Can AI Translators Fully Replace Humans? Not Yet, Experts Say

French speakers have a saying for a downpour: “Il pleut des cordes” — literally, “it’s raining ropes.” For a machine translation system, that phrase is a minefield.

François Yvon is a research director at the CNRS, France’s National Centre for Scientific Research, where he specialises in automatic translation. He says the expression illustrates a dilemma facing AI systems every day. A well-functioning system knows perfectly well that ropes don’t fall from the sky, so it swaps in the English equivalent, “it’s raining cats and dogs.” But, Yvon asks, “is it going to find the equivalent when it’s translating into Indonesian?”

A weather remark carries little risk. The stakes are rising elsewhere, though, as machine translation takes on an ever-larger share of global communication. Google reported in April that users now translate a trillion words a month through Google Translate alone, a figure that excludes DeepL, Microsoft, Claude, ChatGPT, Gemini and countless bespoke corporate systems. Direct sales of machine-translation tools, separate from general AI subscriptions, were worth roughly $1.4bn (£1bn) last year.

From hobby project to Wikipedia’s biggest contributor

Jonathan Zimmermann founded OKA (Open Knowledge Association), a Swiss non-profit that translates Wikipedia entries into other languages. He argues raw AI output is now perfectly adequate for low-stakes material such as internal paperwork or product listings. An encyclopaedia, he says, is a different matter: “Wikipedia runs on trust. A wrong fact can sit there for years and get cited by journalists and AI systems.”

OKA has published more than 10,000 articles, mostly in Spanish, Portuguese and English, making it, by Zimmermann’s account, the largest single contributor of translated pieces to the encyclopaedia, a feat only possible with AI assistance. The process starts with a machine-generated draft that a human translator then checks line by line, since doing the job entirely by hand would cost millions. Even so, Zimmermann says that human layer is shrinking: a couple of years ago translators using tools like DeepL still reworked large portions of each draft; today the output frequently needs no 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,” he says.

Where mistakes still slip through

Translators insist the answer is not yet. Ariella Baltié, who translates court documents from English into French, was recently asked to check papers a client had already run through AI. She found two serious mistakes near the very start: the legal term “le demandeur” (the plaintiff) rendered as “the requester”, and “Madame X, épouse Y” — Ms X, wife of Y — garbled into “Ms X marries Y”, mistaking a noun for a verb entirely. Such slips are more than embarrassing; in the US, errors in machine-translated asylum paperwork have reportedly delayed the release of detainees.

In medicine, the margin for error is even smaller. Carol Bereuter, a specialist medical translator, once found a system had rendered the Greek letter μ as “milli” instead of “micro”, turning a microgram dose into a milligram one, a thousandfold difference. She also warns that spotting such errors has become harder as the technology has improved: fluent, natural-sounding output no longer gives itself away with the clunky phrasing that once flagged a bad translation.

Context remains AI’s weak spot, Yvon notes, a machine translator asked to “translate into Polish” might just as easily think the sentence concerns shoe polish. A study OKA ran last year, funded by Wikimedia and covering 119 articles across ten language pairs, found that 27% of AI-generated text still needed correction before publication, three-quarters of it to fix errors the AI itself had introduced, with 5-6% of mistakes altering the original meaning entirely. Yvon doubts such systems will ever hit 100% reliability, since they are built on probability rather than certainty, and what’s statistically likely isn’t always what’s correct.

Source: BBC News