Open any CAT project and you get the same picture: a grid. Source on the left, target on the right, rows numbered like a spreadsheet. Row 47 sits there waiting for a translation, and it says "Apply." Apply what? A discount? A setting? Nail polish? You scroll up. You scroll down. The tool doesn't care — it just wants row 47 filled and confirmed so the progress bar moves.
We've been working like this for thirty years and calling it efficiency. It isn't. It's the single biggest reason translation quality quietly rots, and everyone in the chain has agreed not to say so out loud.
The grid was never neutral
The segment isn't a unit of meaning. It's a unit of accounting. Somebody decided that a sentence-shaped string was the thing to count, price, match against a memory, and score, and the whole industry organised itself around that decision. Translation memory made it worse, not better. A TM stores strings, so it rewards you for treating text as strings. You get a 95% match on row 47 from a project two years ago, you accept it, and you never notice that the paragraph it now lives in has a completely different subject.
For Turkmen this is sharper than most people realise. Turkmen leans hard on suffixes and on context to mark who's doing what to whom. A verb ending shifts depending on things that live three sentences away — formality, who's being addressed, whether this is an instruction or a description. Hand me an isolated segment and I'm guessing. Half the time I guess right because I've read the file top to bottom before I start. The tool doesn't reward that reading. It rewards me confirming rows.
And here's the part that stings: the QA step inherits the same blindness. Automated checks run per segment. Number mismatch in row 47, tag missing in row 88, terminology deviation in row 12. All string-level. None of it asks whether row 47 makes sense given row 46. The report comes back clean and the translation is still wrong, just wrong in a way no checker was built to see.
The industry finally admits it
Now the vendors are selling the fix, and I have to give them this — they've named the real disease. Translated is rolling out what it calls a context-centric workflow, pitched openly as a replacement for the TM-centric paradigm, working on "full content context" instead of matched strings. The trade press has quietly restyled CAT tools as "AI collaboration platforms" rather than translation memory engines. The framing everywhere is the same: stop feeding the machine isolated rows, feed it the whole document.
There's a related tell in the QA metrics story. COMET and the other automated scores are starting to break on LLM output — Translated published a whole piece on why they fall short. Part of the reason is that those metrics grew up in the segment era. They compare a produced string to a reference string. An LLM that reads context and rephrases a sentence to fit the paragraph gets penalised for not matching the reference word-for-word, even when a human reads it and says yes, that's better. The metric is measuring string similarity. The translator is measuring meaning. When those two disagree, the metric is the one that's wrong.
So the diagnosis is correct. The segment was the weak point. Good.
Don't buy the whole cure
Here's where I get suspicious. "Full content context" sounds like the machine now understands the document. It doesn't. It has more of the document in its window. That's a real improvement and I'll take it — feeding the model the surrounding paragraphs genuinely helps disambiguate that Turkmen verb ending. ModernMT's adaptive engine sitting as the default in Matecat, free, learning from your corrections as you go, is a better starting point than a cold segment match ever was. I'm not romantic about the old way.
But the context that actually decides a Turkmen translation is usually not in the file. It's the screenshot of where the button sits. It's knowing this "Apply" is a submit button, not a verb in a sentence. It's the client telling me the audience is oilfield technicians in Türkmenbaşy, not marketers in Ashgabat. No amount of document-window context reaches that. The connectors going straight into ServiceNow and Figma help a little, because at least the string arrives with a UI around it. But most of my work still lands as a bare table with no picture attached.
And the honest read on where this goes: over 70% of freelancers already run MT, and a lot of them report rate erosion and burnout in the same breath. The context-centric turn is real progress on quality. It is also, conveniently, the story that lets "AI draft plus human review" become the default tier and human-from-scratch become the expensive exception. The workflow improvement and the pay cut are riding in on the same horse.
So when a PM sends me a context-centric project and tells me the model already read the whole file, my answer is the same as it was in the grid days. Send me the file, the screenshots, and who's reading it. Then I'll read row 47 in light of everything else — which is the thing the tool has spent thirty years training me not to do.