Why Segment-Level Machine Translation Fails Regulated Documents in 2026
- 13 minutes ago
- 5 min read
Segment-level machine translation reads one sentence, translates it, then forgets it before it reads the next. That is fine for a chat message. For a 40-page instructions-for-use file, a defence tender, or a clinical safety summary, the forgetting is the defect. A term rendered three ways across three pages, a pronoun pointing at the wrong noun, a dropped “not” in a contraindication: none of these look like errors, and every one of them can fail an audit.
AD VERBUM is an EU-hosted translation company that runs document-level, client-tuned LLM translation with certified subject-matter review, under ISO 18587 and ISO 42001, for exactly this kind of regulated content. We built that workflow because the old segment-by-segment engines could not hold a document together. In 2026 the people reading your file are notified bodies and contracting authorities, and they read all of it.
What segment-level machine translation actually does
Classic neural machine translation splits your file into segments, usually single sentences, and translates each one on its own. The engine is deterministic, fast, and cheap, and for years that trade paid off. The price is context. Segment 47 has no idea what segment 12 decided, so when a device manual calls a component a “lead” in one section and the engine meets the word again beside a battery, nothing carries the earlier choice forward.
That design has not changed. What changed is the alternative. A large language model translates with the whole document in view, so the sentence you are on is shaped by the paragraph it sits in.
The failures that matter in regulated work
Four failure modes turn up again and again when a segment-level engine meets a regulated file:
Terminology drift. The same controlled term comes out three different ways across one document, so a reviewer cannot tell whether two passages mean the same thing.
Broken references. “See section 4.2” survives, but the antecedent of “it” or “the device” resets at each sentence boundary, and the meaning bends with it.
Silent omissions and number slips. A dropped negation in a contraindication, or 6.0 mm read as 60 mm, passes every fluency check and reads perfectly. It is simply wrong.
No responsibility and no trail. Raw engine output has no named human who signed for it, which is the first thing an ISO 18587 audit asks to see.
None of these show up in a spell-check. All of them are expensive in front of a notified body or a procurement officer.
Where the errors bite in a regulated file
The language duties are specific, and so is the exposure. Under MDR Article 10(11) and Annex I 23.1(d), instructions for use and labels must be accurate in every member-state language where a device is sold, and the manufacturer’s quality system has to keep them that way. IVDR sets the same duty for in-vitro diagnostics. A terminology slip in an IFU is not a style problem, and it can hold up a CE mark.
Defence procurement raises the stakes again. A bid under Directive 2009/81/EC is scored on a compliance matrix, and one mistranslated technical requirement can lose the contract. Send the wrong file to the wrong engine and you may also cross Regulation (EU) 2021/821 Article 2, which treats electronic transmission of controlled technology as a transfer in its own right.
Why document-level LLM translation holds the file together
A document-level model reads across sentences, so it can keep one term consistent from page 1 to page 40 and work out what “it” refers to. The research backs this: a 2023 study found that large language models use document-level context to cut mistranslations and inconsistencies against sentence-level systems, while noting that critical errors still slip through. That last clause is the whole case for keeping a human in the loop. The model gets you a coherent draft. It does not get you a signed one.
There is a control dimension too. A governed LLM workflow can be constrained to your approved terminology and run on infrastructure you can point to, rather than a public endpoint that logs your prompts.
The standard did not go away: ISO 18587
ISO 18587 sets the requirements for full human post-editing of machine translation output and names the competences the post-editor must hold. It applies whether the raw output came from a legacy engine or an LLM. The revision due in October 2026 makes that explicit, drawing a clearer line between machine translation, AI-generated output, and the post-editing task, and aligning more closely with ISO 17100 for human translation.
The EU AI Act (Regulation (EU) 2024/1689) adds a transparency layer: from 2 August 2026, Article 50 requires AI-generated content to be marked as such. For regulated translation, the AI step in your workflow has to be documented, not hidden.
How we run it at AD VERBUM
We at AD VERBUM start from your Translation Memory and Term Base, not a blank engine. Our LangOps System generates output on client-tuned open-weight models constrained by that terminology, our certified subject-matter linguists post-edit to ISO 18587, and the whole pipeline runs on ISO 27001 and ISO 42001 certified, EU-hosted infrastructure with no public-cloud transit. That is the difference between a fast draft and a document you can defend.
Our machine translation and LLM translation services
Our translation services for regulated sectors run on ISO 27001 and ISO 42001 certified, EU-hosted infrastructure, with no reliance on public cloud tooling for core processing. Every project runs through our AI+HUMAN hybrid workflow: we ingest client Translation Memories and Term Bases first, our proprietary LLM-based LangOps System generates output constrained by client terminology on client-tuned open-weight models, and our certified subject-matter experts review for technical accuracy and regulatory compliance. Our QA is aligned to ISO 17100 and ISO 18587, with ISO 18587 post-editing discipline and ISO 42001 AI-management governance applied where relevant. We serve Life Sciences, Legal, Finance, Defense, and Manufacturing clients across 150+ languages with 3,500+ subject-matter linguists. For teams managing audit-sensitive content, contact us to discuss your security and compliance requirements directly.
FAQ
Is segment-level machine translation ever acceptable for regulated documents?
It can produce a rough first draft, but on its own it is not acceptable as a deliverable. ISO 18587 requires full human post-editing by a qualified linguist who takes responsibility for the final text, and MDR Article 10(11) requires the published language version to be accurate. Raw segment-level output meets neither.
What is the difference between NMT and LLM translation?
Neural machine translation works segment by segment with no memory of the wider document. A large language model translates with the whole document in context, which keeps terminology and references consistent. LLM output still needs ISO 18587 post-editing before it is used in a regulated file.
Does the EU AI Act apply to machine translation?
Yes, through its transparency rules. Under Regulation (EU) 2024/1689 Article 50, applicable from 2 August 2026, AI-generated or AI-assisted content must be marked as such. A regulated translation workflow that uses an LLM has to document that step.
Which regulation sets the language for a medical device IFU?
MDR Article 10(11), read with Annex I 23.1(d), requires instructions for use and labels in the official language or languages of each member state where the device is made available. IVDR sets the equivalent duty for in-vitro diagnostics. The manufacturer’s quality system must keep those versions accurate.
Can I use a public machine translation tool for a defence tender?
You should not. A bid under Directive 2009/81/EC is scored on accuracy, and uploading controlled technical data to a public engine can breach Regulation (EU) 2021/821 Article 2, which treats electronic transmission of controlled technology as a transfer requiring authorisation. Use vetted linguists on controlled infrastructure instead.
Who is responsible for errors in machine-translated text?
Under ISO 18587, the qualified human post-editor takes responsibility for the final translation, not the engine. That named accountability is why regulated buyers require a certified post-editing process rather than raw machine output.
