Google said in April that its Translate service now processes one trillion words every month, according to the company, a scale that has sharpened debate about how long people will remain essential to translation.
That total excludes use of DeepL, Microsoft systems, Claude, ChatCGT and Gemini alongside bespoke corporate platforms, the BBC report noted. Direct sales of machine translation outside general AI subscriptions were estimated at $1.4bn (£1bn) last year, according to estimates cited in the report.
Researchers say idioms still expose weaknesses. François Yvon, a research director specialised in automatic translation with the French National Centre for Scientific Research CNRS, pointed to the French expression Il pleut des cordes for heavy rain, which literally refers to ropes. Yvon said a well-trained system renders it with the English cats-and-dogs equivalent but may struggle to find a matching expression in Indonesian.
John Zimmerman, founder of Swiss non-profit Open Knowledge Association, which translates Wikipedia articles, said automated output now suits many routine purposes. Zimmerman said multiplying spending a hundredfold to eliminate the final small share of mistakes makes little sense for internal papers or product listings, while accuracy matters more for an encyclopedia built on trust because false information can persist for years and be repeated by journalists and AI systems.
Zimmerman said OKA has published more than 10,000 items, mainly in Spanish, Portuguese and English, making it the largest provider of translated entries to the collaborative encyclopedia. He said the organisation could not have reached that volume without AI, as conventional translation at that scale would have cost millions, and described its workflow as AI drafting followed by sentence-by-sentence human review.
Zimmerman said reliance on people is declining. He said translators rewrote large portions of drafts from tools such as DeepL two or three years ago, whereas prose today often requires no correction. He said the industry question has therefore shifted from whether AI translation performs well to whether it performs well enough to operate without people, a prospect he described as both plausible and tempting.
Legal specialists warn that remaining mistakes carry high stakes. Ariella Baltié, a translator and legal expert specialised in English-to-French court documents, said a client recently asked her to review AI-translated papers that contained two major mistakes near the start. She said Le demandeur had been rendered as the requester rather than the plaintiff, which she described as highly inappropriate for legal use, and Madame X, épouse Y had been rendered as Ms X marries Y, turning a noun describing marital status into a verb and losing the sense.
The report noted that faulty machine translation of asylum applications in the United States has delayed release of detainees, illustrating wider legal risks.
Medical translators report similar dangers. Carol Bereuter, a specialist in medical translation, said she encountered an automated text that converted the Greek letter μ to milli for one thousandth rather than micro for one millionth. Bereuter said administering milligrams instead of micrograms would deliver a thousandfold overdose.
Bereuter said improved fluency makes detection harder. She said older Google Translate output often sounded awkward or ungrammatical and stood out, while current output reads smoothly and naturally, allowing mistakes to pass unnoticed.
Yvon warned that reputational and practical risks rise when a person signs a translation or bases decisions about treatment, contracts or large purchases on it. He said words with several meanings, usually resolved easily by people through context, can mislead machines, citing translate into Polish being interpreted as concerning shoe polish. He added that without recognising discussion of a football match, a system can mishandle taking a corner.
OKA examined performance in a Wikimedia-funded study last year covering 119 articles, 10 language pairs and more than 1,000 hours of translator labour, according to Zimmerman. He said 27% of machine-drafted wording required amendment before release, three quarters of those amendments addressed AI-introduced mistakes, and 5 to 6% of mistakes altered the source meaning.
Yvon said the basic probabilistic method, which selects one word over another based on likelihood, makes full reliability difficult to envisage because likely wording does not equal correct wording, particularly in subtle and unpredictable language. He predicted systems that assess and revise their own work through several AI components acting as drafter, critic and reviser. He said the decisive advance will arrive when systems communicate uncertainty as well as translations.
COMMENTS