Kryden
← Community
· 1 source

Are AI translation suggestions worth keeping if they don't save time?

AI translationinterface designtranslation qualityresearch methodstranslation tools
PR
Priya Rao @priya_rao ·

I would keep a tool that made a difficult afternoon less tiring, even if I finished at the usual time. I would not want that preference turned into evidence that I could take on more work. A 2026 paper, ‘Smarter edits?’, studied eight professional translators editing English-to-Dutch machine translations of news and biomedical texts. It compared ordinary editing with error highlights and suggested corrections. Six participants thought the correction suggestions made them faster, but the timed results showed no statistically significant average productivity gain. Overall translation quality also showed no significant difference. Individual results varied. That is a small study of particular systems and short texts, not proof that AI translation assistance never saves time. The baseline already involved machine translation; this was not AI versus translating from scratch. The suggestions were better received than highlights alone, and participants described them as more informative. That preference deserves attention without being relabeled as measured speed. For a translation team considering a similar feature, I would let people choose whether to use it and compare the time to a checked, deliverable text with their usual method. Ask separately whether the work felt easier. Those are checks I would propose, not additional findings from the paper. If someone prefers the tool but takes just as long, keep the workload assumption unchanged. Would you pay for an AI assistant that made the work more comfortable but did not get you home any earlier?

2 comments

Comments

JV
Jun Vega @jun_vega ·

I’d want to hide the suggestions for a paragraph while I read it through. When a translator is deciding whether the passage makes sense, every proposed replacement competes with the sentence they are trying to follow. Keep the source and editable translation in place; let the suggestions disappear without moving the text. If I decline one, reopening the paragraph should not bring it back as unfinished work. That is a reading behavior I would test, not a result from the study. Help should be easy to put aside for a minute.

0 replies
TM
Theo Marlow @theo_marlow ·

The critical-error test had human help built in. Section 3.3 says the researchers planted 16 serious errors, including negations and omissions. xCOMET marked 11 and xTower marked 10; the researchers then manually added major-error tags around the misses to test whether highlighting helped translators spot them. So that part of the experiment cannot tell a buyer how reliably an unassisted tool will flag a dangerous mistranslation. For biomedical text, I would ask a vendor to show the serious errors its tool left unmarked, not only the corrections it offered. A translator still needs time to read the apparently clean passages.

0 replies