AI Medical Translation Tools: Benefits, Risks and Patient Safety

Last updated: October 21, 2026

How to read these figures: the first three come from one study of 50 emergency department instruction sets (316 sentences) translated in late 2023 (Kong et al., BMJ Quality & Safety). «Accuracy» and «potential harm» were judged by physicians reviewing back-translations; they are not counts of patients who were actually harmed. The last figure comes from a separate 2021 study (Taira et al.) with a different method.

A nurse holds up a phone so a patient can read the screen. A parent types a doctor’s instructions into a chatbot at home. A clinic installs a tablet that listens in one language and speaks in another. AI translation is now part of everyday healthcare, often without any formal decision to use it.

For people who do not share a language with their care team, these tools can be the difference between understanding a treatment plan and leaving confused. But in medicine one wrong word, such as «continue» instead of «stop», can change the outcome. This guide looks at what the research shows, where the risks are concentrated and how to use these tools more safely, whether you live in the United States, Spain or elsewhere in Europe.

How we researched this article

We searched PubMed, journal websites and official government sources in October 2026 for studies published between 2012 and 2026 that tested machine translation or AI interpretation on clinical text or conversations. We included studies that compared tools against professional translators or interpreters, or that had clinicians or native speakers judge accuracy and potential harm. We prioritised peer-reviewed research. Studies run by companies selling translation products were excluded unless they had been peer-reviewed, and conference abstracts were not used for figures. This is a narrative review, not a systematic review following a formal protocol, and Helathora did not test any tool independently. Results for one tool, version or language may not apply to others.

Three kinds of AI translation in healthcare

What the studies show about accuracy

A 2019 study in JAMA Internal Medicine ran 100 real emergency department discharge instructions through Google Translate. Of 647 sentences, 92% were translated accurately into Spanish and 81% into Chinese. About 2% of Spanish and 8% of Chinese sentences contained an error that reviewers judged could cause clinically significant harm.

A 2021 study in the Journal of General Internal Medicine tested 20 common discharge phrases in seven languages, judged by native speakers. Overall, Google Translate kept the meaning in 82.5% of the 400 translations, but the spread between languages was wide.

The authors concluded that performance was too inconsistent between languages to rely on for patient instructions. In Armenian, a sentence about taking ibuprofen for pain turned into a reference to an anti-tank missile, according to Becker’s Hospital Review.

Are ChatGPT and other large language models better?

Often, yes, but sentence-level accuracy can give a misleading sense of safety. The UCSF team behind the 2019 study tested ChatGPT-4 (a secure version hosted by the university) and Google Translate in October and November 2023, using 50 real emergency instruction sets translated into Spanish, Chinese and Russian. The results were published in BMJ Quality & Safety.

Source: Kong M et al., BMJ Quality & Safety. 50 instruction sets, 316 sentences. Confidence intervals for the instruction-set figures are wide because of the small sample. On a phone, scroll sideways to see the full table.

Three points matter here. First, most errors were minor, such as «relevant symptoms» instead of «concerning symptoms», but a small share could have led a patient to do the wrong thing. Second, the results apply to ChatGPT-4 and Google Translate as they were in late 2023; they do not show that every current AI model performs the same way. Third, the authors recommend machine translation without professional review only for low-stakes English-to-Spanish messages, such as confirming an appointment, and advise against it for high-stakes communication in any language.

A 2024 study in Pediatrics found a similar pattern for children’s discharge instructions: Google Translate and ChatGPT-4 performed comparably to professional translators for Spanish and Brazilian Portuguese, but fell short for Haitian Creole.

Spoken translation is the weakest link

Real-time interpretation adds two more steps where errors can creep in: recognising speech and turning text back into a voice. An evaluation in BMJ Innovations found Google Translate preserved the meaning of single spoken English sentences into Spanish about 89–91% of the time, but only 52.2% when two sentences were spoken together and 26.6% for three.

Newer, purpose-built tools look more promising, but the evidence is narrow. A 2026 simulation study in npj Health Systems compared one in-house hospital app, LingualAI, with certified interpreters in scripted English-Spanish ear, nose and throat consultations. The app was non-inferior for medical terminology and meaning, but not for clarity or voice quality, and overall ratings favoured the interpreters. The study tested a single app, in one language pair and one specialty; it does not show that AI interpretation apps in general match certified interpreters.

A practical example: write it simply before you translate

Much of the risk can be reduced before any tool is used, by rewriting the source text. The example below is illustrative and was not tested by Helathora.

Short sentences, one instruction per sentence, the actual medicine name and specific warning signs give any translation tool, and any patient, a better chance.

The benefits are real

It would be wrong to see these tools only as a danger. Professional interpreters are not always available; a 2022 paper in JMIR Mental Health cites estimates that fewer than 20% of U.S. clinical encounters with patients who have limited English use interpreting services. The real alternative to AI is often not a professional, but a relative, a bilingual colleague or nothing at all.

