Machine translation and AI in the language chain
Machine translation and language models can produce multilingual drafts quickly, but their usefulness depends on the text, language pair, prompting, confidentiality arrangements and expert review of the result.
Capabilities and their limits
In 2012, Google Translate could provide the gist of simple French or German text but mishandled tenses and sometimes invented meaning: a German account of visiting a brother with a caravan in Croatia became an English sentence about running and a bear owning the caravan. By December 2020, fully automatic translation was still considered unreliable for professional use because understanding the original's context required a person. Human revision remained necessary, although integration with translation memory or globalisation management systems could reduce the work.
By May 2024, neural machine translation could match human translation in some domains, with quality depending on the engine, training data and its curation, use, domain and language pair; English–Spanish had more high-quality resources than languages of lesser diffusion. Language models differed because they could be prompted for context, register and audience, such as an informed consent form for children aged eight to eleven. Their quality depended on training data and prompting skill; without suitable prompting, they were no better than machine translation and certainly no better than professional human translation.
AI may assist with initial translation or adaptation drafts, terminology and consistency checks, source–target comparisons and repeated lower-risk content. These uses depend on clear source writing, stable terminology, sufficient context, a defined audience and purpose, an agreed route through translation or adaptation, and human review. Ambiguous or inconsistent inputs can spread problems faster, and fluent output does not establish accuracy, usefulness or appropriateness. Combining human work and AI may be appropriate where the tools are used ethically.
Prompts for translation and revision
In May 2024 demonstrations, conventional machine translation returned the same output for the same input and retained the source's structure, including a parenthetical exception for Japan in an incidence sentence. A simple instruction to translate produced similar output from a language model, so a more elaborate prompt needed to improve on a machine translation available immediately without prompting.
Asking the model to rewrite in the target language as an educated native speaker and expert medical writer, with natural wording that does not sound translated, improved the demonstrated output. In the incidence example, a rewrite prompt removed the parentheses and made the sentence flow. Specifying the original's tone, register and intent changed the wording again, including replacing “physician” with “doctor”.
A literal prompt can help back-translation reveal defects by preserving the source word for word. In a demonstration, missing Spanish text produced an English assertion that breast lumps did not occur in old people; an ordinary translate prompt instead supplied “all individuals”, repairing the wording and concealing the missing source information. Literal output can therefore expose a problem during checking while carrying it into a translation intended for use.
“Improve the syntax and readability” can help review a translation whose meaning is correct but whose wording reads poorly, although it may change very little. “Rewrite for clarity” made a demonstrated incidence figure more visible in its sentence. The translator still has to decide whether the revision is better. Models can also help explain a difficult passage or summarise a long legal paragraph in bullet points before translation, provided it is not confidential.
Human involvement therefore includes deciding whether the prompt fits the task and recognising AI-generated source text, alongside the earlier decisions to post-edit a draft or discard it and translate again. Rapid output in several languages is no assurance of quality. Someone who cannot read the output language cannot judge it adequately and should involve an expert native linguist.
Evidence from abstracts and plain language summaries
An August 2022 pilot compared Google Translate output for five Ipsen-sponsored articles' abstracts and plain language summaries of similar length, across therapy areas and study types. All had medical writing assistance and uniformly high-quality English. Translation into French, German, Mandarin and Slovenian produced twenty translations of each document type, evaluated by one bilingual reviewer with a scientific background per language using error counts and five-point scales for accuracy and likelihood of misinterpretation. The assessed categories covered untranslated words, misspellings, additions and deletions; phrases and tone; grammar, syntax, punctuation, symbols, units, numbers and dates; and clarity.
Both types were considered accurate and readable, with plain language summaries scoring better across most measures. Word insertions, omissions and misspellings were rare and mostly had little or no effect on meaning. The results for summaries and abstracts, respectively, were:
| Measure | Plain language summaries | Abstracts |
|---|---|---|
| Inappropriate inserted words | 0.2% | 0.1% |
| Omitted words | 0.2% | 0.2% |
| Misspellings | 0.05% | 0.02% |
| Untranslated words | 0.6% | 1.0% |
| Mistranslated words | 0.6% | 1.0% |
| Incorrect phrases per 100 words | 0.35 | 0.41 |
| Grammar/syntax errors potentially causing misinterpretation | 17.2% | 40% |
| Mean accuracy score, out of 5 | 4.60 | 4.30 |
| Mean score for unlikelihood of misinterpretation, out of 5 | 4.55 | 4.25 |
| Rated very accurate | 60% | 45% |
| Rated definitely unlikely to be misinterpreted | 60% | 45% |
Grammatical and syntax error rates were similar, and no translation was rated inaccurate or very inaccurate, or probably or definitely likely to be misinterpreted. No language stood out, but the sample could not detect inter-language differences. Differences were generally small and no statistical tests established significance.
Reviewers found active-voice text translated more naturally and readably than passive-voice text. Acronyms were inconsistent: Google Translate often recognised them when defined but lost them on later use or when pluralised with an “s”. Abstracts' complex sentences, abbreviations and scientific terminology may explain the differences, and clear-language recommendations for translation-friendly summaries could also be applied to full articles.
