Where the Corrections Used to Go

A three-sentence advertisement for a restaurant, Russian into Turkish, December 2025. The Russian reads: we make any sushi cake you like, from 1500 lira a kilo, order 24 hours ahead. Eleven words in the professional Turkish version. Sixteen in the one the machine produced, and four separate things broken in those sixteen.

The offer lost its subject. We make became you can order, and the seller disappeared from his own sentence. The text swelled by forty-five per cent without one added fact. The lead time, a neutral piece of information, came back as an obligation: you are required to place your order 24 hours in advance. And the register moved from a friendly shop notice to something a utility company sends you.

Nothing here is a mistranslation. Every word means what it should. What did not survive the crossing is everything above the word: who is speaking, how close they stand to the reader, whether this is an invitation or a rule. Pragmatics, in the old vocabulary, and it is the part a business actually pays for.

Now the part that makes this more than a complaint about a free tool. In 2009 Google ran Translator Toolkit: a browser-based editor where you corrected machine output, kept a translation memory and built glossaries, all free to anyone. It ran in a browser, which meant it ran on a tablet in a cafe, while the professional alternatives of that decade wanted a paid licence and a desktop machine. The barrier to doing the work properly was an account. It shut down on 4 December 2019, on the reasoning that usage had declined and better tools existed elsewhere. The engine stayed. The place where a professional handed corrections back went away.

To be exact, human correction did not vanish from the company, it moved: into paid cloud products with glossaries and into third-party CAT systems that cost money. What disappeared was the free public loop, the one any working translator could reach. Six years later the engine turns a shop notice into a utility bill, which is roughly the failure it had then.

A theory about the decline, mine and unproven, offered as a working translator's reading rather than a documented cause. Confidentiality is not a preference in this trade, it is a clause. Clients forbid passing source documents to third parties, some forbid machine translation outright, and a translation memory is the client's asset before it is yours. Uploading any of that to a company that sells advertising, with no clear commitment about what happens to the text, breaks the contract you signed. The GDPR came into force in May 2018, eighteen months before the shutdown, and made that calculation sharper for every agency in Europe. Local CAT tools did not survive because they were better. They survived because they were the only ones a professional could legally use. Usage may well have declined exactly as Google said; the question their reasoning leaves alone is why.

What closed was not a feature. It was the low barrier: the one arrangement where correcting a machine cost nothing but attention.

One more turn, and the circle closes. The models that now advertise flawless translation learned it from parallel corpora built by translators: documents, books, subtitles, memories assembled one corrected segment at a time. What they display is aggregated human skill, produced back when correcting the machine was somebody's paid job. The loop was closed in 2019 and the output kept improving, because it was still being fed work done while the loop was open.

Which leaves an arithmetic worth watching. A growing share of text on the web is now machine output, so the corpus behind the next model contains more of its own product, and the errors nobody flags are exactly the ones that survive that cycle. A sentence with the seller removed from it and an invitation turned into an obligation does not look wrong. It looks finished, it gets published, and it becomes training data.

The lesson is not that machine translation is bad. It is that a system which generates without a place to receive corrections will hold its failures for a decade, and they will be the failures nobody reports, because the output looks finished. A reader who knows only one of the two languages sees fluent Turkish and accepts it. The loss is visible only from the position that was removed.