Machine translation of health materials approaches professional quality in high-resource languages and degrades sharply in low-resource ones

Narrative synthesis #

Across five studies using three different instruments (rubric scoring, blinded sentence-level rating, CIoL assessment), machine translation of written health materials approaches professional quality for high-resource target languages (Spanish, Chinese) but degrades sharply for low-resource ones. The low-resource failure is independently confirmed on the same language: Das found Google Translate least accurate for South and Southeast Asian languages including Bengali, and Hibbs found a machine-translated Bengali SACT booklet introduced 11 critical errors and failed formal CIoL assessment — the one place two groups converge on a single language. Chen's Chinese fluency result is the softest evidence here (two sentences only).

Merge candidates for a human pass (proposal only — do not touch): C-0197ClaimC-0197Initial AI draftUnsupervised machine translation fails to meet professional-quality standards for medical safety information in most non-English languagesUnsupervised free machine translation (e.g., Google Translate) fails to reach professional-quality translation standards for medical safety information across most non-English languages, with a large fraction of output c…, C-0198ClaimC-0198Initial AI draftMachine translation of patient-education material into Spanish preserves information and meaning as accurately as professional translationUnsupervised machine translation of written patient-education material into a high-resource language (Spanish) preserves the information content and the meaning/intent of the source as accurately as professional human tr…, C-0133ClaimC-0133Initial AI draftUnsupervised machine translation of medical information into low-resource languages introduces more meaning-changing errors than professional translationUnsupervised machine translation of written medical information into low-resource languages (e.g. Bengali) introduces more critical, meaning-changing errors than professional human translation., C-0201ClaimC-0201Initial AI draftMachine translation of patient-education material into a high-resource language does not introduce more clinically dangerous errors than professional translationUnsupervised machine translation of patient-education material into a high-resource language (e.g. Spanish) does not introduce more serious, clinically dangerous errors than professional human translation.. This is drafted apart from the companion mechanism claim (Machine translation of health text degrades readability before it degrades meaning…), with which it shares four EVDs; collapsing the two into one conditional-quality claim is a reasonable maintainer call.