Machine translation is now good enough that skipping the human editor has become a "real temptation" across the industry, even as errors in law and medicine can still change meaning, the BBC reported. The French idiom "il pleut des cordes" translates word for word as "it's raining ropes." François Yvon, a research director in automatic translation at France's CNRS, said a well-trained system should instead use "it's raining cats and dogs" in English — and then asked whether it would find the right equivalent in Indonesian. In April, Google said people were translating a trillion words a month with Google Translate, a figure that does not include DeepL, Microsoft, Claude, ChatGPT, Gemini, or custom corporate systems. Direct sales of machine translation, outside general AI subscriptions, were estimated at $1.4 billion last year.
Jonathan Zimmermann, founder of the Swiss nonprofit OKA, which translates Wikipedia articles, said raw AI output can be the right call for internal documents or product listings, where multiplying costs by 100 to catch the last few percent of errors is not rational. An encyclopedia is different, he said, because "Wikipedia runs on trust" and a wrong fact can sit for years and be cited by journalists and AI systems. With more than 10,000 published pieces, mostly into Spanish, Portuguese, and English, OKA says it is the largest contributor of translated Wikipedia articles — a scale it could not have reached without AI drafts that a human then checks sentence by sentence. Two or three years ago, with tools such as DeepL, translators still rewrote a good chunk of every draft. "Today the prose often doesn't need fixing at all," Zimmermann said.
Legal and medical translators argue it is not time to drop the human. Ariella Baltié, who translates court documents from English into French, was asked to check a file already run through AI and found two major errors at the start: "le demandeur" (the plaintiff) rendered as "the requester," and "Madame X, épouse Y" turned into "Ms X marries Y," mistaking a noun for a verb. In the United States, the BBC noted, errors in machine-translated asylum applications have delayed the release of detainees. Carol Bereuter, a medical specialist, described a text that rendered the Greek letter mu as milli instead of micro — a thousandfold dose error. As the prose has grown more fluent, she said, mistakes are harder to spot. A Wikimedia-funded OKA study of 119 articles, 10 language pairs, and more than 1,000 hours of work found that 27% of AI-generated text was still corrected before publication, that three-quarters of those corrections were errors the AI introduced, and that 5 to 6% of errors changed the original meaning.
Yvon said the systems still work by probability, which is not the same as being correct, and that a sentence such as "translate into Polish" can be read as a request about shoe cream. He expects later tools to critique and rewrite their own output, but argued that the real breakthrough will come when they can also say when they are not sure. The verified picture is rapid, cheap machine translation at huge scale, and a still-unresolved question about when the remaining errors are acceptable.