Machine translation is one of the quietest AI success stories. It powers menus, product listings, support articles and casual conversation across dozens of languages, usually well enough that no one thinks about it. That everyday fluency is precisely what makes its high-stakes failures so dangerous: because the technology is so obviously useful for low-risk text, organisations extend it to medical, legal and safety-critical content without pausing to ask whether the same tool is fit for a very different job. It usually is not, and the gap does not announce itself.
The accuracy cliff
General-purpose translation models perform impressively on ordinary prose — reported accuracy of roughly 80–85% is common for everyday text. But that headline figure collapses in specialised domains. For dense medical or legal terminology, independent assessments have put accuracy closer to 55–60%. The problem is not just the lower number; it is that the errors cluster exactly where they do the most harm and are hardest for a non-expert to catch.
In healthcare, a mistranslated discharge instruction, dosage or consent form can lead directly to misdiagnosis, incorrect treatment or a patient agreeing to something they did not understand. In law, a reversed clause, a mistranslated name or a term of art rendered with its everyday meaning can invalidate a contract, make evidence inadmissible, or — in documented immigration cases — contribute to a wrongly denied asylum claim. These are not edge cases dredged up to make a point; they are the predictable result of using a general tool for expert work.
Two features of machine translation make this especially treacherous. First, the output is always fluent. A wrong translation reads just as smoothly as a right one, so nothing in the text signals danger. Second, the person best placed to notice the error — a reader who speaks only the target language — has no access to the source and no way to know that something was lost or reversed. Fluency hides the defect from exactly the people it harms.
Why "just add a human" is not the whole answer
The instinctive fix is a human reviewer, and human oversight is indeed essential — but it has to be the right kind of human, doing the right kind of review. Regulators have started to say so explicitly. Guidance under Section 1557 of the US Affordable Care Act now requires that communications critical to a patient's rights, benefits or care be reviewed by a qualified translator, and is pointed about what "qualified" means: being bilingual is not enough. The reviewer needs specialist terminology knowledge and professional accountability. The EU AI Act, for its part, treats migration, asylum and border-control uses of AI — which lean heavily on translation — as high-risk, subject to inspection before deployment.
The lesson is that oversight is a design problem, not a box to tick. A reviewer skimming fluent output for obvious errors will miss exactly the subtle, high-consequence mistakes that matter: the omitted exception, the term of art with a specific legal meaning, the dosage unit silently converted wrong. Effective review requires people with domain expertise, access to the source, and enough time to check meaning rather than merely read for flow.
Building assurance into translation workflows
Treating translation as an assurance problem rather than a convenience turns a vague worry into a set of concrete controls.
Classify content by stakes before choosing a workflow. Not everything needs the same treatment. Marketing blurbs and internal notes can run on machine translation with light review; discharge instructions, contracts, consent forms and safety documentation cannot. The single most valuable step is deciding, deliberately and in advance, which content is high-stakes — and routing it to qualified human review by default.
Require domain-qualified review for high-stakes content. For regulated and safety-critical material, the reviewer must have both the target-language fluency and the subject-matter expertise to catch meaning-level errors, working from the source rather than polishing the output. Bilingualism alone does not qualify anyone to sign off on a medical or legal translation.
Maintain curated glossaries and terminology. Much domain error comes from terms of art rendered with their everyday meaning. A managed glossary of approved translations for critical terms, applied consistently, removes a whole class of recurring mistakes.
Preserve an audit trail. Regulated communication needs a record: source, machine output, human revisions and who approved the final version. Beyond compliance, this is what makes it possible to investigate a harm, learn from it and demonstrate due care. Many consumer-grade tools keep no such trail — and quietly retain the data you feed them, which is its own compliance and privacy problem.
Check data handling before use. Feeding a patient record or a confidential contract into a general translation service can breach privacy rules on its own, regardless of the output quality. High-stakes workflows need tools with appropriate data-protection guarantees, not whatever is most convenient.
The pattern beneath the sector
AI translation is a clean illustration of a rule that holds far beyond language: a model that is excellent on average can be dangerous at the margin, and the margin is where the stakes live. The fluency that makes machine translation so useful for the ninety per cent of content that does not matter much is exactly what disguises its failures on the ten per cent that matters enormously. Assurance is the discipline of telling those two categories apart — deciding deliberately where a fast, probabilistic tool is genuinely sufficient, and where meaning is too consequential to trust to anything that cannot be held accountable for getting it wrong.