Google Translate handles one trillion words a month, while direct sales of machine translation hit an estimated $1.4bn last year. As raw AI output reaches high quality, translators and researchers debate whether machines are finally reliable enough to remove humans from the loop entirely.
Machine translation has moved far beyond the days of clunky phrasing and obvious grammatical mistakes. According to Bbc reporting, Google’s translation service alone handles one trillion words per month, while direct sales of machine translation—excluding general AI subscriptions—reached an estimated $1.4bn, or £1bn, last year. Other tools like DeepL, Microsoft, Claude, ChatCGT, and Gemini join custom corporate systems in processing a massive global volume of text.
For many everyday tasks, industry experts say raw machine output is already the most practical choice. John Zimmerman, founder of the Swiss non-profit Open Knowledge Association (OKA), points out that spending heavily to catch minor errors makes little sense when dealing with internal documents or product listings.
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.
High Stakes in Wikipedia and Legal Translation
When context shifts from product listings to encyclopedias or courts, the tolerance for error drops sharply. OKA serves as the single largest contributor of translated articles to Wikipedia, with more than 10,000 published pieces mostly into Spanish, Portuguese, and English. Zimmerman notes that traditional translation methods would cost millions for that volume, a feat made possible only because AI generates an initial draft that human translators then check sentence by sentence.
The Hidden Dangers of Fluent Prose
Medical translations carry risks that can be a matter of life and death. Carol Bereuter, a specialist in medical translations, highlights how automatic tools handle the Greek letters frequently used in units for dosages and blood tests. One automatically translated text swapped the Greek letter μ for milli (one thousandth) instead of micro (one millionth)—an error that would multiply a prescribed dose by one thousand.
As translation software improves, catching these flaws becomes increasingly difficult. Unlike older tools where clumsy phrasing or grammatical errors stood out instantly, modern systems produce prose that sounds completely natural, fluid, and legitimate. François Yvon, a research director specializing in automatic translation at the French National Centre for Scientific Research CNRS, warns that even straightforward sentences can trip up machines when context is missing. A system asked to translate into Polish
might mistakenly assume the text is entirely about shoe cream, while failing to recognize a football match can lead to mistranslating taking a corner
.
What Wikimedia Studies Reveal About Error Rates
A study run last year by OKA and funded by Wikimedia examined 119 articles across 10 language pairs, involving more than 1,000 hours of translator work. The results showed that 27 percent of the AI-generated text still required correction before publication. Three quarters of those corrections addressed errors introduced by the machine itself, and 5 to 6 percent of those errors actually changed the original meaning of the text.
Achieving complete reliability remains a fundamental challenge due to how these systems operate. Yvon explains that probability dictates how an algorithm selects each word, meaning what is calculated as probably correct is not necessarily what is actually correct in nuanced human language.
As things stand, I think the fundamental way these systems function makes it quite hard to imagine how they could become 100% reliable.
Looking ahead, Yvon expects future technology to improve by using multiple AI tools simultaneously—one to generate the translation, a second to critique it, and a third to produce a revised version. Ultimately, he suggests the true breakthrough will arrive only when translation systems gain the ability to recognize and express their own uncertainty.