Language barriers in healthcare: a clinician's guide

Language barriers in healthcare are a patient-safety problem, not an inconvenience. When a clinician and patient do not share a language, the history gets thinner, consent gets weaker, and the error rate rises: in a study of adverse events across six US hospitals, 52.4% of events involving patients with limited English proficiency were attributable to communication errors, against 35.9% for English-speaking patients, and those events were far more likely to cause physical harm (Divi et al., International Journal for Quality in Health Care, 2007). The fix is procedural: use a qualified interpreter, run the consultation in a way that survives interpretation, and document what you used.

This is a practical reference for clinicians working across a language gap — in primary care, outpatient clinics, emergency departments and telehealth. It covers who should interpret and who should never, a protocol for the interpreted consultation, what the evidence currently says about AI and machine translation, and how to record all of it in the note.

If you only have a minute, the short version:

• Default to a qualified professional interpreter. Untrained interpreters are more likely to make errors and breach confidentiality.
• Never use a child as an interpreter except in a genuine emergency; some jurisdictions prohibit it outright.
• Speak to the patient, not the interpreter — first person, short thought groups, one question at a time, no idioms.
• Confirm understanding with teach-back, because a nod is not comprehension.
• Treat machine translation as a stopgap for brief, simple exchanges, and note that accuracy is markedly worse translating into English than out of it.
• Document the language, the interpreter's identity and the modality in the note — it is a clinical and medico-legal fact about the encounter.

How much harm do language barriers actually cause?

Enough to change your practice. The best-known quantification remains the Joint Commission pilot study of adverse events in six accredited US hospitals: 49.1% of events involving patients with limited English proficiency (LEP) involved physical harm, compared with 29.5% of events involving English-speaking patients, and communication errors accounted for 52.4% of LEP events versus 35.9% otherwise (Divi et al., 2007).

That is a pilot study across six hospitals, so the precise percentages should be read as indicative rather than definitive. The direction, however, is consistent with what the primary-care literature describes: inadequate histories, unnecessary testing, poorer adherence, lower satisfaction and greater malpractice exposure when language is handled informally. The American Academy of Family Physicians' review notes that use of professional interpreters is associated with fewer communication errors, better satisfaction and fewer adverse events, and that trained interpreter use has been linked to shorter hospital stays and reduced 30-day readmission (Juckett and Unger, American Family Physician, 2014).

How common is a language barrier in practice?

Common, and increasing. Globally, the World Health Organization estimates that around one in eight people — over a billion — are on the move, with 304 million international migrants in 2024, roughly double the 154 million recorded in 1990, and 117.3 million people forcibly displaced. WHO states plainly that refugees and migrants "often experience poorer health outcomes due to language and cultural barriers, institutional discrimination and restrictive policies" (WHO, Refugee and migrant health fact sheet, March 2026).

Nationally the picture is the same. When the AAFP reviewed US practice, more than 25 million Americans spoke English less than "very well" and more than 60 million spoke a language other than English at home, per the Census Bureau's American Community Survey. Layer on health literacy — more than a third of Americans have limited health literacy, the same review notes — and the communication problem is broader than translation alone. A patient can share your language and still not share your vocabulary.

In much of the world the arithmetic is different again: the clinician is multilingual, the patient speaks one of several regional languages, and there is no interpreter service to call. That reality is why the techniques below matter more than any procurement decision — they work whether your interpreter is a certified professional on a video link or a trained member of your own reception staff.

Who should interpret — and who should never?

In order of preference: a qualified professional interpreter (in person, by telephone or by video), then a formally tested and trained bilingual member of staff. Untrained ad hoc interpreters — bilingual colleagues pulled from a corridor, family members, friends — are a last resort. Children should not interpret except in a genuine emergency.

