AI Medical Translation: Why Vietnamese Still Needs People
💡 TL;DR: 2026 made AI medical translation sound solved, with voice tools claiming near-perfect scores. But a peer-reviewed JAMA Network Open study tells a harder truth: when AI translated hospital discharge instructions, results were comparable to humans for Spanish yet markedly worse for Vietnamese, Chinese and Somali. For Vietnamese, AI produced clinically impactful errors in 41% of sections, against 14% for professional human translators. For anything a patient acts on, human medical translation is still the safe default.
- A 2026 JAMA Network Open study tested AI vs human translation of 148 pediatric discharge sections in four languages.
- For Vietnamese, AI produced clinically impactful errors in 41% of sections versus 14% for human translators.
- Chinese was 52% vs 20% and Somali 92% vs 13%, while Spanish was comparable at about 7% on both sides.
- The dangerous errors are small and fluent: flipped negations, drifted doses or frequency, and swapped symptoms.
- Route consent forms, discharge and dosing instructions to certified human Vietnamese translation; use AI only as a reviewed drafting aid.
AI translation just had its loudest year yet
If you only read the headlines, you would think the language barrier fell in 2026. Real-time voice tools went mainstream, and DeepL reported that 96% of professional linguists preferred its DeepL Voice product in blind evaluations, with quality scores near 96 out of 100 and a 76% drop in critical errors versus rivals. Voice-to-voice translation now spans 40 plus languages, and meeting integrations rolled out through the first half of the year. I covered the live side of this shift in my piece on Gemini 3.5 live translation, and the progress is real.
So it is fair to ask the uncomfortable question. If AI is this good at conference calls and travel chat, is AI medical translation good enough to hand a patient their discharge papers without a human in the loop? The honest answer arrived not from a press release but from a hospital study.
What happened when AI translated hospital discharge papers
In JAMA Network Open, a research team led by Melissa Martos and K. Casey Lion ran a clean comparison. They took 148 sections drawn from 34 real pediatric discharge instructions and translated each into four languages spoken by US patient families: Spanish, Simplified Chinese, Vietnamese and Somali. One version came from a widely used neural machine translation engine, Azure AI Translator. The other came from professional human translators.
Then two professional translators per language scored every section blind on a five point scale for fluency, adequacy, faithful meaning and error severity, and they flagged whether any error was clinically impactful, meaning it could change what a parent actually does at home. This is the right test, because discharge instructions are not marketing copy. They tell a family how much medicine to give, how often, and which warning signs mean go back to the emergency room.
Vietnamese is exactly where the cracks appear
The aggregate numbers already favored people: across all languages, human translations beat AI on fluency, adequacy, meaning and error severity, every gap statistically significant. But the language breakdown is what matters for Vietnamese medical translation. For Spanish, AI was essentially comparable, with clinically impactful errors in about 7% of sections for both AI and humans. Vietnamese was a different story.
- Clinically impactful errors: 41% of AI translated Vietnamese sections versus 14% for human translators.
- Faithful meaning, fluency and error severity: all meaningfully lower for AI, each difference statistically significant.
- Chinese fared similarly poorly, with clinically impactful errors in 52% of AI sections versus 20% for humans, and Somali was worst at 92% versus 13%.
| Language | AI: clinically impactful errors | Human: clinically impactful errors |
|---|---|---|
| Spanish | ~7% | ~7% |
| Vietnamese | 41% | 14% |
| Chinese | 52% | 20% |
| Somali | 92% | 13% |
Read that again. In nearly half of the Vietnamese sections, the AI version contained a mistake serious enough to potentially harm a child. That is not a typo problem. That is a patient safety problem.
Why Spanish is fine and Vietnamese is not
The split is not random. Machine translation quality tracks how much high quality bilingual data a model has seen. Spanish and English share an enormous, well aligned corpus: parallel laws, medical leaflets, news and books. Vietnamese, Chinese and Somali are comparatively low resource for clinical text, so the model has thinner ground to stand on exactly where precision matters most.
Vietnamese adds its own traps. Tone and diacritics carry meaning, so a dropped mark can turn one word into another. The language leans on classifiers and context rather than the explicit plurals and tenses English uses, which makes dosage and frequency easy to blur. And clinical Vietnamese vocabulary is genuinely specialized: the everyday word and the correct medical term are often not the same, and a general engine reaches for the everyday one.
