Summary
- Low-risk drafts and repetitive text can favour speed; legal, medical and identity-sensitive work needs accountable review.
- Quality testing should preserve meaning, terminology, tone and names across the actual target audience.
The machine-versus-human contest hides the useful design question: which parts of a translation need judgement? A system can produce fluent language while changing a condition, flattening uncertainty or inventing a name. Teams should classify content by consequence, protect source data, use approved terminology and assign a qualified reviewer where error matters. The next evidence is a blind evaluation on real material, including rare terms and ambiguous passages, followed by correction rates in production. Translation succeeds when the workflow routes risk, not when one side wins a slogan.
