The French say "Il pleut des cordes" when it's raining hard - literally "it's raining ropes." It's the kind of idiom that sends AI translators into a tailspin. "You can imagine the dilemma this throws AI into," says François Yvon, a research director specialising in automatic translation at the French National Centre for Scientific Research (CNRS). A well-trained system, he notes, will know ropes aren't literally falling from the sky and offer "it's raining cats and dogs." But can it find the equivalent in Indonesian? Yvon wonders.
Weather commentary is low-stakes, but machine translation is increasingly handling far riskier material. In April, Google announced that people are now translating one trillion words a month with Google Translate - and that's before counting DeepL, Microsoft, Claude, ChatCGT, and Gemini, plus bespoke corporate systems. Direct sales of machine translation outside general AI subscriptions were worth an estimated $1.4bn (£1bn) last year.
"For plenty of use cases, raw AI translation is now the right call: multiplying your costs by 100 to catch the last few percent of errors isn't rational if you're translating internal documents or product listings," says John Zimmerman, founder of OKA (Open Knowledge Association), a Swiss non-profit that translates Wikipedia articles. "For an encyclopedia it's different. Wikipedia runs on trust. A wrong fact can sit there for years and get cited by journalists and AI systems." With over 10,000 published pieces, mostly into Spanish, Portuguese, and English, OKA is, Zimmerman says, the single largest contributor of translated articles to the collaborative encyclopedia - something impossible without AI. "Translating 10,000 articles the traditional way would cost millions. In our process, the AI produces the draft, then a human translator checks every sentence."
But that human touch is fading. "The machine output is now good enough that skipping the human entirely has become a serious option, and a real temptation across the industry. Two or three years ago, with tools like DeepL, our translators still rewrote a good chunk of every draft. Today the prose often doesn't need fixing at all. So, the question is no longer whether AI translation is good, it clearly is, but whether it's good enough to remove the human," Zimmerman says.
Ariella Baltié, a translator and legal expert who specialises in court documents from English to French, was recently asked to check AI-translated papers. "There were two major errors right at the beginning," she says. The AI rendered "Le demandeur" (the plaintiff) as "the requester" - "totally awkward for legal papers." Worse, "Madame X, épouse Y" (Ms X, wife of Y) became "Ms X marries Y," mistaking a noun for a verb and mangling the meaning. In the US, machine-translation errors in asylum applications have delayed detainee releases.
In medicine, accuracy is life or death. "The units for dosages and blood tests often include Greek letters," says Carol Bereuter, a medical translation specialist. She found a text where the Greek letter μ was translated as milli (one thousandth) instead of micro (one millionth). "That's a serious error. If you give someone a milligram when it's meant to be a microgram, you are giving them a thousand times the dose." As AI improves, catching errors gets harder. "Whereas in the past, with things like Google Translate, the sentence would jump out at you because there'd be a grammatical error or it would sound clunky, today it sounds perfectly legitimate. It sounds fluid. It sounds natural. So the errors are a lot harder to spot."
"If it's something that will bear your signature, there's a risk for your reputation. And if decisions are going to be taken on the basis of a translation concerning a medical procedure, a contract you might sign or a major purchase you might make, the risk is even greater," warns Yvon. Even clear sentences can trip up AI: "When you have words with different meanings, it's usually very clear for a human which one applies because they understand the context, but when a machine translator comes across the sentence 'translate into Polish', for example, it is quite capable of deciding it must be all about shoe cream," he says. Context is crucial - without grasping that a football match is being discussed, AI might mis-translate "taking a corner."
Last year OKA ran a Wikimedia-funded study of 119 articles, 10 language pairs, and over 1,000 hours of translator work. "27% of the AI-generated text still got corrected before publication. Three quarters of those corrections were errors the AI introduced, and 5 to 6% of errors changed the meaning of the original text," Zimmerman says. Eradicating those errors is tough. "As things stand, I think the fundamental way these systems function makes it quite hard to imagine how they could become 100% reliable," Yvon says, because AI translates based on probability - what's probably correct isn't the same as what is correct, especially with language's subtleties.
Yvon foresees improvement from a new generation of tech: "We will see translation systems that are increasingly capable of evaluating what they do and correcting themselves using different AI tools simultaneously: one to do the translation, a second to criticize it and a third to produce a new version, for example." But the real breakthrough will be when AI can express doubt. "The real breakthrough won't come until these systems are not merely capable of translating but also of telling us when they're not sure," Yvon says.