Machine translation has evolved from a niche experimental tool to a global workhorse, with Google alone reporting that users now run one trillion words through its Google Translate service every single month. That massive volume does not even include translations processed by competing platforms, from DeepL and Microsoft’s translation tools to generative AI systems like ChatGPT, Claude and Google’s own Gemini, plus the custom-built corporate translation solutions that power countless international businesses today. By 2023, the direct market for standalone machine translation, separate from general AI subscription services, reached an estimated value of $1.4 billion (£1 billion), reflecting just how central automated translation has become to global communication.
Idiomatic language has long been one of the biggest hurdles for AI translation. Take the common French phrase for heavy rain: “Il pleut des cordes,” which translates literally to “it’s raining ropes.” While a well-trained AI system will recognize this as an idiom and substitute the equivalent English phrase “it’s raining cats and dogs,” experts question whether it can consistently find the right equivalent in less widely spoken languages, such as Indonesian. What seems like a trivial example for weather commentary points to a broader challenge: as AI translation moves into high-stakes sectors from legal work to healthcare, small mistakes can carry severe consequences.
For many low-stakes use cases, industry leaders agree that raw AI translation is already the most practical choice. Jonathan Zimmermann, founder of Swiss non-profit Open Knowledge Association (OKA) – the largest single contributor of translated articles to Wikipedia – explains that for internal company documents or public product listings, spending 100 times more to catch a tiny share of remaining errors is economically irrational. Even for OKA’s Wikipedia work, AI has been transformative: translating 10,000 articles via traditional human-only methods would cost millions of dollars, but OKA’s hybrid model, where AI produces a first draft and human translators review every line, makes the work accessible.
What has shifted most dramatically in just the last two to three years is the quality of AI output. Zimmermann notes that when tools like DeepL first became widespread, human translators still had to rewrite large portions of every AI draft. Today, many AI-generated translations require no editing at all. This shift has led to a growing industry-wide temptation to cut human oversight out of the process entirely, raising the core question: is AI translation now good enough to replace human translators entirely?
Specialized translators working in high-stakes fields say the answer is still no. Ariella Baltié, a legal translator who converts court documents from English to French, recently reviewed a set of AI-translated legal papers that contained two critical errors in the opening lines alone. The AI incorrectly rendered “Le demandeur” – the legal term for plaintiff – as “the requester,” a mistranslation that is entirely inappropriate for formal legal documents. It also mistook the noun phrase “Madame X, épouse Y” (meaning “Ms X, wife of Y”) for a verb phrase, outputting “Ms X marries Y” instead. Mistranslations of this type in legal work are not merely embarrassing: in the United States, errors in machine-translated asylum applications have already led to prolonged delays in detainee release.
In the medical field, translation errors can be life-threatening. Carol Bereuter, a specialist in medical translation, recalls an AI-translated document that misrendered the Greek letter μ – the standard symbol for “micro,” meaning one-millionth of a unit – as “milli,” which equals one-thousandth of a unit. That error changes a prescribed microgram of medication to a milligram, meaning a patient would receive 1,000 times the intended dose. Bereuter adds that modern AI’s improved fluency has actually made errors harder to catch: early machine translation often produced obviously clunky, grammatically incorrect text that stood out to human reviewers, but today’s AI output sounds natural and coherent, hiding subtle but dangerous mistakes.
François Yvon, a research director specializing in automatic translation at France’s National Centre for Scientific Research (CNRS), warns that context-dependent ambiguity continues to trip up even the most advanced AI systems. A human translator immediately understands that the phrase “translate into Polish” refers to converting text to the Polish language, but Yvon notes that AI can easily misinterpret the phrase to refer to polishing shoes, translating the sentence to be about shoe polish. Similarly, without understanding that a text refers to a football match, AI will often misinterpret “taking a corner” as a physical action rather than a soccer play.
A 2023 study run by OKA and funded by Wikimedia underscores the scale of remaining errors. Researchers analyzed 119 articles across 10 language pairs, tracking more than 1,000 hours of human translator work, and found that 27% of AI-generated text still required corrections before publication. Three-quarters of those corrections fixed errors introduced by the AI itself, and between 5% and 6% of those AI errors fundamentally altered the meaning of the original text.
Yvon argues that fully eliminating these errors is a fundamental challenge given how current AI translation systems work. Large language models select words based on statistical probability, choosing the word that is most likely to fit rather than the word that is definitively correct. That discrepancy matters in language, a domain defined by subtlety and nuance that can easily throw AI off course.
Looking forward, Yvon says next-generation systems are likely to improve performance through hybrid AI architectures, where one model generates the translation, a second critiques it for errors, and a third produces a revised final version. But the true breakthrough, he argues, will not arrive until AI systems can communicate when they are uncertain about a translation. Until that day, even as AI handles more and more translation volume globally, human oversight remains critical for any high-stakes use case.