Machine translation is probably already in your regulatory workflow, even though no one approved it. A distributor’s email goes into a browser tab, and a complaint from another market goes through a free tool to see what it says. Each is a reasonable shortcut, and each happens outside your quality system with no controlled tool, no record, and no review.
The trouble starts when that same shortcut drifts toward something a regulator will read: instructions for use (IFU), a label, a field safety notice. Now somebody wants to know how the translation was produced, who checked it and why anyone trusted it, and “the tool is quite good these days” won’t cover it.
So, this post is about where AI translation fits regulatory teams, and where it doesn’t. We’re in favor of it for the right jobs, and we’ll say which they are. But it doesn’t replace a controlled workflow. Qualified linguists, managed terminology, human review or documented validation, and the rules around device information assume somebody is doing that work.
What AI Translation Is Good at
It’s fast, and regulatory teams are rarely short of things to translate.
It’s at its best when the aim is to understand something rather than publish it. A complaint lands in Portuguese. A literature search throws up 40 papers in four languages. A distributor asks a question, and you need to know by lunchtime who should answer it. A rough machine draft tells you what you’re looking at and who to hand it to, and that’s often all you need from it.
It also earns its keep when a qualified linguist is going to work on the output anyway. If the engine starts from your approved terminology and translation memory, the linguist spends less time typing and more time on the calls that need a human: is “adverse reaction” the right term here, and does that warning still read like a warning?
A rough rule is that the closer a text gets to someone who has to act on it safely, the less AI should be doing alone. In practice:
Content | What AI can do | What still needs a person |
Incoming emails, complaints, literature | Give you the gist and help you sort what matters | Anything you act on goes to a qualified linguist or a subject expert |
Internal working documents | Produce a first draft | A reviewer who knows the product |
IFU, labels, packaging, patient-facing material | Produce a draft inside a controlled workflow | A qualified medical linguist, approved terminology, independent review, sign-off |
Field safety notices, clinical and safety documents | The same, often against a deadline | The same checks, plus a record you could show an auditor |
Fluent Isn’t the Same as Right
Obvious machine translation mistakes get caught. The fluent ones don’t. A unit slips, a negation drops out, “contraindication” softens into something vaguer, and the sentence around it still reads perfectly well, so nobody stops to look.
A controlled workflow comes first, and it’s less glamorous than it sounds. Someone has decided which engine is used for which type of content, who sets it up, what happens when the source changes and who signs at the end. Without that, the same document gets translated three ways in one month depending on who got to it first.
People are the harder part. Qualified linguists who know the language and the subject, working to a recognized standard. ISO 17100 is the usual reference for translation services and expects a second qualified person to check the first one’s work, while ISO 18587 covers full human post-editing of machine output. The reviewer matters more than it might seem. They need the knowledge and the authority to reject a translation, and they need to read it against the source and against the person it’s written for. A surgeon and a lay patient need different language, and the tool can’t tell which of them is holding the page.
Terminology drifts if nobody owns it. Device names, components, warnings and symbols need approved terms, held in a term base and a translation memory and used the same way every time. Left alone, an engine will quite happily say “single-use” in one paragraph and “disposable” in the next, and your user (and your notified body) will wonder whether you mean the same thing.
Validation is the easiest piece to put off. How did you decide that this engine, in this workflow, was good enough for this content? Representative samples tested, error types logged, acceptance criteria agreed in advance, all written down. Engines also change without much warning, so the record needs revisiting when they do, and again when the content changes, say from a packaging insert to a software interface. “We looked and it seemed fine” isn’t an answer if someone asks.
What MDR Says About IFUs, Labeling and the Rest
The Medical Device Regulation (MDR) doesn’t mention AI translation, and it doesn’t tell you how to translate anything. The Johner Institute makes the same point: neither MDR nor the In Vitro Diagnostic Regulation (IVDR) sets explicit requirements for translating an IFU, but the IFU still has to be understandable to the intended user in every target language. It points to ISO 17100 and IEC/IEEE 82079-1 as the standards people tend to use. So you have an outcome to deliver and a method you’ll have to defend.
Language comes first, and it’s decided country by country. Article 10(11) requires the IFU to be supplied in the official language(s) of the country where the device is made available, and Member States fill in the detail. Germany, for example, only allows devices to be supplied to users and patients if the information meant for them is in German (Johner Institute). Fimea in Finland accepts Finnish, Swedish or English for most information but wants Finnish and Swedish for anything needed for safe use and leaves the manufacturer to decide through a risk analysis what that covers. Same regulation, different answers, which is why “translated” can’t be a single checkbox.