• Professional in-person interpreter — best for complex, emotive or high-stakes consultations: breaking bad news, consent, psychiatric assessment, safeguarding. Non-verbal cues survive.
• Telephone interpreting — fast, available out of hours and across rare languages; the AAFP cites a cost of roughly $2 to $3 per minute without a contract. Weakest for non-verbal information.
• Video interpreting — a reasonable middle ground, and the only realistic route to a qualified interpreter in sign language or a rare language at short notice.
• Trained dual-role staff — a bilingual employee who has been formally tested and trained as an interpreter. Genuinely useful; see the training evidence below.
• Untrained bilingual staff — convenient and risky: untrained interpreters are more likely to make errors, violate confidentiality and increase the risk of poor outcomes (AAFP).
• Family members — may carry their own agenda, offer unsolicited advice, cannot guarantee confidentiality, and often limit the scope of questioning on sexual, psychiatric or intimate-partner topics. If the patient insists, the AAFP advises documenting that in the chart.
• Children — should not be used except in emergencies, given their limited grasp of adult and clinical material; several US states forbid the practice.

There is also a legal floor in many systems. In the United States, Title VI of the Civil Rights Act requires interpreter services for patients with limited English proficiency in organisations receiving federal financial assistance (with an exception for Medicare Part B), and the AAFP review states that failing to provide them where necessary is "considered discriminatory and illegal". The same review is blunt about the economics: because Medicaid, Medicare and most private insurers generally do not reimburse interpretation, it functions in most states as an unfunded mandate. Know your own jurisdiction's rule — it is usually stricter than clinicians assume.

How do you run a consultation through an interpreter?

Brief the interpreter before you start, speak directly to the patient in the first person, deliver one short thought group at a time, keep to three key points or fewer, avoid idiom and jargon, and close with teach-back. Allow extra time and seat the interpreter next to or slightly behind the patient so eye contact stays between you and the patient. These steps come from the AAFP's practical guidance and cost almost nothing to adopt.

Before the consultation

• Identify the patient's preferred language at registration, not in the room — and flag it in the record so the next clinician does not rediscover it.
• Book the interpreter ahead where the visit is planned; interpreted visits take longer, so schedule the extra time rather than absorbing it.
• Meet the interpreter briefly first: give the clinical background, agree the goal of the consultation, and establish rapport.
• Position the interpreter beside or slightly behind the patient, so the conversation runs between you and the patient.

During the consultation

• Address the patient directly and use first-person statements — "I would like to examine your abdomen", not "tell her I want to examine her".
• Speak in short sentences or thought groups, ask one question at a time, and pause for interpretation after each.
• Limit yourself to three key points or fewer; drop idioms, acronyms, jargon and humour, none of which interpret reliably.
• Prefer sentence-by-sentence interpretation, which prevents summarising and tangents.
• Watch the patient, not the interpreter, while the patient speaks — you are still looking for the clinical signs you would normally read.
• Check understanding with teach-back or "show me": patients will nod out of politeness, and a nod is not comprehension.
• Allow extra time overall, and build in roughly a 10-minute break for every hour of continuous interpretation (AAFP) — interpreting is cognitively punishing and quality degrades.

After the consultation

• Record the interpreter's name and the modality used in the note.
• Debrief with the interpreter: cultural context, anything they struggled to render, anything the patient said that did not reach you cleanly.
• Make sure the written instructions the patient leaves with are usable — a perfectly interpreted plan delivered on an English-only handout is not a plan.

Can a bilingual clinician or bilingual staff member just interpret?

Not on self-assessment alone. Conversational fluency and clinical interpreting competence are different skills, and the evidence suggests people cannot judge their own. In an Annals of Family Medicine study of a community health centre's dual-role interpreter programme, self-reported ability was a poor predictor of measured proficiency and interpreting skill (Hardin et al., 2025).

The encouraging finding in that study is how fixable this is. Of 137 eligible bilingual staff, 87 completed a 40-hour online medical interpreter course with pre- and post-testing, and 63 of those 87 (72%) reached the qualifying standard — at a cost of roughly $500 per employee, with mean interpretation scores rising from 7.34 to 8.46. The authors recommend formal language testing rather than relying on employees' self-reports, and position the approach as a way for clinics to meet language-access requirements with the staff they already employ. For a small clinic with two bilingual receptionists, that is a far more realistic plan than a contract nobody can fund.

Can AI translation tools close the language gap?

