The Brief
Work & Economy 5 min read

One Trillion Words a Month: Who Still Needs a Human Translator?

NAVION

Share

Google Translate now processes one trillion words every month. That number, disclosed by Google in April, captures something larger than a product milestone. It marks a threshold: machine translation has moved from a curiosity to infrastructure, handling a volume of language that no human workforce could match. The question that follows is not whether AI can translate. It clearly can. The question is what it still gets wrong, and whether those errors matter enough to keep humans in the loop.

The Gap Between Fluent and Correct

The most important shift in AI translation over the past few years is not speed or scale. It is the disappearance of obvious errors. Earlier systems produced output that was easy to flag: clunky grammar, unnatural phrasing, sentences that sounded like they had been fed through a machine. That made quality control straightforward. Reviewers could spot problems quickly.

Today, the errors are harder to find precisely because the prose sounds natural. Carol Bereuter, a specialist in medical translations, describes the change directly: where older tools produced sentences that would “jump out” due to grammatical problems, current AI output sounds fluid and legitimate. The errors are still there. They are just no longer wearing a sign.

This is not a minor inconvenience in high-stakes fields. Bereuter encountered an automatically translated medical text that rendered the Greek letter mu as “milli” rather than “micro.” The difference is a factor of one thousand. In a dosage context, that is not a translation error. It is a patient safety risk. The sentence read correctly. The meaning was catastrophically wrong.

Legal translation carries similar risks. Ariella Baltié, a translator specialised in court documents, reviewed an AI-translated legal filing and found two significant errors near the opening. The word “demandeur” (plaintiff) became “requester,” a term that carries no legal weight in the target language. A phrase describing a woman as the wife of someone was rendered as a description of her marrying him, confusing a noun for a verb. In asylum proceedings in the United States, errors of this kind have contributed to delays in the release of detainees.

Where AI Works, and Where It Doesn’t

Not every translation carries these stakes. John Zimmerman, founder of OKA (Open Knowledge Association), a Swiss non-profit that translates Wikipedia articles, draws a practical line. For internal documents or product listings, he argues, multiplying costs to eliminate the last few percentage points of error is not rational. AI translation is, in those contexts, the right call.

OKA has published more than 10,000 translated articles, mostly into Spanish, Portuguese, and English, making it the single largest contributor of translated content to Wikipedia. That scale would have been impossible without AI. Translating that volume through traditional methods would have cost millions. The current process uses AI to produce a draft, which a human translator then reviews sentence by sentence.

But even that model is under pressure. Zimmerman notes that two or three years ago, translators were rewriting a substantial portion of every AI draft. Today, the prose often requires no fixing at all. The temptation to remove the human reviewer entirely has become, in his words, “a serious option” across the industry.

OKA ran a study funded by Wikimedia examining 119 articles across 10 language pairs, drawing on more than 1,000 hours of translator work. The findings are instructive: 27% of AI-generated text was corrected before publication. Three quarters of those corrections addressed errors the AI itself had introduced. Between 5 and 6% of errors changed the meaning of the original text.

That last figure is the one that matters. A translation can be 94% accurate and still mislead a reader on the point that counts most.

What the Reliability Ceiling Reveals

François Yvon, a research director specialised in automatic translation at the French National Centre for Scientific Research (CNRS), identifies a structural reason why current systems are unlikely to reach full reliability. AI translation is fundamentally probabilistic: the system selects the most likely word given the context it has processed. What is statistically probable is not the same as what is correct. Language is too subtle, too context-dependent, and too full of exceptions for probability alone to be sufficient.

Yvon points to a simple example: the instruction “translate into Polish” could, without sufficient context, lead an AI to interpret the word “Polish” as referring to shoe cream. A human reader would never make that mistake. The machine, operating on probability, might.

His forecast for the next generation of systems involves layered AI tools: one to produce the translation, a second to critique it, a third to generate a revised version. That architecture could reduce errors. But the real breakthrough, in his view, will come when systems can express uncertainty, when they can flag the sentences they are not confident about rather than producing fluent-sounding text regardless of whether they have understood it.

That capability does not yet exist at scale. Until it does, the human role in translation is not about doing what AI cannot. It is about catching what AI does not know it got wrong.

In Short

AI translation is fast, cheap, and increasingly fluent. For low-stakes, high-volume content, it is already the rational default. But fluency is not accuracy, and the errors that remain are harder to spot than ever. In legal, medical, and any context where meaning carries consequences, human review is not a redundancy. It is the mechanism that catches the mistakes the machine makes confidently. The trillion-word figure tells you how far the technology has come. The 5 to 6% of meaning-changing errors tells you why the story is not over.

Based on reporting from BBC News.

Written by

NAVION