A 2025 study found that even the newest large language models “tend to produce considerably more literal and less diverse translations compared to humans.” Publishers are trying them anyway. In January 2026, HarperCollins France confirmed it was testing machine translation on Harlequin romance novels, with freelancers editing the output. So can AI translate a novel well enough to publish? Here’s what the evidence says, where machines fail, and which book translation jobs you can safely hand to AI.
Can AI Translate a Novel? The Short Answer in 2026
AI can produce a readable draft of a novel translation in days, but it can’t reliably preserve a novel’s voice, subtext and consistency unless a human translator stays in control. Amazon says its Kindle Translate tool returns finished translations within several days. So the verdict is pretty simple: AI can draft a novel translation, but it can’t yet own one.
What “translate” means for a novel vs a document
Literary translation covers fiction, poetry and other creative writing, and the goal there is to recreate an effect on the reader, not just carry meaning across. A contract succeeds if the facts survive. A novel only succeeds if the narrator still sounds like that narrator and the jokes still land.
Large language models (LLMs) and neural machine translation handle the first job well. But fluent isn’t the same as faithful, and a 300-page book gives fluent-but-wrong text plenty of room to hide.
What the 2025 NAACL study found
The strongest recent evidence is a 2025 NAACL paper by Zhang, Zhao and Eger, which tested nine machine translation systems on literary text in four language pairs. Published human translations consistently beat LLM output, which came out considerably more literal and less varied. Put simply, machines get the words and humans get the book.
Where the Industry Actually Stands: Publishers, Translators and the Data
Publishers are testing AI translation, mostly on commercial fiction, while translators report lost work and little time saved. A heads-up on dates, though: several of the key figures below are older than they look.
Publishers testing AI-assisted translation
HarperCollins France stated in January 2026 that “no Harlequin collection has been translated solely using machine translation generated by artificial intelligence.” According to ATLF, the French literary translators’ association, freelancers edit the machine output to cut working time, and the publisher pointed to falling Harlequin sales and a wish to keep prices low.
Veen Bosch & Keuning, the Netherlands’ largest publisher and a Simon & Schuster company, announced an AI-assisted trial in November 2024 on fewer than ten titles, all commercial fiction. “We are not creating books with AI, it all starts and ends with human action,” a spokesperson said. No results have been published yet.
Then in November 2025, Amazon launched Kindle Translate in beta for self-publishing authors, labelling AI-translated books for readers. That same month, CEATL (Europe’s council of literary translators’ associations) and the European Writers Council, together representing around 260,000 authors and translators from 91 associations, urged the book sector to back human translators.
What translators report: lost work and lost time
A Society of Authors survey of 787 UK members, run in January 2024 and published that April, found that 36% of translators had lost work to generative AI, 43% had seen their income fall and 77% expected lower income in future. Nothing newer has been published, so treat this self-selected sample as an early-2024 snapshot.
Broader 2025 data from Acolad found that 84% of UK translators, across all fields, expect demand for human translation to fall. And a CEATL survey from October 2023 found roughly 70% of about 1,500 literary translators considered AI not at all or hardly useful, though that’s a baseline only.
What the studies find
In March 2024, Authors of Switzerland, the Swiss writers’ association, published a study funded by Pro Helvetia, the Swiss Arts Council. Five translators post-edited DeepL output, and the verdict was blunt: editing AI pre-translation “at best, results in no time saved and, at worst, leads to a significant loss of time.” The sample is tiny, but it echoes what translators elsewhere say.
A 2023 UMass Amherst study by Karpinska and Iyyer, presented at the WMT conference, found that critical errors persist, including occasional content omissions, even when models translate whole paragraphs. It tested GPT-3.5, so NAACL 2025 better reflects current models. For authors, the message is straightforward: an AI draft is cheap to produce, not cheap to fix.
