The best sentence in the BBC's recent report on machine translation has nothing to do with idioms. It is about fluency. Carol Bereuter, a medical translator, says the older tools gave themselves away: the grammar was clunky, so the errors jumped out. Today the output "sounds perfectly legitimate. It sounds fluid. It sounds natural. So the errors are a lot harder to spot."

That is the whole story in miniature. Better translation does not simply shrink the error rate. It disguises the errors that remain.

What the report says

The reporting covers France, Switzerland and the United States. It does not cover India, so I will keep its findings and my Indian context apart.

The scale is enormous. In April, Google said people were translating one trillion words a month with Google Translate, and that figure excludes DeepL, Microsoft, Claude, ChatGPT, Gemini and bespoke corporate systems. Direct sales of machine translation, outside general AI subscriptions, were estimated at $1.4bn last year.

The humans have not vanished, though. Jonathan Zimmermann of OKA, a Swiss non-profit that translates Wikipedia articles, makes a sober economic case. For internal documents or product listings, paying 100 times more to catch the last few percent of errors is not rational. An encyclopedia is different: a wrong fact can sit there for years and get cited.

OKA's own workflow is the telling case. The AI drafts; a human checks every sentence. Translators once rewrote a good chunk of each draft, Zimmermann says, and now the prose often needs no fixing. Yet a study OKA ran last year, funded by Wikimedia, covering 119 articles, 10 language pairs and more than 1,000 hours of translator work, found that 27% of the AI text still got corrected before publication. Three quarters of those corrections fixed errors the AI introduced, and 5 to 6% of errors changed the meaning of the original.

Why the errors persist

François Yvon of the CNRS names the root cause. These systems choose each word by probability, and probably correct is not the same as correct. He finds it hard to imagine that approach ever reaching 100% reliability.

The examples are small and vivid. Ariella Baltié, a legal translator, saw "le demandeur" (the plaintiff) rendered as "the requester", and "Madame X, épouse Y" turned into "Ms X marries Y", a noun mistaken for a verb. Bereuter found the Greek letter μ read as milli instead of micro: a thousandfold dose difference. And Yvon points to context failures, such as a machine deciding "translate into Polish" concerns shoe cream, that a human resolves instantly.

Now, India

From here on, the inference is mine, not the BBC's. The Constitution lists 22 scheduled languages, and many people deal with courts, hospitals and government offices in a language other than the one the paperwork was written in. I have no sourced figures on the size of India's translation market or its workforce, so I won't invent any.

What I can do is run the report's logic. Three parts of it travel well.

First, the stakes sort the work. A translated product listing and a translated bail order are different animals. The report's rule of thumb (raw AI where an error is cheap, human review where it is not) maps neatly onto Indian life: a restaurant menu on one side, a medical consent form or a land record on the other.

Second, the idiom problem Yvon raises for Indonesian is a problem of pairs. Will a system find the right equivalent in a genuinely different language? Many Indian language pairs look nothing like English to French, and legal and bureaucratic registers carry terms of art with no tidy counterpart. I can't say how any particular tool handles them, and the report doesn't test it. That gap is precisely why a reviewer who knows both the language and the domain matters.

Third, the fluency trap works the same in any language. If a reader cannot easily check a smooth Hindi, Tamil or Bengali output, polish will pass for accuracy. The most exposed person is the one signing a document they can only partly read. That is where a polished error does its worst.

Where judgement earns its fee

The report suggests the human role is moving from writing to verifying, and that verifying is the scarcer skill. I would expect that to hold in at least three places:

• Legal documents, where a single term changes rights. The report notes that in the US, errors in machine-translated asylum applications have delayed the release of detainees.
• Medical text, where units, dosages and a Greek letter carry the risk.
• Anything that bears your signature, as Yvon puts it, where reputation and decisions rest on the translation.

Look at what the fee actually buys. Not typing speed. Accountability and context: knowing the court, the clinical setting, the football match behind "taking a corner."

The danger is not that machines are bad. It is that they are good enough to make careful checking feel like waste.

The temptation

Zimmermann's warning bears repeating: skipping the human is now "a real temptation across the industry." The machines are not the problem. Their competence is, because it makes checking feel pointless right up until the one error that matters.

I have been cheerfully wrong about technology before, so I hold this loosely. Yvon expects a new generation of systems to improve things, and the report's text cuts off before he says how. But today's evidence favours a split by consequence, not by capability. Let the machine draft the menu. For the plaint, the prescription and the property deed, keep a person who will answer for the words.