person with limited English proficiency, including companions. Language services must be free, accurate, timely, and must protect the patient's privacy and independent decision-making.
• Qualified interpreters when interpretation is needed, and qualified translators when translation is needed.
• No reliance on minors, unqualified adult companions, or unqualified staff to interpret, except temporarily in an emergency.
• Minimum quality standards for video and audio remote interpreting: real-time, clear transmission, and trained users.
• Machine translation: if a covered entity uses machine translation for text that is "critical to the rights, benefits, or meaningful access" of a person with limited English proficiency, where "accuracy is essential," or where the material is "complex, non-literal or technical," then "the translation must be reviewed by a qualified human translator."

Does this cover a small practice? Very likely yes. The 2024 rule changed HHS's long-standing position and now treats Medicare Part B payments to providers as federal financial assistance, so a practice that bills Part B is generally a covered entity. Some obligations depend on size: practices with 15 or more employees must designate a Section 1557 coordinator and adopt grievance procedures. The language-access duties themselves apply regardless of size. The compliance dates for written language-access policies, staff training, and the multilingual Notice of Availability passed on July 5, 2025.

There is new federal law on the way, too. The SPEAK Act, signed February 3, 2026 as part of the Consolidated Appropriations Act, 2026, requires HHS to publish guidance on language access in telehealth by February 2027. That guidance is expected to cover integrating qualified interpreters into telehealth visits and multi-party multilingual calls. If you use a scribe on video visits, watch for it (see also our telehealth scribe guide).

Why this matters for scribe features

Look at the scribe features above against § 92.201(c)(3). Discharge or after-visit instructions written in the patient's language by an AI model are machine translation. Medication directions, warning signs that mean "go to the ER," and follow-up instructions are about as clear an example of text where "accuracy is essential" as you could find. On the plain text of the rule, those instructions need review by a qualified human translator before the patient relies on them.

The research backs up the rule's caution. A 2024 Pediatrics study (Brewster et al., 154(1):e2023065573) compared professional, Google Translate, and ChatGPT-4 translations of 20 pediatric discharge instructions. For Spanish and Portuguese, the machine translations rated similarly to professional ones. For Haitian Creole, potentially clinically significant errors appeared in 33.3% of ChatGPT translations and 23.3% of Google Translate translations, and professional translations scored significantly better. Several scribes list Haitian Creole as a supported language. The languages your community speaks least often are where machine output is weakest and where human review matters most.

The scribe's English note is a different situation. It is an internal clinical record, not a document given to the patient, so (c)(3) applies less directly. But the accuracy duty in § 92.201(b) and ordinary documentation standards still apply. What the Rabotin study found, an invented "no hearing problems," is exactly the kind of error a payer audit or a malpractice claim would read as fact.

This is our reading of the regulation's text, not legal advice. Practices with real exposure should confirm with health-care counsel.

A visit workflow that works

1. Choose the language mode before the visit starts

For every visit with a patient who has limited English proficiency, decide which of three situations you're in.

• Language-concordant clinician. The clinician speaks the patient's language and qualifies under your language-access policy as qualified bilingual staff. This is the best case for a scribe: there's no interpreter layer to propagate errors, and the scribe only has to produce an English note from a single conversation.
• Qualified interpreter, in person, by video, or by phone. Use the scribe, but treat the patient-history sections as high-risk (step 3).
• No qualified interpreter available. The scribe is not an interpreter. Ambient scribes listen and write a note for the clinician. They don't interpret in real time for the patient, and a fluent English note doesn't mean the patient understood anything. Don't let a scribe's language list become the reason a visit goes ahead without an interpreter.

2. Make sure the scribe can hear the interpreter

With phone or video interpreters, the interpreter's voice often comes out of a speakerphone or a laptop. Echo suppression and microphone placement can drop that audio, which is the same problem we describe for telehealth visits. Run a test visit with your actual interpreter vendor's setup before you rely on it. If the scribe hears only the clinician clearly, the note will be built mostly from the clinician's side of the conversation.

