Modern machine translation rarely produces obvious gibberish. It produces fluent, confident output that is sometimes wrong — and fluency is exactly what stops a reviewer from noticing. Why the failure mode changed, and what to do about it.
- The failure mode moved
- Why fluency defeats review
- What actually catches it
The failure mode moved
A decade ago, machine translation failed loudly. The output was clumsy, word order was wrong, and you could see at a glance that it needed a human. That visibility was, paradoxically, a safety feature: nobody shipped it by accident, because it obviously looked broken.
Modern neural and LLM-based translation fails quietly. The output is fluent, grammatical and confident — and every so often it is confidently wrong. It drops a negation, swaps a dosage, renders a term with its everyday meaning instead of its regulated one, or smooths over an ambiguity in the source by silently picking one reading. The sentence reads perfectly. That is the problem.
Why fluency defeats review
A reviewer’s attention is drawn to what looks wrong. When the text reads naturally, the eye slides over it, and the one clause that inverts the meaning gets the same half-second as the ten that are fine. The more fluent the output, the harder it is to review well — the surface quality actively suppresses the scrutiny the content needs.
This is worst exactly where it matters most. In a marketing blurb, a fluent near-miss is cosmetic. In an IFU, a torque spec, a contraindication or a safety warning, a fluent near-miss is a defect that has camouflaged itself as correct. The domains with the highest stakes are the ones where confident-but-wrong is most dangerous, because the reader trusts the fluent text and acts on it.
What actually catches it
Review against the source, not the target alone. Reading the translation for whether it "sounds right" is precisely the trap. The check that works is bilingual: does each segment carry the meaning of the source, verified by someone who can read both and knows the domain.
Anchor the terms. Machine output drifts on exactly the terms that matter — the ones with a specific regulated meaning. A termbase check catches the term-level errors that fluent grammar hides.
Match the check to the risk. Not all content deserves the same scrutiny; that is the whole logic of tiered service. Repetitive internal content can ride on light review. A filing, a label, or a safety instruction cannot — there, the accountable human pass is not a cost you are trimming, it is the control that stops a fluent error from reaching a reader who will trust it.