Partly, and not yet for the consultation itself. The current evidence supports machine translation for brief, simple, low-stakes exchanges and for material at a distance from the patient — leaflets, terminology, educational content — but not as a substitute for a qualified interpreter in a diagnostic or consent conversation. The accuracy is also asymmetric, which is the detail clinicians most need to know.

A systematic review of AI translation in clinical settings screened 1,095 reports and analysed nine studies across the US, France, Switzerland and South Africa. Reported accuracy was 83–97.8% translating from English, but only 36–76% translating into English. Usability scored well (76.7–96.7%), but patient satisfaction (84–96.6%) consistently outran clinician satisfaction (53.8–86.7%). The authors conclude that AI handles simple, brief clinical communication adequately, remains inadequate for extensive discussion, is inferior to human interpretation, and is best combined with human services (Genovese et al., Annals of Translational Medicine, 2024).

A scoping review of 33 studies reaches the same conclusion more sharply: machine translation is "still unsuitable for direct use in patient interactions" because of the clinical risk created by insufficient accuracy, while showing more promise further from the bedside (Merx, Phillips and Suominen, Studies in Health Technology and Informatics, 2024). And newer is not automatically better: a 2025 pilot comparing large language models with traditional machine-translation tools on medical consultation summaries in Arabic, Chinese and Vietnamese found the traditional tools generally performed better on complex texts, with LLM output inconsistent — and warned that the automated metrics used to score translations do not capture clinical relevance (Li et al., arXiv preprint, 2025).

Read those three findings together and a usable rule falls out. Machine translation is acceptable where the content is short, concrete and verifiable — "where is the pain?", "take one tablet twice a day" — and where a misunderstanding would be caught immediately. It is not acceptable where meaning is layered: consent, risk, psychiatric assessment, safeguarding, breaking bad news, or anything you would want defensibly recorded. The asymmetry matters because the patient's history is the leg that is weakest: translating into English is precisely where these tools fail most.

How should language and interpretation be documented?

Treat it as a clinical fact about the encounter, not an administrative footnote. A reader of your note six months later — a colleague, a lawyer, a coroner — needs to know how the history was obtained, because it bears on how much weight that history carries. Record, at minimum:

• The patient's preferred language, and whether an interpreter was needed or offered.
• The interpreter's name and the modality (in person, telephone, video, trained staff member).
• If the patient declined an interpreter, or insisted on a family member, that the offer was made and declined — the AAFP specifically advises documenting this.
• That understanding was confirmed, and how (teach-back, show-me, repeat-back of the plan).
• Which language the written instructions were given in.

In a SOAP-structured note, most of this belongs at the top of the Subjective — it frames everything that follows. Our guide to how to write a SOAP note covers the structure, and the free printable SOAP note template gives you a place to put it. If the extra documentation feels like one more thing in a day that is already full, that is a real problem and not a small one: see the hidden cost of clinical documentation.

An illustrative interpreted consultation

Illustrative example — not a real patient. A 58-year-old man attends with three weeks of epigastric pain. Reception has flagged his preferred language and a telephone interpreter is connected before you enter the room. His adult daughter is present and offers to interpret.

• You decline the offer kindly and keep the professional interpreter, explaining in one sentence that it is routine and that she is welcome to stay as family. This also reopens the questions a daughter's presence would have closed — alcohol, melaena, weight loss, mood.
• You brief the interpreter in 30 seconds: epigastric pain, you are screening for red flags, you will use short sentences.
• You ask one thing at a time: "When did the pain start?" — pause — "Does it wake you at night?" — pause. Idiom is dropped: not "has it been playing up", but "is it worse now than three weeks ago?"
• You use teach-back for the plan: "Tell me how you will take this medicine tomorrow." He repeats it as once daily; it is twice daily. You correct it. Had you accepted his nod, he would have been under-treated for a fortnight.
• You document it: preferred language, telephone interpreter and name, daughter present, interpretation declined for the family member, teach-back completed, instructions supplied in his language.

Nothing in that encounter required new technology. It required the interpreter to be arranged before the room, not after — which is a scheduling decision, not a clinical one.

What does good language access look like at clinic level?