3. Review the note where the errors cluster

Generic advice like "review the note" isn't enough here. The research points to specific places to check:

• The HPI and review of systems. These come from the patient's speech, where 80–100% of interpreter errors carried through in the Rabotin study.
• Pertinent negatives. Look for "denies…" or "no…" statements you don't remember the patient making. An omission by the interpreter can turn into an invented negative in the note.
• Medications, doses, and frequencies. Mixed-language speech garbles drug names ("los zartan"), and interpreters substitute numbers.
• Words that sound alike in two languages. Presión and depresión is the standard example. If a new diagnosis appears that you didn't discuss, check the audio.

If your scribe links note sentences to the source audio (Abridge's linked evidence does this), use it on anything that drives a diagnosis or an order.

4. Don't hand out machine-translated instructions without human review

If your scribe produces patient instructions in the patient's language:

• Treat them as a draft that needs qualified human review when they cover medications, warning signs, procedures, or anything else critical, which is most of what instructions cover.
• Build a library of professionally translated templates for your most common instructions (new diabetes medications, post-procedure care, return precautions) in your top languages. Use those instead of AI-generated text where you can.
• Use the interpreter for teach-back before the patient leaves. Teach-back catches what a translated handout misses.
• Patient-portal messages count too. If you use AI-drafted portal replies in other languages, the same rule applies (see our patient inbox workflow).

5. Document the interpreter yourself

Scribes summarize the conversation into the standard note sections. They don't reliably record who interpreted, how (in person, video, or phone), and the interpreter's ID number. That information supports your § 92.201 compliance and protects you if the accuracy of the visit is questioned later. Add it to your note template as a required field, or use a smart phrase.

6. Get consent in the patient's language

Recording consent only counts if the patient understood what they agreed to. Have the interpreter deliver your scribe disclosure and consent script, or use a professionally translated written version. State disclosure and recording-consent requirements apply no matter what language the patient speaks; see our guides to patient consent and recording laws and state AI disclosure laws. Patients who say no need a clear path to a visit without the scribe, and that choice shouldn't make the visit any worse.

7. Measure use by language

Use the ED data as your warning. Every quarter, pull scribe use by patient preferred language and by interpreter use. If interpreted visits use the scribe far less, ask clinicians why. The cause is usually something fixable, such as audio setup, review burden, or lack of training on the multilingual workflow.

Questions to ask your scribe vendor

Add these to the evaluation checklist in our guide to choosing a scribe:

1. Which of your listed languages have you benchmarked, and what were the error rates? Ask specifically about the top three non-English languages in your patient panel.
2. How does the model weight interpreter speech against the patient's original-language speech? Can it flag places where the two disagree?
3. Can I see the original-language transcript and audio linked to each note sentence?
4. Are patient instructions in other languages machine-generated? Can I turn that feature off, or force a human-review step?
5. How do you handle code-switching within a single sentence or between speakers?
6. Does your BAA and data-retention policy cover third-party interpreter audio captured in the recording? (Our HIPAA basics guide covers what to look for in a BAA.)

Where this matters most

Practices that see a lot of patients with limited English proficiency, including community family medicine, pediatrics, and federally qualified health centers, get the most from a scribe that handles languages well, and carry the most risk from one that doesn't. Pediatric practices have a second layer: when parents rely on a child to interpret, § 92.201 prohibits that outside emergencies, and the child is not a substitute for a qualified interpreter. Our adolescent confidentiality guide covers the other scribe risks in pediatric visits.

To compare scribes for your setting, see AI medical scribes for family medicine, AI medical scribes for small clinics, the pediatricians specialty page, or browse all AI medical scribe tools and patient communication tools.

The bottom line

• Multilingual scribes are useful. A language-concordant clinician using a scribe that writes an English note is a real improvement for patients with limited English proficiency.
• In interpreted visits, the scribe documents the interpretation, mistakes included. Check the patient history, the negatives, and the medications more closely than usual.
• Patient-language instructions from AI are machine translation. Under 45 CFR 92.201(c)(3), critical content needs review by a qualified human translator, and the less common the language, the more that review matters.
• The scribe is never the interpreter.
• Track who gets the scribe. If interpreted visits are left out, the equity gap stays hidden until you measure it.

---

This article is for informational purposes only and is not legal, medical, or compliance advice. Regulations, vendor features, and language support change frequently; vendor details here reflect their published documentation as of October 2026. Verify current capabilities directly with each vendor, and consult qualified health-care counsel about your obligations under Section 1557 and state law.