Can AI replace humans in translation? Experts warn of high-stakes errors

As machine translation volume climbs to one trillion words per month via Google Translate—and with a $1.4bn (£1bn) market for direct sales of translation tools—industry experts are debating whether AI can reliably replace human translators. According to reporting by the BBC, while tools like DeepL, Microsoft, Claude, ChatGPT, and Gemini are increasingly capable, their reliance on probability rather than context creates significant risks in high-stakes environments.

In medicine, the consequences of machine error can be life-threatening. Medical translator Carol Bereuter cites a case where an AI mistranslated the Greek letter μ (micro) as “milli,” a error that could lead to a thousand-fold dosage overdose. Similarly, in the legal sector, translator Ariella Baltié reports that AI consistently fails to identify context, such as misidentifying a plaintiff as a “requester” or mistaking nouns for verbs. These errors have already caused tangible real-world consequences, including delays in the processing of US asylum applications.

Even as AI prose becomes more natural, experts warn that this fluency makes errors harder to detect. François Yvon, a research director at the French National Centre for Scientific Research (CNRS), explains that current systems lack the ability to show doubt when they are uncertain, a limitation he considers a major barrier to reliability. A study by the Swiss non-profit OKA, which analyzed 119 articles, found that 27% of AI-generated text still required human correction, with 5% to 6% of errors altering the fundamental meaning of the original text.

John Zimmerman, founder of OKA, argues that while raw AI output is efficient for low-stakes tasks like internal documentation or product listings, it is not yet sufficient for critical work. For projects like Wikipedia, where OKA has contributed over 10,000 translated articles, Zimmerman notes that the current industry temptation to skip human verification ignores the reality that machine output remains imperfect. As Yvon summarizes, while future tech may involve systems that critique their own translations, the fundamental hurdle of achieving 100% reliability in the face of linguistic nuance persists.