Individual technique only goes so far; most failures are systems failures. The clinic-level moves that make the biggest difference are unglamorous:

• Capture preferred language at first contact and store it as structured data, so it can be used for booking, reminders and audit rather than living in a free-text note.
• Make the interpreter the default, not the escalation — if arranging one takes three phone calls, clinicians will stop arranging one.
• Train and formally test your bilingual staff rather than relying on them informally; the dual-role model above is cheap relative to its risk reduction.
• Budget honestly. At roughly $2–3 per minute for telephone interpreting without a contract, and no reimbursement in most US settings, language access is a real line item; pretending otherwise is how ad hoc interpreting becomes policy by default.
• Build bilingual patient-facing material for your three or four commonest languages — instructions, consent summaries, safety-netting advice — reviewed by a human, since this is exactly the lower-risk territory where machine translation performs best.
• Audit it. Count how often a patient with a recorded language need was seen without an interpreter. The number is always higher than clinicians expect, and it is the only metric that changes behaviour.

Frequently asked questions

Can I use a family member to interpret if the patient asks me to?

Sometimes, but understand the trade-off and record it. Family interpreters cannot guarantee confidentiality, may have their own agenda, and typically narrow what you can ask about sexual, psychiatric or safeguarding matters. The AAFP advises documenting in the chart that the patient insisted on a family member, and a professional interpreter should still be offered.

Is it ever acceptable to use a child as an interpreter?

Only in a genuine emergency. Children lack the vocabulary and the maturity for clinical material, the role is harmful to them, and several US states forbid the practice outright. If a child has interpreted during an emergency, arrange a qualified interpreter and revisit the history as soon as the situation allows.

How accurate is AI translation for medical conversations?

Good enough for short, simple exchanges and clearly not good enough for consultations. A 2024 systematic review found accuracy of 83–97.8% translating from English but only 36–76% translating into English, and concluded AI remains inferior to human interpretation for extensive discussion. A separate scoping review calls machine translation unsuitable for direct use in patient interactions.

Do interpreted consultations take longer?

Yes, and the time should be scheduled rather than absorbed. The AAFP advises allowing extra time for interviews conducted through an interpreter and roughly a 10-minute break for every hour of interpretation. Booking the interpreter before the appointment, rather than hunting for one once the patient is in the room, recovers most of the overrun.

Related reading

• What is an AI medical scribe? A guide for doctors — what ambient documentation does and does not do.
• How to write a SOAP note — where language and interpreter details belong in the record.
• The hidden cost of clinical documentation — the admin load that interpreted visits add to.
• Physician burnout: causes, signs and solutions that work — why time pressure drives the shortcuts this article argues against.

References

• Divi C, Koss RG, Schmaltz SP, Loeb JM. Language proficiency and adverse events in US hospitals: a pilot study. International Journal for Quality in Health Care. 2007;19(2):60–67. academic.oup.com/intqhc/article/19/2/60
• Juckett G, Unger K. Appropriate use of medical interpreters. American Family Physician. 2014;90(7):476–480. aafp.org/pubs/afp/issues/2014/1001/p476
• World Health Organization. Refugee and migrant health (fact sheet). March 2026. who.int/news-room/fact-sheets/detail/refugee-and-migrant-health
• Genovese A, Borna S, Gomez-Cabello CA, et al. Artificial intelligence in clinical settings: a systematic review of its role in language translation and interpretation. Annals of Translational Medicine. 2024;12(6):117. atm.amegroups.org/article/view/132363/html
• Merx R, Phillips C, Suominen H. Machine translation technology in health: a scoping review. Studies in Health Technology and Informatics. 2024. openresearch-repository.anu.edu.au
• Li A, Zhou W, Hoda R, Bain C, Poon P. Comparing large language models and traditional machine translation tools for translating medical consultation summaries: a pilot study. arXiv preprint. 2025. arxiv.org/abs/2504.16601
• Hardin DM, Osornio JA, Peña JR III, Hur E, Hardin KJ. Medical interpreting in primary care: design and validation of a replicable training program. Annals of Family Medicine. 2025;23(6):507–514. annfammed.org/content/23/6/507