Why Literary Translation Needs More Than Neural Networks

You've heard it a hundred times: "Machines are taking over translation." Neural networks, zero-shot translation, interlinguas—the hype is real. But here's the contrarian truth: for literary translation, the machine is not your replacement. It's your raw material. The best literary translators in 2025 are not the ones who reject MT; they're the ones who use it as a rough draft and then rewrite it with intent.

This is a position piece, so let me be blunt: if you're translating literature, stop treating MT output as a starting point. Treat it as a source of inspiration, a scaffold to be torn down. The machine's job is to get you 80% there. Your job is the other 80%.

The Machine Is a Brilliant Idiot

Neural Machine Translation (NMT) is genuinely impressive. It analyzes whole sentences, not just words, and it's the engine behind Google Translate and DeepL (Machine translation, Wikipedia). In 2016, Google's GNMT cut translation errors by 55–85% on major language pairs compared to phrase-based systems (GNMT at production scale, Google Research). But here's what those numbers don't tell you: they're measured on news and Wikipedia text, not on metaphor, rhythm, or cultural allusion. A machine can nail the denotation and butcher the connotation.

For literary work, that's fatal. The machine doesn't know that "the sky wept" isn't just rain. It doesn't feel the weight of a subjunctive in a language that doesn't have one. It's a brilliant idiot—fluent but clueless.

Post-Editing Is Not Translation

In the industry, the standard response is post-editing. ISO 18587:2017 defines the requirements for full human post-editing of MT output (ISO 18587:2017, ISO). And sure, post-editing works for user manuals or patent claims. But literary translation isn't about conveying information; it's about crafting experience. If you're just fixing the machine's errors, you're not translating—you're proofreading.

Consider the translation of a novel's opening line. The machine might give you a grammatically correct, semantically accurate sentence. But it won't capture the voice, the rhythm, the deliberate ambiguity. That's not something you can post-edit into existence. You have to start over, guided by the machine's output as a map of the terrain, not as a destination.

What the Machine Gets Right (and Wrong)

Now, let's address the counter-argument: "But machines are so good now! DeepL is chosen three times more often than Google by professional translators in blind tests" (DeepL Translator, Wikipedia). True. And NLLB-200, Meta's model, improves quality by 44% on average for low-resource languages (No Language Left Behind, Nature). These are real gains.

But watch what happens with literary text. The machine flattens stylistic quirks. It normalizes unusual syntax. It misses the pun, the cultural reference, the intentional ambiguity. The evaluation metrics that show machine progress—BLEU, chrF, COMET—are all designed for information transfer, not literary effect. BLEU compares n-gram matches to a reference, and even its creators at IBM admit it correlates poorly with human judgment (IBM Research). COMET and BLEURT are better, but they still can't judge whether a translation makes a reader feel something.

So yes, the machine is getting better. But "better" in this context means "closer to literal," and literary translation is about everything but literal.

Your Competitive Advantage Is Judgment

Here's the practical advice: in a world of MT, your value isn't in producing text—it's in making choices. The machine offers you a draft, but you decide which words carry the weight, where to break the sentence, how to render the untranslatable. That's not a skill you can automate. It's judgment, taste, and cultural fluency.

And that judgment is scarce. The language services industry is big—CSA Research pegged it at $49.68 billion in 2023 (CSA Research)—but it's shrinking. MT is eating the commodity work, the routine translations. What's left is the high-stakes, high-creativity work: literature, marketing, legal nuance. That's where you should aim.

Here's a concrete scenario: you're translating a contemporary Spanish novel into English. The machine gives you a serviceable draft. But the protagonist has a habit of using a particular colloquialism that the machine renders as "you know?" in every instance. You know that in English, repetition like that is a stylistic choice—but it's the wrong choice here because the character's speech should feel natural, not robotic. So you vary it: "right?", "see?", "okay?"—each with a different nuance. That's the kind of micro-decision that no metric can capture, and it's exactly what a literary translator is paid for.

The Bottom Line

Use MT as a research assistant, not a ghostwriter. Run the text through the machine to get a quick sense of the source, then close the tab and write your own translation. The machine's output is a map, not the territory. Your job is to walk the terrain and describe it in your own words. That's the only way to produce literature, not just translated text.

Bottom line: Treat MT as a brainstorming partner, then rewrite from scratch with your own voice.

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/
• ISO 18587:2017 (ISO) - https://www.iso.org/standard/62970.html
• DeepL Translator (Wikipedia) - https://en.wikipedia.org/wiki/DeepL_Translator
• No Language Left Behind (Nature) - https://www.nature.com/articles/s41586-024-07335-x
• IBM Research - https://research.ibm.com/blog/bleu-nlp-benchmark-anniversary
• CSA Research - https://csa-research.com/l/media/Language-Services-and-Technology-Industry-Faces-Revenue-Decline-but-Remains-Poised-for-Transformation