AI produces the first draft in most multilingual projects today. That has turned expert human validation from a quality preference into a documented compliance obligation on both sides of the Atlantic. Here's how the workflow runs, and what each step protects you from.
AI Didn't Remove the Work in Multilingual Translation Workflows. It Moved It.
Until recently, most of the effort in a multilingual project went into producing the first version. Drafting, adapting, and formatting content for each target language absorbed the bulk of the timeline, and everything downstream, from review to layout and publication, was planned around it.
That has changed. Neural machine translation and large language models now generate a usable first draft in minutes, across dozens of languages at once. Authoring platforms have absorbed the capability directly; e-learning tools, content management systems, and documentation platforms increasingly ship with one-click translation built in.
But the effort didn't disappear when the first draft got faster. It shifted from producing content to verifying it.
When the first draft is instant, the question stops being "can we produce this?" and becomes "can we demonstrate this is correct?" In regulated content, that is not a linguistic question. It's a domain question, which is why subject-matter expert (SME) validation, a service the industry had quietly folded into other line items, is now appearing at the center of multilingual programs again.
SME Validation vs. Post-Editing vs. Linguistic Validation: What Each One Actually Means
Three services get used interchangeably, and the confusion has consequences: organizations buy one and expect another.
Service
What Happens
Who Performs It
Post-editing
A professional linguist edits machine output, correcting grammar, terminology, fluency, and mistranslation.
Trained translator / post-editor
SME validation
A domain expert confirms the content is factually and technically correct, safe, and appropriate for the intended audience in that language.
Practicing expert in the field, with target-language fluency at a professional register
Linguistic validation
A formal, multi-stage methodology, forward and back translation, clinician review, cognitive debriefing with real respondents, reconciliation, used for clinical outcome assessments and patient-reported outcome instruments.
Specialized team plus recruited respondents
The distinction that matters: a post-editor checks whether the translation says what the source says. An SME checks whether what it says is true, safe, and usable in practice. Those are different questions, and a fluent, grammatically flawless translation can fail the second one completely.
Why AI Translation Quality Control Matters More Than It Used To
Fluency and Accuracy Have Come Apart
This is the shift most organizations haven't fully absorbed. Machine output is fluent in every language it supports, including the ones where it's least accurate. The two qualities used to travel together: a bad translation read like a bad translation. That signal is gone.
The practical consequence is counterintuitive. In the languages where AI performs best linguistically, residual errors are harder to catch, not easier, because there's nothing in the surface of the text to flag them. A reviewer without the right expertise, reading polished prose, has no reason to stop. The errors that survive AI translation are precisely the ones that read well.
Research supports a hybrid model over either extreme. A 2025 study in npj Digital Medicine compared standalone GPT-4o, professional human translation, and a human-in-the-loop workflow where linguists reviewed and edited AI output. The hybrid model was rated best overall, preferred by evaluators 46.5% of the time versus 28.4% for professional-only translation, and it was substantially faster. The study also found that machine performance varied significantly depending on how much digital training data existed for a given language, which is why a single uniform process across a language set tends to over-serve some languages and under-protect others.
The Regulatory Floor for Machine Translation Moved
Rules written before generative AI are now being applied to it, and newer ones address it directly. In several frameworks, human review of machine-translated content is no longer a quality choice: it's the condition under which using AI is permitted at all.
Framework
What It Requires
Where Validation Fits
Section 1557, Affordable Care Act (U.S.) — 45 CFR § 92.201(c)(3)
Where a covered entity uses machine translation and the text is critical to an individual's rights, benefits, or meaningful access, where accuracy is essential, or where the source is complex, non-literal, or technical, the translation must be reviewed by a qualified human translator.
Human review of MT is a regulatory floor, not an upgrade.
EU AI Act — Article 50(4) (transparency obligations applicable since August 2, 2026)
AI-generated text published to inform the public on matters of public interest must be disclosed as artificially generated, unless it has undergone human review or editorial control and a person holds editorial responsibility for the publication.
Validation is the named exemption. The review must be substantive, and someone must own it.
EU MDR / IVDR
Labelling, instructions for use, and related documentation must be available in the official language(s) of each member state where the product is placed on the market, and must be accurate and comprehensible.
Language errors here are regulatory findings, not customer-service issues.
ISO 17100 / ISO 18587
Translation services require a revision step by a second qualified person; post-editing of machine output requires post-editors with full translator competences.
Gives you a recognized process to point to when asked how AI output was handled.
HIPAA / GDPR
Govern where and how sensitive content may be processed and retained.
The engine belongs in the compliance review, not only its output.