Where Machines Fail When AI Translates a Novel: Five Recurring Problems
Translators and researchers point to five recurring failures, and each one is hard to spot unless someone reads the source and the translation side by side.
1. Voice: flattening a narrator’s register
Voice is the sound of a particular narrator: their rhythm, vocabulary and attitude. Speaking to The Markup in April 2025, translator Kira Josefsson described what automated tools strip out as “that thing that pierces the humdrum of the world and lends the world specificity.” The NAACL finding of more literal output points the same way. If every narrator drifts toward a tidy middle register, your gruff detective and your wistful teenager start to sound alike.
2. Wordplay, idiom and rhythm
Puns, idioms and sentence rhythm rarely survive word-for-word rendering, though documented examples of a specific pun failing are scarce. What matters is the mechanism. A joke built on one language’s sound needs a brand-new joke in the other language, which is closer to transcreation, recreating a text’s tone, intent and emotional effect for a new audience, than to translation.
3. Cultural subtext and register (honorifics, dialect, taboo)
The Markup documented one telling case. Russell Scott Valentino, a translator and former president of the American Literary Translators Association, tested ChatGPT on a Croatian short story by Miljenko Jergović set in Yugoslavia. It rendered a “dark-skinned” character as “Black,” when the original most likely pointed to a Jewish or Romani identity.
In the same article, linguist Emily M. Bender raised a hypothetical risk, not an observed error: a reference to Amsterdam “may be improperly translated to Washington, D.C. or London.”
Errors like Valentino’s are dangerous because the English still reads naturally. Honorifics, dialect and taboo carry the same risk, since they depend on social context that a model can misjudge without any warning.
4. Long-range consistency across 80,000 words
A character’s nickname or a running motif has to stay stable from chapter one to chapter forty. This is the failure with the thinnest evidence: the WMT and NAACL studies tested paragraphs, not books, and no public study has yet measured name and term drift across a whole novel. The safeguards are a shared glossary and a translator who reads the whole manuscript, picked through sound quality management.
5. Quiet omissions and invented smoothness
The 2023 UMass Amherst study found that occasional omissions persisted even with paragraph-level context. And because the output still reads well, nobody who reads only the translation knows a sentence has vanished.
Translator Liesl Yamaguchi put the stakes bluntly: “A bad translation is worse than no translation because it’s going to block the way to a good translation being produced.”
Here’s the thing, though: these five failures do the least damage in plain, plot-driven fiction, which is exactly where publishers have started.
What AI Does Well in Book Translation (and Why That Matters)
AI is most useful for plain, plot-driven prose, first drafts and terminology lists, and it costs far less than full human work. That first point is our reading of where publishers deploy it, not a study finding.
Plot-driven prose and dialogue-light narration
Where the story carries the book and the style stays plain, the failures above matter less. That’s why Harlequin romance and Dutch commercial fiction came first.
Nuanxed, a Swedish company offering AI-assisted book translation mostly for commercial fiction, had produced about 900 translations across more than 60 language combinations by April 2025, The Markup reported.
One caution on “blind tests.” GlobeScribe, a UK AI translation startup, launched in 2025 at a stated $100 per book per language and claims that native speakers in its own tests could not reliably tell AI from human translation. It hasn’t published its methodology.
How you evaluate matters, as the NAACL authors showed. Student raters using MQM (Multidimensional Quality Metrics), a detailed error-marking scheme, judged nearly 60% of human translations no better than machine output. But with Best-Worst Scaling, where raters simply pick the best and worst of several versions, they spotted the human translation 80% to 100% of the time.
First drafts, glossaries and terminology lists
AI also helps around the translation itself: extracting character names, building a glossary and producing a rough first pass for a translator to rewrite, which is called post-editing.
Speed and cost: AI vs MTPE vs human
Machine translation post-editing (MTPE) means a human editor corrects and rewrites machine output instead of translating from scratch. In short, MTPE costs less per word than full human translation, but voice fidelity drops along with the price.
