AI translation cuts costs but leaves a costly checking problem

AI translation can substantially reduce the cost of multilingual work, but the evidence does not establish when businesses can dispense with human reviewers. Reporting by BBC Technology shows why: systems increasingly produce convincing prose while still making errors that alter legal meaning or medical quantities. The commercial calculation depends less on whether a translation reads well than on what happens if it is wrong.

What happened

The scale of machine translation is already considerable. In April, Google said its translation service was handling one trillion words a month. That excludes competing services, general-purpose AI assistants and bespoke corporate systems. BBC Technology also reports estimated direct machine-translation sales of $1.4bn (£1bn) last year, excluding spending through general AI subscriptions. Neither measure, however, establishes how much output is accurate enough to publish unchecked.

One organisation testing that boundary is the Open Knowledge Association, or OKA, a Swiss non-profit that translates Wikipedia articles. Its founder, Jonathan Zimmerman, says it has published more than 10,000 translated pieces, mainly in Spanish, Portuguese and English, making it Wikipedia’s largest contributor of translated articles. AI produces the drafts; human translators check every sentence.

Zimmerman says translating that volume conventionally would have cost millions. This is an important distinction in the adoption debate: automation is not merely replacing an existing workflow more cheaply. At OKA, it has enabled a volume of work the organisation says would otherwise have been beyond its means.

Yet its own research shows that drafting and publication remain different jobs. A Wikimedia-funded study conducted by OKA last year covered 119 articles, ten language pairs and more than 1,000 hours of translators’ work. According to Zimmerman, 27% of AI-generated text was corrected before publication. Three-quarters of the corrections addressed errors introduced by AI, and 5–6% of errors changed the original meaning.

Those figures are evidence from one workflow, not a universal error rate for translation software. The account does not provide a vendor-by-vendor comparison or establish how the results would transfer to contracts, product catalogues or medical records. It does show that substantial editing can remain necessary even when the initial draft appears accomplished.

The background

Translation requires choosing meanings, not merely exchanging words. François Yvon, a research director specialising in automatic translation at France’s CNRS, uses a French rain idiom to illustrate the problem. A well-trained system may recognise its English equivalent without handling the same expression successfully in Indonesian. Competence in one language pair is not proof of competence in another.

Context creates similar difficulties within apparently ordinary sentences. Yvon notes that a system may confuse the Polish language with shoe polish, or misunderstand a reference to taking a corner if it has not recognised a football discussion. A grammatically plausible result can therefore conceal a failure to identify the subject at all.

Specialist terminology raises the stakes. Ariella Baltié, a translator and legal expert specialising in court documents, told the BBC that a prospective client brought her AI-translated papers containing two major errors near the beginning. One rendered the legal term for a plaintiff as a general requester; another turned a description identifying a woman as someone’s wife into a statement that she marries him.

These were not simply inelegant choices. One failed to preserve a legal role; the other changed the relationship expressed by the sentence. BBC Technology reports that errors in machine-translated US asylum applications have delayed detainees’ release. In such settings, the cost of translation is only a small part of the cost of a mistranslation.

Medical translation offers an especially stark example. Carol Bereuter, a specialist in the field, encountered an automatically translated text that interpreted the Greek letter μ as milli, meaning one-thousandth, rather than micro, meaning one-millionth. Confusing those units in a dosage would create a thousandfold difference. The source describes an error in a text, not a patient receiving that dose.

Bereuter also identifies a less obvious consequence of technical progress: mistakes are becoming harder to notice. Earlier output often advertised its weaknesses through awkward grammar. Today’s fluent wording removes that warning. A reader who can judge whether a sentence sounds natural may still be unable to judge whether it faithfully preserves the original.

What people are saying

Zimmerman’s position is pragmatic rather than protective of human translation for its own sake. He argues that unreviewed AI output is already appropriate for many uses, including internal documents and product listings, where the expense of removing the remaining errors may be disproportionate. That is his assessment of the trade-off, not evidence that every document in those categories carries little risk.

For Wikipedia, he reaches a different conclusion. An erroneous fact can remain available for years and be repeated by journalists or AI systems. The relevant cost is therefore not confined to correcting the original page: republication can extend the error’s life and reach. OKA’s sentence-by-sentence review reflects that exposure.

Zimmerman says the quality of drafts has improved markedly. Two or three years ago, translators using tools such as DeepL still rewrote substantial portions; now the prose often needs no repair. He sees removing the reviewer as a serious option, and a growing temptation across the industry. His account distinguishes stylistic improvement from demonstrated factual reliability.

Yvon places the dividing line at responsibility and consequential decisions. A translation carrying someone’s signature creates reputational exposure; one used to decide on a medical procedure, contract or major purchase creates more. This reframes the buying decision. The cheapest way to obtain readable foreign-language text need not be the cheapest way to complete the underlying transaction safely.

He is also sceptical that current systems can become completely reliable. Their probabilistic operation makes the most likely wording different from wording that is certainly correct. This is an expert assessment of the technology’s limitations, rather than a measured ceiling on future performance.

Murray Wilkinson, founder of Found Revenue, an AI business services provider with three UK locations, says firms that see value usually pick one process, measure it before and after, and keep a named person accountable for the outcome.

What happens next

Yvon expects systems to become better at evaluating and revising their own work. He describes a possible arrangement in which one AI tool translates, another criticises the result and a third prepares a revised version. Such a workflow could move some checking into software, but the BBC report supplies no performance results showing how much risk it would remove.

A more consequential advance, in his view, would be a system that can reliably signal uncertainty. That would change the economics of review: instead of examining every sentence equally, specialists could concentrate on passages the system cannot confidently resolve. The commercial value would depend on whether those warnings catch consequential mistakes, not simply on whether a warning feature exists.

For now, the missing evidence is as important as the improvement in prose. The reported study does not establish a safe threshold for dispensing with reviewers, and the specialists’ examples do not measure how frequently comparable failures occur. There is no supported timetable for eliminating human translation work. What the evidence supports is a narrower shift: machines can do much of the drafting, while people remain responsible for deciding which errors the business cannot afford.