What a clinically impactful error looks like
After years of English to Vietnamese translation in medical and legal files, the failure patterns are familiar. The dangerous errors are rarely exotic. They are small, fluent and confident:
- Negation flips: do not resume the medication becomes resume the medication. The sentence still reads smoothly, which is what makes it dangerous.
- Dose and frequency drift: twice daily quietly turns into every two days, or 5 mL becomes 5 mg. Units and intervals are where machine output is least reliable.
- Anatomy and symptom swaps: a near synonym replaces the precise term, so a warning sign loses its edge.
A parent does not get to see the English original. They trust the Vietnamese in their hand. When that text is fluent but wrong, fluency becomes a liability, because it removes the very confusion that might otherwise prompt a question.
Where AI does earn its place
None of this means AI has no role in healthcare language work. Used with judgment, it is a strong assistant. It can draft a first pass that a qualified human medical translator then reviews and corrects, which is faster than starting from a blank page. It can power instant, low stakes communication, such as helping front desk staff understand a simple request. It can triage a backlog so humans spend their hours on the documents that carry real risk.
The principle is matching the tool to the stakes. For internal notes, ephemeral chat and gisting, machine output is often fine. For anything a patient signs, swallows or acts on, the workflow needs a credentialed human who owns the result. The JAMA authors land in the same place, urging thoughtful, inclusive implementation rather than blanket deployment.
A safer playbook for health content in Vietnamese
If your organization serves Vietnamese speaking patients, a few habits remove most of the risk. First, route all high stakes documents, consent forms, discharge and medication instructions, dosing and legal disclosures, to certified Vietnamese translation by a qualified human, not raw machine output. Second, ask for back translation on critical passages so an independent translator can confirm the meaning survived. Third, build and reuse a medical glossary so terms stay consistent across every document, the same discipline I described in my guide to Vietnamese translation rates. Treat AI as a drafting accelerator inside that human owned process, never as the final voice for clinical content. The cost of professional medical translation is trivial next to the cost of a single avoidable readmission.
FAQ
Is AI good enough for medical translation into Vietnamese in 2026?
Not on its own for documents patients act on. A 2026 JAMA Network Open study found AI produced clinically impactful errors in 41% of Vietnamese discharge instruction sections, versus 14% for professional human translators. For consent forms, discharge and dosing instructions, use certified Vietnamese translation by a qualified human, and keep AI as a drafting aid that a person reviews.
Why is AI medical translation worse for Vietnamese than for Spanish?
Because machine translation quality depends on how much high quality bilingual medical data a model has seen, and English-Spanish has far more of it than English-Vietnamese. Vietnamese also relies on tones, diacritics and context rather than explicit plurals and tenses, so dosage, frequency and negation are easy for an engine to blur. The same study found AI comparable for Spanish but markedly worse for Vietnamese, Chinese and Somali.
What is a clinically impactful translation error?
It is a mistake serious enough to change what a patient or caregiver does, such as a flipped negation, a wrong dose or frequency, or a swapped symptom. These are dangerous precisely because the surrounding text still reads fluently, so nothing signals that something is wrong. Professional medical translators are trained to catch them.
Should hospitals stop using AI translation entirely?
No. AI is useful for low stakes, internal or ephemeral communication, and as a first draft that a qualified human reviews. The safe rule is to match the tool to the stakes: machine assistance for gisting, and certified human translation for anything a patient signs, swallows or acts on.
Source: JAMA Network Open
About the author
Dao Huy (Lucas) is a professional translator working across English, Vietnamese, Chinese and French, with 7+ years in medical, legal, financial and academic translation. The discharge instruction errors in this study are the exact failure modes that careful human medical translation is built to prevent: faithful meaning, correct doses and warnings that do not quietly flip.
If you need English to Vietnamese translation, certified Vietnamese translation of medical or legal documents, or multilingual localization that reads naturally and safely, I am happy to help. Get a quote at daohuy.com and tell me what you need translated.
Written by Dao Huy (Lucas), Vietnamese translator & localization specialist (EN · ZH · FR → Vietnamese). See translation services →