Per-word ranges come from 2026 pricing guides by ICS Translate and Alconost for commercial translation. Other guides differ and literary rates are negotiated, so treat these as indicative. Totals are our arithmetic for an 80,000-word novel.
And those savings can shrink once a book needs heavy editing, as the post-editing studies above suggest.
Which Novels Can AI Translate? A Genre-by-Genre Decision Guide
Plot-driven commercial fiction is the safest candidate for AI-assisted work, while voice-driven literary fiction, poetry and dialect-heavy books should stay fully human. The table is our synthesis, not a published finding, and it uses publishers’ actual choices as a signal.
Green, amber and red: category-by-category verdicts
Green means the failure types hurt least, so AI plus a careful edit is reasonable. Amber means the book leans on tone or idiom, so a literary editor needs to rework the draft line by line. Red means the book’s value lives in exactly what machines flatten.
The sources show the same pattern. Veen Bosch & Keuning limited its trial to commercial fiction, and Harlequin tested category romance. Even GlobeScribe’s founders concede, “There will always be a place for expert human translation, especially for highly literary or complex texts.”
A 5-minute sample test before you commit to book translation
Pick the most voice-heavy 1,500-word chapter in your book and run it through your chosen tool. Then ask a native reader to mark every instance of the five failure types. If you can, add a best-versus-worst comparison against a human sample, since the NAACL results show detailed scoring sheets can mislead inexpert raters. And if the reader flags voice or omission problems on the first page, move your book one colour towards red.
Before You Let AI Translate a Novel: Run a Hybrid Workflow
A sound workflow starts with human decisions, uses AI for the draft and puts a literary translator in charge.
Step-by-step: style guide, glossary, draft, literary post-edit, native review
- Write a style guide. Describe the narrator’s voice, register and any dialect, with sample lines.
- Build a glossary. List character names, places and recurring terms before translation begins.
- Generate the draft. Apply the style guide and glossary, keeping the source close.
- Literary post-edit. A qualified translator rewrites for voice and rhythm and restores omissions; providers of book translation services that also offer editing can support this stage.
- Native review. A second reader checks the final text against the source; translator profiles help you match language pair and specialisation.
Why bad MTPE can cost as much as translating from scratch
The Pro Helvetia study found post-editing can save no time, and CEATL has reported that post-editing is often as much work as literary translation, or more, for less pay. Translator Sean Bye told The Markup in 2025, “Very often you’re better off redoing the whole thing from scratch.” Checking a subtly wrong draft line by line can easily take longer than translating fresh.
Rights, privacy and disclosure: what to check first
Before uploading a manuscript, check three things: ownership, data handling and disclosure.
- Pro Helvetia (March 2024) said machine-generated translations remain in a “legal grey area of ‘intellectual creation’.” Newer rulings and EU AI Act duties could not be verified here, so take legal advice.
- Data handling. Confirm whether a tool stores uploaded text; Pro Helvetia noted free DeepL stores input, while the paid version does not.
- Check your contract and retailer rules. Amazon’s Kindle Translate label shows platforms are starting to flag AI translation to readers.
Conclusion: AI Can Draft a Novel Translation, but Someone Has to Own It
So, can AI translate a novel? It can produce a fast first draft, but a human translator has to own the finished book.
The decision rule in three sentences
If your novel is plot-driven and plain in style, AI plus a careful edit may be enough. If it leans on tone, humour or idiom, use MTPE with a literary editor. And if its value lives in voice, rhythm or dialect, commission full human translation.
Not sure which category your manuscript falls into? Run the 1,500-word sample test above, then request a free, no-obligation quote from Ulatus for that chapter or the whole book, delivered within 24 hours.
![[2025] How to Make Your Book an International Bestseller](https://www.ulatus.com/translation-blog/wp-content/uploads/2025/03/Ulatus-global-blog-page-image-1280-×-720-px-12-300x172.png)