The study was limited to five similarly written publications from one sponsor, four languages and one reviewer per language. French, German and Mandarin probably had extensive training coverage, while languages with less internet content might perform worse. Only Google Translate was tested; research cited in the 2023 discussion suggested ChatGPT did not consistently outperform Google Translate or Microsoft Translator and performed worse with less widely spoken languages. Short text summaries were easier to handle than long documents or material embedded in tables and figures, and a blinded back-translation could supplement the survey assessment. In 2023, browser translations were not considered suitable without professional review, particularly for pharmaceutical reporting bound by accuracy requirements, and replacing specialist translation of scientific articles was considered a distant possibility, if achievable at all. Publishers and the scientific community were encouraged to make professional translations of technical publications available for training translation software.
Suitability and post-editing
In late 2024, one medical translation agency offered machine translation with full post-editing, but neither light post-editing, which had not produced good results, nor machine translation alone, which it considered inadequate for medical work. Its experience was that full post-editing could deliver quality equal to human translation while saving time.
The agency assessed suitability per text with a metric yielding an approximate percentage and assigned post-editing only when the text passed. Medical marketing texts with different styles or structures, missing context, unsupported language pairs and insufficient trained post-editors could make machine translation unsuitable. A full translation memory could also make it uneconomic; the account described a 100% match as costing more than machine translation output. The project manager checked whether post-editing would take longer than proofreading or translating outright. Additional proofreading could still be needed, so a machine translation project was not necessarily shorter. Engines could be trained and add-ons developed, while the shortage of trained post-editors was changing quickly as workflows shifted towards post-editing.
A second-hand report of EMWA's translation group's November 2024 discussion was more cautious: AI could provide inspiration, but translation use was not recommended because editing it often took longer than human translation from the start. The group nevertheless treated AI as an important tool rather than rejecting it altogether. These experiences differ on time savings, while the agency's assessment explicitly tests for cases where post-editing costs more effort.
Expert-facing, syntactically complex and terminology-dense texts tend to need more post-editing, and consent forms add their own genre difficulties. As of March 2024, generative AI nevertheless showed promise for simplifying complex patient-facing information, including consent forms. Technology training can help translators recognise machine translation error types and improve post-editing and responsible use.
Productivity and terminology
A March 2024 account put typical productivity gains nearer 20% than the 50% advertised by some software makers, with greater potential for experienced medical translators. One experienced translator completed a 6,000-word consent form in about four hours using translation memories, termbases and a neural engine, compared with an estimated three days before those tools. An engine trained on high-quality human-translated and revised consent forms produced output needing only small syntax and terminology changes during careful bilingual revision.
A language model can extract key terms into a two-column table and propose translations, which should be carefully revised before import into a termbase. By March 2024, machine translation was making large volumes of medical information available quickly in multiple languages, and doctors were increasingly using AI translation with patients speaking other languages. In 2022, rapid multilingual document production using translation memories and machine or deep learning was described as depending on standardised concepts and terms, with CDISC and MedDRA enabling coded exchange of research terms and adverse events.
Review and confidentiality
Complex medical concepts, changing terminology, neologisms, abbreviations, incomplete sentences and shorthand can defeat machine translation. Depending on the errors, a text may need heavy post-editing or complete retranslation. French “intoxication”, meaning poisoning, can be wrongly rendered as English intoxication, meaning inebriation. Improving BLEU scores measure increasing similarity to high-quality human translations, but convincingly medical wording does not assure accuracy. Errors can differ unpredictably by platform and language, and searching online for generated phrases does not validate them because matching wording may itself be wrong in context. Output should be checked against credible sources, with less experienced translators especially vulnerable to false confidence. Unfamiliar terminology makes errors easy to overlook, and surviving errors can have regulatory consequences; AI's training may also contain less medical than non-specialist material.
Translators should master grammar, syntax, style, language transfer and their specialty before incorporating technology as an aid, with a human providing final validation. Review requires the cultural background to judge tone and meaning in both languages, and even human translations can contain errors. Native-speaker review and verification remain recommended because linguistic and cultural nuances are difficult to automate. EMA guidance discussed in late 2024 stressed close human supervision of drafting, compiling and translating medicinal product information.
The agency described in late 2024 always sought permission before using a cloud-based engine, and some clients' data-protection policies prevented its use. A March 2024 warning described free public tools as permitting entered text to be shared or used for training, and recommended a secured paid engine such as DeepL Pro for confidential medical text. Medtronic had developed an internal engine, Medtronic Translate, as an alternative to public translation tools that could not be used internally for confidentiality reasons.
A 2022 US Department of Health and Human Services proposal called for qualified human review of machine translation where text was critical to rights, benefits or meaningful access for someone with limited English proficiency, where accuracy was essential, or where the source was complex, non-literal or technical. A December 2025 account identified a qualified-human-review requirement in the May 2024 revision of Section 1557 of the Affordable Care Act.
In a September 2024 clinical research data-literacy project, Latin American Spanish was selected for infographics after assessing the audience and intended reach, with other languages to be considered later. Exported English text was translated with Google Translate, then both versions went to a bilingual partner for back-translation, editing and confirmation before the translated infographic was produced.