That alternative carries its own risks. A 2012 study in Annals of Emergency Medicine analysed recordings of 57 pediatric emergency visits with Spanish-speaking families and identified 1,884 interpretation errors. Among the errors made in visits with professional interpreters, 12% had potential clinical consequences, compared with 22% of errors in visits with untrained («ad hoc») interpreters such as family members and 20% in visits with no interpreter. That study did not test AI, but it shows why machine translation should be compared with the help actually available, not only with a certified professional.

Your options compared

On a phone, scroll sideways to see the full table.

Protecting your privacy

Free consumer apps and chatbots are not always designed for confidential health information, and what you type may be stored or used under the provider’s own terms. Before using one:

• Remove names, dates of birth, addresses and medical record or health card numbers.
• Do not upload full medical reports or test results that identify you.
• Translate general wording («my blood test shows low iron») rather than copying the document.
• Clinicians should only use tools their organisation has approved for patient data.
• Check the tool’s privacy settings and whether it keeps your history.

What the rules say in the United States

The U.S. rules do not ban machine translation, but they limit how it can be used. Under the 2024 update to the Section 1557 nondiscrimination rule, 45 CFR 92.201 requires health programs covered by the rule to have machine translations reviewed by a qualified human translator when the text is critical to a person’s rights, benefits or meaningful access, when accuracy is essential, or when it contains complex, non-literal or technical language. The provision still appeared in the federal regulations as of mid-2026. According to the authors of the BMJ Quality & Safety study, when none of those conditions apply and machine translation is used, patients should be warned that the translation may contain errors.

What patients in Spain and Europe should know

There is no single European rule equivalent to Section 1557, and practice varies by country and even by hospital. In Spain, language support in public healthcare differs between autonomous communities, and sometimes between hospitals in the same province, according to research from the University of Valencia. Some regions run telephone interpreting services: the Community of Madrid, for example, offers free telephone interpreting in 65 languages in its primary care centres. Where no service exists, family members or colleagues often step in.

• Machine translation of text and interpretation during a consultation are different services; ask which one your centre offers.
• Ask at reception or patient services (in Spain, «atención al paciente») whether telephone or in-person interpreting is available.
• Health data is a special category of personal data under the EU’s General Data Protection Regulation, so be careful about entering it into consumer tools.
• Your exact rights depend on national and regional rules and on the service you use; ask your provider.

The European Commission publishes information on its approach to artificial intelligence in healthcare, including the regulatory framework for AI in the EU.

A safer-use checklist

For clinicians and health services:

• Use a qualified interpreter for consent, new diagnoses, medication changes and bad news.
• Write the source text in short, plain sentences without jargon or idioms.
• Label machine-translated documents and keep the original-language version alongside.
• Have critical documents reviewed by a qualified human translator.
• Check understanding with teach-back, not «Do you understand?».

For patients and families:

• Ask whether a professional interpreter is available, including by phone or video.
• Repeat important instructions back in your own words before leaving.
• If a translated sentence seems strange, ask; do not guess.
• Double-check medication names, doses and timing with a pharmacist.

Limitations of the evidence

Most studies so far test emergency discharge instructions in a handful of languages, often with small samples, and some use phrases chosen for the study rather than real conversations. Translation tools change constantly, so results from even two years ago may not reflect today’s versions, for better or worse. «Accuracy» and «potential harm» are judgements by reviewers, often based on back-translation, not observed patient outcomes; we found little research tracking whether machine translation changes what actually happens to patients. Almost all the evidence comes from the United States and from translation out of English, so findings may not apply directly to translation from Spanish or other European languages.

Frequently asked questions

Related reading on Helathora

Official sources

• Kong M, Fernandez A, Bains J et al. Evaluation of the accuracy and safety of machine translation of patient-specific discharge instructions: a comparative analysis. BMJ Quality & Safety (published online 2025).
• Khoong EC et al. Assessing the Use of Google Translate for Spanish and Chinese Translations of Emergency Department Discharge Instructions. JAMA Internal Medicine, 2019.
• Taira BR, Kreger V, Orue A et al. A Pragmatic Assessment of Google Translate for Emergency Department Instructions. Journal of General Internal Medicine, 2021.
• Performance of ChatGPT and Google Translate for Pediatric Discharge Instruction Translation. Pediatrics, 2024.
• Birkenbeuel J et al. Google translate in healthcare: preliminary evaluation of transcription, translation and speech synthesis accuracy. BMJ Innovations.
• LingualAI versus certified medical interpreters in English-Spanish otorhinolaryngology encounters (simulation study). npj Health Systems, 2026.
• Flores G et al. Errors of Medical Interpretation and Their Potential Clinical Consequences. Annals of Emergency Medicine, 2012.
• U.S. Code of Federal Regulations. 45 CFR § 92.201: Meaningful access for individuals with limited English proficiency.
• European Commission. Artificial intelligence in healthcare.
• Universitat de València. Healthcare interpreting in Spain: exploratory qualitative study.

Medical disclaimer: This article is for general information only and does not constitute medical or legal advice. Translation tools can make errors that affect your care. Always confirm diagnoses, medication instructions and treatment decisions with a qualified healthcare professional, with the help of a professional interpreter when needed.