Machine translation is everywhere, but experts say people still matter for critical texts

Artificial intelligence has taken over much of the translation business, but people remain necessary where mistakes carry real consequences, according to a report by BBC News. Researchers and professional translators told the broadcaster that machine output has improved sharply, while its remaining errors have become harder to detect.

Scale of machine translation

The volume is enormous. Google said in April that users of its translation service process one trillion words each month, and that figure leaves out rival tools, general AI chatbots and custom systems built for companies. Direct sales of machine translation, counted separately from broader AI subscriptions, were estimated at $1.4bn last year.

For routine material such as internal documents or product listings, some in the field now consider unedited AI output a sensible choice, since paying for full human checking costs far more than the small gain in accuracy justifies.

Where errors still appear

A study run last year by the Open Knowledge Association, a Swiss non-profit that translates Wikipedia articles and was funded by Wikimedia, looked at 119 articles across 10 language pairs. It found that 27% of AI-generated text was corrected before publication, and that between 5% and 6% of errors altered the meaning of the original.

Specialists in legal and medical translation described cases where machines mistranslated legal terms or confused units of measurement, which in medicine could mean a dose a thousand times too large. They added that because modern systems produce smooth, natural-sounding sentences, such slips no longer stand out the way clumsy grammar once did.

François Yvon, a research director at France's CNRS, said the probability-based way these systems choose words makes complete reliability hard to imagine. He expects future tools to combine several AI models that translate, critique and revise one another's work, but argued the key advance will come when systems can signal uncertainty. In his words, the breakthrough requires them to start "telling us when they're not sure."

What it means for technology users

For people who rely on online translation tools, the findings suggest a practical division: automated output is often adequate for everyday or low-risk content, while contracts, medical information, official paperwork and published reference material still benefit from review by a qualified human. As the tools become more fluent, checking important translations carefully becomes more, not less, important.