A PO landed last week with the task type marked "linguistic review" and a note that the file was "already translated, just needs a check." Eight cents a word. I opened it. It was machine Turkmen. Not a human draft that needed polish — raw output with the tell-tale limp: correct-looking words in the wrong case, a genitive chain that collapsed halfway, terminology that shifted three times in two paragraphs because the model has no memory across segments. Nobody had touched it. The "already translated" was a machine talking to itself.
This is the part of vendor life nobody puts in the onboarding pack. The task says one thing. The file says another. And the gap between them is money out of my pocket.
The quiet swap
Machine post-editing isn't a niche anymore. A GTS survey this spring put MTPE adoption among freelancers at nearly 88% — 48% doing it frequently, another 40% occasionally. Nimdzi's numbers say 60% of translators now use MT and 80% of agencies do. So the machine is in the room. Fine. I've made my peace with that; I post-edit when the source is worth post-editing and I say no when it isn't.
What I haven't made peace with is the relabeling. When a PM calls a post-edit a "review," two things happen. The rate drops to review money, and the effort jumps to translation-from-scratch effort — because fixing fluent-and-wrong Turkmen is slower than typing it fresh. I've written before about how cleaning up machine Turkmen can cost more than starting over. Undisclosed MTPE is that trap with the label peeled off.
And the frustrating thing is, the industry already decided this is wrong. Acolad's 2025 translator survey found content is often already managed by AI even when it isn't disclosed, and that knowing which tools were used is how you assess whether the output can be trusted. Slator's read on the year was that transparency became the norm. The norm. So why am I still reverse-engineering whether a file came out of a CAT tool or a black box on a Tuesday afternoon?
Why the anchor makes this worse
There's a reason the swap keeps happening, and it isn't malice. It's the price anchor. Clients now know machine translation is nearly free, so any human hour feels expensive by comparison. Acolad's survey put translator rates down 30–50% in some pairs since 2023; Slator says 57% of language companies reported revenue declines. The PM sitting between me and a squeezed end client is under pressure to make human work look like a light touch-up. "Review" is cheaper to sell than "translation." So the file gets rebranded on the way to me.
I understand the pressure. I don't accept the accounting. Because for Turkmen specifically, the machine's output isn't a 90%-there draft I nudge. It's confidently broken. The morphology is where these models fall apart — six grammatical cases, vowel harmony, possessive suffixes stacking on case endings — and the machine produces something that reads smooth to a monolingual PM and is quietly wrong to anyone who speaks the language. That's the worst kind of wrong to fix, because you can't skim it. You have to re-derive every clause.
What I ask now, before I quote
So I stopped guessing and started asking. Three questions, sent back before I accept anything labeled review, edit, or check:
Was this machine-translated, and with what? DeepL doesn't do Turkmen. Google and the big multilingual models do, badly. If the answer is "we ran it through X," I know exactly what kind of damage I'm walking into and I price the post-edit, not a review. If the answer is "a translator," great — then I hold them to that when the file says otherwise.
Can I see a sample before I commit to the rate? Two hundred words tells me everything. I'll spot machine output in the first segment. If a PM won't send a sample, that's an answer too.
Is the rate a review rate or an edit rate? Naming it forces the honesty. Nobody has to lie; they just have to say the number out loud next to the real task.
Half the time the PM genuinely didn't know. The file came from the end client already "done," got passed down the chain, and the label stuck by accident. That's the good outcome — we fix the PO and move on. The other half, the answer gets vague, and vague is data. I've turned down two jobs this quarter where "review" meant "salvage," and both times the agency came back with the actual rate once they realized I wasn't going to eat the difference.
Here's what I keep telling PMs I like: I'm not against the machine touching a file. I'm against the machine touching a file in secret and me paying for the surprise. Disclosure isn't a favor to the translator. It's the input that lets me quote you a number that's still true when I open the document. Hide it, and every quote I give you is a guess — and I'll guess high, because I've been burned.
Tell me what the machine already did. Then we can talk about what I'm worth on top of it.