The intended reader matters just as much. Annex I, section 23.1(a) says the IFU must be “written in terms readily understood by the intended user” (Johner Institute). A translation can be grammatically clean and still miss that standard, with clinical shorthand in front of a lay patient or a casual tone in a warning meant for surgeons. An AI tool doesn’t know who’s reading unless someone tells it, and even then, someone has to check.
Then there’s what happens when the information is wrong. Under Article 2(64), an “incident” includes “any inadequacy in the information supplied by the manufacturer” (Fimea). Whether a particular translation error is reportable is a call for your vigilance team. But a dropped precaution or a wrong value in one language looks a lot like the inadequacy that definition describes, so agree now on how you would assess one.
What the EU AI Act Changes (and Doesn’t)
Short version: the AI Act probably isn’t about your translation tool, but its way of thinking is worth borrowing.
The dates have moved, so check the ones in your planning documents. The Digital Omnibus on AI, Regulation (EU) 2026/1744, took effect on July 27, 2026 and pushed back the main high-risk deadlines. Stand-alone high-risk systems now have until Dec. 2, 2027. AI in products covered by sectoral legislation, medical devices included, has until Aug. 2, 2028 (AI Act Explorer). If your own devices contain AI, that’s a separate conversation for your notified body and your legal team, and how the Act and MDR fit together there is still being worked out.
A translation tool your regulatory staff use is another matter. It isn’t part of a device, and it doesn’t obviously land in the Act’s high-risk categories, although your legal colleagues should confirm that for your setup. A couple of parts of the Act still touch how you work. Article 4 on AI literacy has applied since Feb. 2, 2025, and as amended it requires providers and deployers to “take measures to support the development of AI literacy of their staff,” without guaranteeing a specific level for any one person. For a regulatory team that means asking whether the people choosing, using and checking AI translation know what it does well and where it breaks, and training records are a sensible start. The Article 50 transparency obligations have applied since Aug. 2, 2026 and are mostly about AI that talks to people and AI-generated content, so ask legal whether anything your company publishes falls inside them.
The bit that matters most is human oversight. For high-risk systems, the Act is built around real people with the competence, training and authority to understand what the AI produced and step in when it’s wrong. That’s very close to what a qualified linguist and an independent reviewer already do.
What a Defensible Setup Looks Like
Our rule of thumb is people first, then AI, then people again. People set up the work and own it, AI carries the volume in the middle, and a person checks what comes out. In practice that’s six steps.
- A person sets the stakes. What is the content, who reads it, which markets is it for and does it suit AI-assisted translation at all? Warnings, contraindications and anything else tied to safe use go into a higher-control lane, or stay fully with human translators.
- A person prepares the ground. The source text is finalized and checked, and approved terminology, translation memory and style notes are loaded. A vague sentence in the source becomes a vague sentence in every language, so this is the cheapest place to catch problems.
- AI does the heavy lifting. The engine produces a first draft from that approved material, quickly.
- A qualified linguist post-edits. Someone with subject knowledge and the target language compares the draft with the source and fixes what needs fixing, including the fluent-but-wrong sentences.
- An independent reviewer signs off. A second qualified person checks the result, with in-country review where the risk justifies it. Their name and the date go on the record.
- A person keeps the evidence. That means the engine and version, the terminology version, who edited, who reviewed, what changed and the validation behind the whole setup. Revisit it when the engine or the content type changes.
Seven Questions for Your Language Partner
These work whether you use an outside provider or an in-house team. They’ll show you quickly how much control sits behind the words “AI-assisted.”
- Which content types do you run through AI, and which do you keep fully with human translators? What drew that line?
- Who are the linguists, and which standards do you work to? ISO 17100 for translation services and ISO 18587 for post-editing are good starting points.
- How do you manage terminology and translation memory for our products, and who approves a change?
- Who reviews the final text, and is it someone other than the person who edited it?
- What validation sits behind your AI-assisted workflow, when was it last repeated, and what happens when an engine is updated?
- What records do we receive for each translation, and how long are they kept?
- Where does our content go? Which engines process it, where are they hosted and is anything we send used to train them?
If the answer to the first one is “everything goes through AI,” you already know what you needed to.
How BIG Can Help
BIG works with regulatory, quality and medical affairs teams who want the speed of AI without losing control of the result. Our life sciences linguists translate and review medical device and pharma content from your approved terminology and translation memory. Where AI helps, it handles the volume in the middle of a process that starts with people who set the stakes and ends with people who check the result and sign it off, with the records to match.
If you’re working out where AI belongs in your regulatory content, or you’d like a second opinion on the workflow you already have, get in touch with the BIG team



