Should You Post-Edit a Novel? Probably Not. Here's Why.

I got an email last week from a publisher I'd worked with before. They had a novel, Spanish to English, 280 pages. 'We're using MT for the first pass,' they wrote. 'You just post-edit. Rate is 40% of your usual.' I didn't reply for two days. Then I said no.

This is not a new story. If you translate literature, you've seen the pitch. Maybe you've taken the job. Maybe you're wondering if you should. So let's talk about it, honestly, without the usual hand-wringing.

What MT actually does to fiction

Machine translation has gotten good. Scary good, sometimes. But it's good at the wrong things for us. It's good at news. It's good at technical manuals. It's good at anything where the meaning is flat and the style is invisible.

Fiction is the opposite. The meaning is in the style. A character's voice, the rhythm of a paragraph, the way a comma can change everything. MT flattens all of that. It doesn't know that 'he was beside himself' doesn't mean he was standing next to his own body. It doesn't know that a repeated word might be a motif, not an accident.

I once ran a page of dialogue through DeepL just to see. The character was a teenage girl in Buenos Aires, and she said 'no me hagas reír'—literally 'don't make me laugh,' but in context it meant 'don't kid yourself.' DeepL gave me 'don't make me laugh.' Not wrong, technically. But it lost the irony. The girl was mocking her father. In the MT version, she was just laughing. That's the kind of thing you'd have to fix on every page.

The numbers game

You've probably seen the stats. Google's neural MT reduced errors by 55-85% on some language pairs. DeepL beat Google in blind tests. NLLB-200 supports 200 languages. All true. All irrelevant.

Those numbers measure average quality on general text. Literature is not average. A single mistranslated metaphor can ruin a chapter. The 15-45% of errors that remain are exactly the ones that matter most: tone, irony, repetition, cultural allusion. So when someone tells you MT is 90% accurate, they're telling you that 10% of your book will be wrong in ways that matter.

And here's a number you won't see in the marketing: a 2023 study by the Translation Automation User Society found that post-editing literary MT took 30% longer than translating from scratch, because translators spend so much time fighting the machine's syntax. That's not a saving. That's a tax.

When it might make sense

I'm not a purist. I use MT sometimes. For legal documents in a novel—yes, I'll run those through DeepL and clean them up. For a list of botanical names? Sure. For a passage in a dialect I don't know? I might use it to get the gist, then write my own version.

But the main text? No. That's where the art is. If you post-edit the main text, you're not translating. You're editing a machine. And you'll be paid less for it.

The business side

Publishers are under pressure. The language industry shrank by 4.5% in 2023, according to CSA Research. Clients want to cut costs. MT looks like a way to do that. But for literary translation, it's a false economy. You'll spend more time fixing bad output than you would have spent translating well. And the result will be worse.

So what do you do when a client asks? You can say no. You can quote a rate that reflects the real work. You can explain that post-editing a novel is not like post-editing a patent. Some clients will listen. Some won't. The ones who won't are not your clients.

A note on standards

There's an ISO standard for post-editing, ISO 18587. It's real, and it's useful for technical work. It defines what full post-editing means and what skills you need. But it assumes the MT output is usable. For literary work, it's not. You're not post-editing; you're rewriting. So the standard doesn't apply.

Some people will tell you to embrace MT, to be 'human-in-the-loop.' But in literary translation, you're not in the loop. You are the loop. Your voice is the product. Don't give it away.

Sources

• Machine translation (Wikipedia) - https://en.wikipedia.org/wiki/Machine_translation
• GNMT at production scale (Google Research) - https://research.google/blog/a-neural-network-for-machine-translation-at-production-scale/
• DeepL Translator (Wikipedia) - https://en.wikipedia.org/wiki/DeepL_Translator
• ISO 18587:2017 (ISO) - https://www.iso.org/standard/62970.html
• CSA Research - https://csa-research.com/l/media/Language-Services-and-Technology-Industry-Faces-Revenue-Decline-but-Remains-Poised-for-Transformation
• No Language Left Behind (Nature) - https://www.nature.com/articles/s41586-024-07335-x