Read together, these converge on a single principle: AI may produce the content; a qualified human must be accountable for it. Validation is how that accountability is created and evidenced.
The AI Translation Validation Workflow, Step by Step
A validation-centered workflow isn't the traditional process with a reviewer added at the end. The sequence changes, because decisions that used to happen during translation now have to happen before the engine runs.
1. Prepare the source. Every defect multiplies by the number of target languages. Review the source for idioms and culturally specific references, region-specific regulatory or procedural content, text embedded in images and diagrams that automated pipelines can't reach, and narration scripts that have drifted out of sync with on-screen text. Protects against: systematic errors replicated across every language at once.
2. Build terminology assets before translating. A locked bilingual glossary and a do-not-translate list (product and protocol names, interface strings, regulatory citations, measurement formats) should exist before the AI pass. Modern engines accept a glossary as an input, so the machine makes the right choice the first time, converting expert review from rewriting into confirming. Protects against: terminology inconsistency across modules, reviewers, and future projects.
3. Translate with configured settings, not defaults. Target locale rather than just language. Register and formality. Approved translation memory applied first, so previously validated content is reused rather than regenerated. And a documented decision about which engine processes the content and where that data goes. Protects against: data-handling exposure under HIPAA and GDPR, and avoidable rework.
4. Post-edit where the language requires it. Not every language should go straight from engine to expert. Where machine performance is weaker, a professional linguist should complete a full post-edit first; otherwise, the SME spends their time on basic linguistic correction instead of domain judgment. Protects against: misallocating the scarcest resource in the workflow.
5. Validate in context. The expert reviews the content in the environment where the end user will encounter it, not in a bilingual spreadsheet. Context changes meaning: a term correct in isolation can be wrong for the screen it appears on; a sentence that reads correctly can contradict the diagram beside it; an accurate warning can sit where nobody will read it in time.
Modern in-context review platforms support this directly: source and target side by side, live preview of each edit in the finished layout, section-level progress tracking, and handling for audio scripts and captions alongside on-screen text. Two things they generally don't cover, and need a parallel workflow, are content the automation can't reach (recorded narration, video, text inside images, attachments, right-to-left layout in custom components) and the return path: in many platforms, validated changes are suggestions that an author must manually import back into the source file.
Validation should also be measured against defined criteria agreed in advance: typically zero tolerance for errors that are factually wrong, unsafe, or reversed in meaning; zero tolerance for terminology and regulatory deviations; and reviewer discretion on style. This is what converts a subjective review into a documentable deliverable. Protects against: the errors that matter, and the inability to prove you looked for them.
6. Reconcile, import, and verify. Decide what changes to implement, import them, and check the result again: layout integrity, character rendering, text expansion, working links, audio-to-text alignment. Protects against: approved changes that never reach the published version, or that introduce new problems on the way in.
7. Close the loop. Feed approved segments back into translation memory and the glossary. Validation that ends at "approved" leaves nothing behind; validation that updates the assets reduces the volume requiring review on every project that follows. Protects against: reviewing the same content again with no record of what was already decided.
The Bottom Line
AI didn't make expertise less necessary. It made it more visible by removing everything around it.
When the first draft absorbed most of the timeline, expert review sat at the end as one compressible line among many. Now that the draft arrives immediately, validation is the process, the point where accuracy is established, risk is contained, and an organization's ability to say "we verified this, and here is who is responsible" actually comes from.
Regulators on both sides of the Atlantic have arrived at the same conclusion from different directions. The workflow above isn't a way of slowing AI down. It's what makes the speed defensible.
Arcadia is a global multilingual services company with operations in Argentina, Spain, and the United States. ISO 9001- and ISO 17100-certified and HIPAA-compliant, we design and run hybrid AI + human validation workflows for healthcare, e-learning, technical, and regulated content in 50+ languages. If you're building or reviewing a multilingual workflow, we're happy to walk through it with you.
Sources
"Evaluating human-in-the-loop strategies for artificial intelligence-enabled translation of patient discharge instructions." npj Digital Medicine (2025).
"45 CFR § 92.201 — Meaningful access for individuals with limited English proficiency." eCFR.
"Section 1557 of the Affordable Care Act and Language Access." American Translators Association.
"EU AI Act, Article 50 — Transparency rules" (applicable August 2, 2026).
"EU AI Act transparency obligations take effect 2 August 2026." Cooley.
"Complying with European Medical Device Regulation (MDR) language translation requirements." Language Scientific.
"ISO 18587:2017 — Translation services: post-editing of machine translation output."
"ISO 18587 update: what's changing in post-editing machine translation." ATC Certification.