Machine Translation in Healthcare: When Not to Use It

Machine translation tools such as Google Translate, DeepL and large language models like ChatGPT are now one click away for every clinician, pharmacist and medical writer. This article is for healthcare providers, life-science companies and anyone deciding when these tools are acceptable and when they put patients or compliance at risk.

Machine translation (MT) is translation produced automatically by software, without a human translator. It can be used raw, or as a draft that a professional then corrects, a process called machine translation post-editing (MTPE), explained in our article on what MTPE is.

What does NHS guidance say about machine translation?

NHS England advises against using automated translation tools for patient documents in primary care. Its guidance for commissioners on interpreting and translation services in primary care states:

"Automated on-line translating systems or services such as "Google-translate" should be avoided as there is no assurance of the quality of the translations." (NHS England, Guidance for commissioners: Interpreting and Translation Services in Primary Care)

The same guidance says documents translated for patients "must be translated by competent and appropriately trained translators and not by practice staff". Spoken communication follows the same logic: patients should always be offered a registered interpreter, as covered in our article on working with interpreters in healthcare settings.

How accurate is machine translation for patient instructions?

Machine translation of patient instructions is often accurate in widely spoken languages but unreliable in others, and the errors that remain can be harmful. Three peer-reviewed studies of emergency department instructions illustrate the pattern.

The Taira study concluded that Google Translate for discharge instructions "is inconsistent between languages and should not be relied on for patient instructions". The Kong study, which tested newer tools, supported machine translation only "for low-stakes written communication from English to Spanish", while "strengthening the case for caution and for professional oversight in non-low-risk communication".

Two lessons stand out. First, good results in one language say little about another, so a tool cannot be approved for "all languages" on the strength of one test. Second, a 2% harm rate sounds small until it is applied to every medication instruction a hospital issues.

When should machine translation not be used in healthcare?

Machine translation should not be used without qualified human translation or full expert review for any text where an error could harm a patient or breach a regulatory requirement. That includes:

• Dosing and medication instructions, including discharge letters and prescription labels.
• Consent forms and participant information for treatment or clinical trials.
• Patient information leaflets, labels and product information for medicines, which regulators approve or review; see patient information leaflet translation.
• Medical device labels and instructions for use, which must be readily understood by the intended user under the EU MDR; see EU MDR language requirements.
• Patient-reported outcome questionnaires, which need linguistic validation.
• Adverse event reports, where the reporter's exact words matter for safety assessment.
• Medical records sent to another clinician for treatment decisions.

Where can machine translation be acceptable?

Machine translation can be acceptable for low-risk uses, provided people understand its limits and a qualified human checks anything that leaves the organisation. Examples include getting the gist of a foreign-language journal abstract during a literature search, or triaging which incoming documents need full translation.

When MT is used as a draft for professional post-editing, the final responsibility still rests with a qualified medical translator who checks every sentence against the source. The Khoong study also recommended including English instructions alongside machine-translated text and "automated warnings regarding the use of machine translation" where clinicians use it at all.

Is it safe to paste patient data into online translation tools?

Pasting patient data into public online translation tools creates a data protection risk as well as an accuracy risk. The Information Commissioner's Office lists "data concerning health" as special category data, which it describes as "personal data that needs more protection because it is sensitive".

Before any tool processes patient information, the organisation needs to know where the data goes, whether it is stored or used to train models, and on what lawful basis it is processed. Our guide to data security when outsourcing translation covers the questions to ask any provider, human or machine.

How to set a machine translation policy for a healthcare organisation

1. Classify content by risk: patient-facing clinical, regulated, internal, and informational.
2. Ban raw MT for high-risk categories and require professional translation with independent review.
3. Approve specific tools for low-risk uses, with contracts that prevent storage or reuse of data.
4. Require post-editing by qualified medical translators wherever MT drafts are used for external documents.
5. Train staff on what not to paste into public tools and when to book an interpreter instead.
6. Review incidents where translation contributed to a complaint or near miss, and update the policy.

Common mistakes with machine translation in healthcare

• Judging quality by fluency. MT output often reads smoothly even when the meaning is wrong.
• Testing in one language and assuming the rest. Results for Spanish say little about Armenian or Farsi.
• Feeding in poor source text. The Khoong study linked spelling and grammar anomalies in the English source to inaccurate translations.
• Calling light editing "post-editing". A bilingual colleague skimming the output is not a qualified review.

For a broader comparison of tools, see are DeepL and ChatGPT good enough for business translation. For the full landscape of regulated medical translation, read our medical and pharmaceutical translation guide and our article on why medical translation needs subject-matter experts.

How TranslationWorks can help with safe medical translation

TranslationWorks is a UK-based translation agency that provides professional human translation of patient information, clinical documents and regulated medical content through native-speaking linguists in more than 150 languages. Our medical and pharmaceutical translation service includes independent review, so texts that patients and regulators rely on are checked by people, not left to software.

TranslationWorks holds ISO 17100, ISO 9001 and ISO 27001 registration and Cyber Essentials certification, which supports secure handling of health data. Tell us what you need translated and who will read it, and get a free quote within 24 hours.