Literary Translation Is Not Dead—It's Just Not What You Think

I. "Won't Machines Replace You?"—The Wrong Question

Everyone asks me if machines will replace literary translators. It's the wrong question. I've been in this business long enough to know that the threat isn't from neural networks—it's from translators who refuse to evolve. The machines are here, they're impressive, but they're not literary. They're tools, and tools don't replace craftsmen; they change the craft.

So let's stop panicking and start adapting. In this article, I'll bust the myths that are holding you back and show you what actually matters in the age of neural machine translation (NMT).

II. "MT Can Translate My Novel Perfectly"—No, It Can't

I hear this from clients who've played with Google Translate and think they can save a fortune. Yes, NMT systems like Google Translate and DeepL are remarkable (Machine translation (Wikipedia)). Google's own GNMT reduced errors by 55–85% on many language pairs back in 2016 (GNMT at production scale (Google Research)). But that's still leaving half the sentences wrong. And a novel isn't a set of sentences—it's voice, tone, subtext, cultural nuance. MT output often requires human post-editing, especially for idiomatic expressions and cultural nuance (Machine translation (Wikipedia)). For literary work, that post-editing is a full rewrite.

I've tested DeepL on a passage from a contemporary Spanish novel. The output was grammatically perfect, but it flattened the narrator's ironic edge into something sterile. No metric catches that. BLEU, the old standard, compares n-gram matches to a human reference and is notoriously unreliable (IBM Research). Newer metrics like COMET and BLEURT correlate better with human judgment, but they still can't assess artistry (IBM Research). So no, MT won't translate your novel. It will give you a rough draft that a skilled translator could use as a base—if they're paid for the effort.

III. "But CAT Tools Are Just for Tech Manuals"—False

Another myth: computer-assisted translation (CAT) tools are only for dry documentation. I used to believe that. Then I realized that translation memory (TM) is a goldmine for literary series. If you're translating a trilogy, consistent terminology is a nightmare. A TM database stores previously translated segments, ensuring consistency and reducing repeated work (Machine translation (Wikipedia)). That's not a crutch—that's a safety net.

I've seen translators waste hours re-translating the same character name or invented term. With a CAT tool, you lock that in. And no, it doesn't force you into a mechanical style. You control the segments; the tool just helps you manage them. ISO 17100:2015 even sets standards for translation services that include using appropriate tools (ISO 17100:2015 (ISO)). So don't dismiss CAT tools as unliterary. Use them to free your mind for the creative decisions.

IV. "Post-Editing Is a Dirty Job"—It's a Valuable Skill

When I tell literary translators that post-editing is a legitimate skill, some scoff. They see it as cleaning up after a machine. But ISO 18587:2017 defines full human post-editing of MT output as a professional service with its own competence requirements (ISO 18587:2017 (ISO)). It's not about fixing commas; it's about transforming raw MT into publishable prose while preserving the original's soul.

I've done post-editing for a publisher who wanted to test the waters with an MT-assisted translation of a detective novel. The raw MT was full of howlers: idioms translated literally, names inconsistent. But after my post-edit, the novel was indistinguishable from a traditional translation—except it took less time. That's the future. Literary translators who refuse to post-edit will find themselves obsolete. Those who embrace it will have a new revenue stream.

V. "Machine Translation Is All the Same"—Wrong Again

People assume all MT is interchangeable. They're not. Google's NMT, Microsoft Translator, DeepL—they all use neural networks, but their training data and focus differ. DeepL, launched in 2017, claims its translations are chosen roughly three times more often than Google, Microsoft, or Facebook in blind tests with professional translators (DeepL Translator (Wikipedia)). I've found DeepL often has a better ear for literary prose in European languages, while Google excels in sheer language coverage.

And don't forget specialized engines. WIPO Translate is a neural MT tool specifically for patent and technical documents (WIPO Translate (WIPO)). For a technical manual, that might beat a general-purpose engine. So before you use MT, test different engines on your text. Use the one that suits your genre. It's a tool choice, not a moral one.

VI. "More Languages Means Better Quality"—Not Necessarily

Meta's NLLB-200 supports 200 languages and improves translation quality by about 44% on average compared with previous systems (No Language Left Behind (Nature)). That sounds impressive, but it's an average across many low-resource languages. For major literary languages like French or German, the big commercial engines are still your best bet.

I once worked on a project translating a novel from a minority language. The MT output was laughably bad—no training data. I had to do everything from scratch. So don't assume that because an MT system supports a language, it can handle literature in that language. Check its performance on the FLORES-200 benchmark if you want a hint, but for literary work, trust your own judgment.

VII. "I Don't Need to Know About AI"—You Do

Some translators bury their heads in the sand, thinking AI is a passing fad. It's not. The Transformer architecture, introduced in 2017, is the foundation of modern NMT (Translation terminology). It changed everything. And it keeps evolving. Meta's SeamlessM4T even does speech-to-speech translation for up to 100 languages (SeamlessM4T (Meta AI, arXiv)).

You don't need to understand the math, but you need to understand the capabilities and limits. For example, did you know that multilingual NMT can do zero-shot translation—translating between language pairs it was never explicitly trained on (Zero-Shot Translation (Google Research))? That's amazing, but it also means you have to be vigilant about subtle errors. The more you know, the better you can collaborate with these tools.

VIII. "Literary Translation Is Dying"—The Market Says Otherwise

People love to predict the death of literary translation. But the market data tells a different story. The global language services industry was estimated at US$49.68 billion in 2023 (CSA Research), and while that's a slight decline, it's still huge. And within that, there's a growing demand for high-quality human translation that machines can't provide.

Yes, the industry is changing. But literary translation is a craft that requires human judgment, cultural insight, and artistic sensitivity. Machines can't replicate that. The translators who thrive will be those who use MT as a tool, post-edit when needed, and market themselves as experts in literary nuance. That's not dying—that's evolving.

Quick tip: When you use MT, always treat its output as a draft, never as a final. And if you're a literary translator, learn at least one CAT tool—it'll make you more efficient, not less creative.

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
• No Language Left Behind (Nature) - https://www.nature.com/articles/s41586-024-07335-x
• CSA Research - https://csa-research.com/l/media/Language-Services-and-Technology-Industry-Faces-Revenue-Decline-but-Remains-Poised-for-Transformation
• ISO 18587:2017 (ISO) - https://www.iso.org/standard/62970.html