How long will humans be needed for translation?

A French idiom about rain shows the problem neatly. “It’s raining ropes,” translated literally from “Il pleut des cordes,” makes no sense in English unless the system knows to swap it for “it’s raining cats and dogs.” François Yvon, who studies automatic translation at France’s CNRS, says that gets tricky fast once the target language changes and there isn’t an obvious equivalent to reach for.

And machine translation is no longer a niche tool. In April, Google said people were translating one trillion words a month with Google Translate. That figure doesn’t include DeepL, Microsoft, Claude, ChatCGT or Gemini, nor bespoke systems built for companies. Outside general AI subscriptions, direct sales of machine translation were estimated at $1.4bn last year.

For a lot of ordinary work, that’s enough. John Zimmerman of OKA, a Swiss non-profit that translates Wikipedia articles, says raw AI translation is often the sensible choice when the job is internal documents or product listings. But Wikipedia is a different beast because it runs on trust, and a bad translation can linger for years and end up cited by journalists and AI systems. OKA has published more than 10,000 pieces, mostly into Spanish, Portuguese and English, and Zimmerman says that scale would have been impossible without AI.

Still, the human is far from gone. OKA’s process uses AI for the draft and then a person checks every sentence, though Zimmerman says the need for heavy rewriting has fallen sharply in just two or three years. In legal translation, Ariella Baltié says AI recently turned “the plaintiff” into “the requester” and mangled “Madame X, épouse Y” into “Ms X marries Y.” In medicine, Carol Bereuter points to a system that misread the Greek letter μ as milli instead of micro, a tiny slip with huge consequences.

The broader problem is that modern systems can sound right while being wrong. Yvon says the old clunky giveaway is fading, which makes errors harder to spot. An AI can even get lost on context and decide that “translate into Polish” is about shoe cream, or miss that a football “corner” is not furniture. OKA’s Wikimedia-funded study of 119 articles across 10 language pairs found that 27% of AI-generated text still needed correction before publication, and 5 to 6% of errors changed meaning. That is not close to safe enough for anything where the signature matters.

Yvon thinks the next step is not magic reliability but systems that can check each other and say when they are unsure. That sounds less glamorous than instant perfect translation. It also sounds a lot more honest.