Is machine translation + a human edit ever good enough?

@midnight_edits OP

1mo ago

DeepL has gotten startlingly good, and I keep wondering if I could run my novel through it and pay an editor to clean up the result for a fraction of full translation cost. Is that a real workflow anyone uses, or am I about to embarrass myself in three languages at once? Honest experiences welcome.

11 comments

@comma_kira1mo ago

I do post-editing of machine translation for a living, mostly technical and marketing copy, and I'll be blunt: for literary fiction this almost never saves you money or dignity. The machine flattens voice. It picks the statistically safe word every time, so a character with a sharp, weird way of speaking comes out sounding like a press release. Fixing that isn't editing, it's retranslating with extra steps.

@comma_kira1mo ago

The cruel part is the editor often has to do MORE work cleaning a confident-but-wrong machine draft than translating fresh, because they first have to detect the subtle errors the machine hid under fluent grammar. Fluent and correct are not the same thing, and MT is very good at fluent.

@midnight_editsOP1mo ago

That distinction between fluent and correct is exactly the thing I underestimated. So the smooth output is part of the trap.

@comma_kira1mo ago

Precisely. A human translator who's unsure leaves a comment or asks you. The machine never flags doubt, it just commits to a plausible wrong reading and moves on. You won't catch it because it reads beautifully.

@nocturne_writes1mo ago

Where it genuinely works: non-fiction where information matters more than voice. I used MT plus a careful human edit for a technical manual and a set of help-center articles, and it was fine and cheap. The moment register, rhythm, humor, or subtext carries the meaning, it falls apart. Novels live entirely in that second category.

@midnight_editsOP1mo ago

Okay, consensus landed where I feared but needed to hear: fine for my own comprehension and for dry non-fiction, actively harmful for the novel. I'll budget for a real human and stop trying to outsmart the price.

@the_proofreader1mo ago

Idioms are where it falls on its face most visibly. MT translates the words of an idiom rather than its function, so you get literal nonsense that a reader in the target language has to mentally reverse-engineer. A human knows to swap in the equivalent expression, or to rebuild the sentence so the idiom isn't needed at all.

@comma_kira1mo ago

And register drift is the silent killer. MT can't reliably hold a consistent level of formality across a whole book, so a character slides between street slang and courtroom diction page to page. A reader feels that wrongness even if they can't name it.

@pulp_priya1mo ago

I'll defend one narrow use. I run my own chapters through DeepL to get a rough gist of how a foreign-language review or a translated sample reads, just for my own private understanding. As a comprehension tool for me it's great. As something a reader pays for, never.

@first_draft_dani1mo ago

There's also a rights and contract wrinkle people forget. Some publishers and prize committees explicitly will not consider machine-assisted translations, and a few translators won't put their name on a post-edited MT job because their reputation is the product. So a cheap MT route can quietly close doors you wanted open.

@slowburn_sam1mo ago

One more cost reality: even when MT post-editing is appropriate, good editors charge more per word for it than people expect, because untangling someone else's confident mistakes is slower than it looks. The fantasy is that you pay ten percent of a translation and get ninety percent of the quality. The real curve is nowhere near that kind.