Where machine translation is genuinely usable in a regulated content workflow, and where relying on it creates real clinical and reputational risk.
Machine translation has become good enough that treating it as unusable across the board is no longer a serious position, but treating it as safe everywhere is not one either. For internal drafts, UI strings and low-stakes navigational text, current systems are fast, cheap and accurate enough that a full human translation pass adds cost without adding real value. For prescribing information, dosing instructions and anything a patient or clinician could act on directly, the risk is not a slightly awkward sentence but a clinical safety issue, and no machine system should be trusted unreviewed. Often the more useful question is structural: how much of what looks like a translation problem is actually a content architecture problem that should not need translating independently in the first place.
Machine translation has improved enormously for general content, but a mistranslated dosage instruction or a subtly altered clinical claim carries consequences a mistranslated marketing tagline does not. In most industries, an imperfect translation costs a bit of polish; in pharma, it can mean a patient reading the wrong instruction or a regulator receiving a claim that was never actually approved in that market. The stakes, not just the accuracy percentage, are what should drive the decision here.
Not every piece of content carries the same risk, and treating a UI string the same way as a package insert wastes money in one direction and takes on unnecessary risk in the other. Two patterns cover most of where machine translation genuinely earns its place.
Internal draft content intended for human review before publication, and lower-stakes content such as UI strings or navigational text where a human reviewer will catch and correct any error before it reaches a patient or clinician, is exactly where current machine translation earns its keep. It is fast, cheap and, for this class of content, accurate enough that adding a professional human translation pass on top adds cost without adding real value.
The most cost-effective pattern we see is machine translation used as a first draft even for higher-stakes content, followed by a qualified human reviewer correcting and approving it. Used as a first-pass draft that a qualified translator then reviews, it can genuinely speed up the workflow without compromising accuracy: faster than translating from scratch, and safer than publishing machine output unreviewed, provided the reviewer is a genuine subject-matter translator and not a general proofreader skimming for fluency. This pattern maps closely to the process defined in ISO 18587, the international standard for post-editing of machine translation output, which sets out both the post-editing requirements and the competences a qualified post-editor needs.
The same technology, applied to a different content class, stops being a productivity question and becomes a safety and compliance question.
Prescribing information, dosing instructions, and any content a patient or clinician might act on directly without independent verification sit in a different category entirely. The cost of an error here is not a slightly awkward sentence, it is a clinical safety issue, and no current machine translation system should be trusted unreviewed for this content class. The same applies to anything touching an approved claim or a regulatory statement: the risk is a subtly incorrect medical claim reaching a market without review by an appropriately qualified local reviewer.
As a rough rule of thumb:
The practical rule is simple: machine-translate structure and UI, human-translate anything a regulator or a patient could act on.
Whether the underlying content model actually separates shared technical facts from translatable copy matters more than the translation method itself — see multilingual content architecture. When a specification, a dose, or an authorisation status is stored once and rendered consistently in every language, the translation-accuracy question disappears entirely for that data, because there is nothing left to translate independently per market. Only copy that legitimately varies by language remains, and that is a much smaller and lower-stakes translation problem than re-translating the same facts over and over across markets.
Machine translation as an accelerant for draft content, human professional translation and review for anything clinical or regulatory, and, most importantly, a content architecture that minimises how much genuinely needs translating in the first place by treating facts as shared data rather than as prose translated independently per language. Get the structure right and the translation question shrinks down to the part that was always going to need a human anyway: the copy that legitimately varies by market, not the facts that should not have been drifting apart between languages to begin with.
No — the work is building the content structure and workflow that translated content lives in, alongside your translation supplier or local teams rather than in place of them. That means separating shared technical facts, like a dosage or an authorisation status, from copy that legitimately varies by market, so translation effort concentrates only on the part that needs to differ.
Not unreviewed — a mistranslated dosage instruction or an altered clinical claim carries a safety consequence a mistranslated marketing tagline never would, so no current machine system should be trusted without a qualified human reviewer for this content class. A hybrid workflow, where machine translation produces a first draft that a subject-matter translator then reviews, can speed up the process without giving up that safety check.
Field-level translation settings let a specification or dosage field be locked so every language pulls the same shared value, while the headline or body copy field next to it stays open for independent translation. This is what stops a fact from drifting between languages once someone edits one version without updating the others. See WPML and translation for how that field-by-field setup connects to a translation supplier’s workflow.
Machine translation quality still varies meaningfully by language pair, generally performing stronger on major European languages and weaker on less commonly trained pairs, so quality is worth testing per language rather than assumed uniform across a whole multilingual rollout. That variation is exactly why the hybrid workflow, machine draft plus qualified human review, holds up better across an entire market set than trusting raw output everywhere equally.
Tell us your languages and content types and we will tell you where translation risk actually sits.