When Latin Stops Being Latin: Can AI Translate a Language That Does Not Quite Exist?

What happens when the language you are translating does not quite exist?

Teofilo Folengo’s macaronic verse looks like Latin, behaves partly like Latin, and poses a revealing challenge for AI-assisted literary translation.

We decided to find out.

For this experiment, we gave TranswrAIte a passage by Teofilo Folengo (1491–1544), the extraordinary Renaissance poet better known under his literary persona Merlin Cocai (Merlinus Cocaius).

Folengo is one of the great practitioners of macaronic literature, a deliberately hybrid form of writing in which Latin grammar and literary conventions collide with vernacular Italian, Mantuan dialect, popular expressions, invented words, and the vocabulary of everyday rural life. The result looks like Latin. It behaves partly like Latin. But try translating it and things become considerably less comfortable.

The passage comes from the Zanitonella, more fully Zanitonella sive inamoramentum Zaninae et Tonelli, Folengo's comic-rustic account of the peasant Tonello's love for Zanina. It appears alongside the much larger Baldus, Folengo's celebrated mock-heroic epic, in his collection of macaronic works.

Its opening immediately gives the translator a problem:

Forte caleffabit gens me citadina vilanum,
quod sic Zanninae brusor amore meae,
quod ve, bovum stallas usus nettare boazzis,
sforzor amorosas fora butare doias.

Words such as

citadina, vilanum, brusor, boazzis, sforzor, mangiare...

inhabit a linguistic borderland. Italian and dialect words are pushed into Latin syntax, dressed in Latin endings, or simply allowed to wander into the sentence wearing their muddy boots.

The joke is not merely in what Tonello says: the language itself is the joke!

First experiment: translate the meaning

We began with the conventional task: make the passage readable in English prose while preserving its meaning.

The result was:

Perhaps the city folk will mock me as a country bumpkin because I burn so fiercely with love for my Zannina, because, accustomed as I am to cattle stalls and the nectar of cowherds, I struggle to pour out my pangs of love. Alas, sooner or later we are all caught in love's birdlime: whoever does not love when young, believe me, will love when old. Nor are those who drive well-fed oxen to water, who milk cows, shear sheep with their shears, feed donkeys on grass and goats on leaves, lead lean pigs to acorns, and know how to gather grains of wheat from the woods any less maddened by love than city people are. No one on earth seems able to eat his bread without having his belly pierced by love's sharp needle. Whether a man is a peasant or was born in the city, Cupid eventually drags everyone into his trap. That rascal shoots his darts this way and that, flying like a blackbird, blind and without breeches. And although the pen with which I write is as much a spade to me, and the familiar hoe has hardened my hands with calluses, and although my inkwell is a great rustic barrel from which I drink while beautiful Zannina drinks away my brains, still, alas, I have not been able to escape this terrible burning, through which the spade is finally forced to yield to the pipe.

Something interesting has happened.

The meaning is there. Tonello expects sophisticated townspeople to laugh at a peasant attempting love poetry. He responds that love is hardly an urban privilege. Cowherds, shepherds, pig keepers and farmers become just as ridiculous under Cupid's influence as everyone else.

But Folengo has disappeared, or, more precisely, one important part of Folengo has disappeared. In other words: the English is too well behaved.

Folengo's original depends on collision. Elevated literary Latin crashes into cowsheds, hoes, pigs, goats, breeches and calloused hands. Vernacular words are forced into Latin grammatical machinery. Sometimes they fit, sometimes they bulge spectacularly from the sides but the truth is that a clean semantic translation solves precisely the problem Folengo wanted to create.

So we tried again.

Second experiment: translate what the language is doing

Instead of asking TranswrAIte to reproduce only the meaning of the passage, we asked it to reproduce its linguistic behaviour.

Here was the prompt:

"Translate the following passage by Teofilo Folengo into English, but do not normalize its macaronic language. Recreate in English the comic linguistic friction of Folengo's original: keep Latin syntax, inflections, or Latin-looking forms where they remain intelligible, while mixing them with ordinary English, rustic vocabulary, and deliberately anglicized or hybrid Latin-English words. Where Folengo creates pseudo-Latin forms from Italian or dialect words, invent analogous pseudo-Latin forms from English words when possible. Preserve the rural, bodily, comic register of the original. The result should feel as though a Renaissance peasant who knows just enough Latin to misuse it magnificently were trying to write elevated English love poetry. Prioritize the linguistic effect of the original over grammatical correctness, but keep the passage sufficiently intelligible that an English reader can reconstruct its meaning. Do not turn it into modern slang or parody medieval English. The goal is not “bad English,” but macaronic English: English wearing an oversized Latin toga."

And this time we got something considerably stranger:

Perchance the citadine gens will calefact me for a villanus, because thus for my Zannina I burnor amorously, and because, accustomed to the stallas of oxen and the nectar of cowmen, I sforz myself to buttare forth these amorous dolours. Alas, once at least are we all vesciated in the birdlime of Amor: he who loveth not as a puttus, credite me, shall love as a vecchius. Nor less mattified than urban gentes are those who menate the pastured oxen to beverate, who mung the vaccas, who with forfices tosate the sheep, who pasture asinos upon grass and capras upon fronds, who know how to ducere lean porcos unto the acorns, and who know how from the busches to cavate the grains of wheat. Nemo upon terra seemeth able to mangiate his bread whose panza is not perforated by the puntigerous needle. Be he a contadinus, be he natus in urbe, once into his trapola Cupido trahit every man. He tirates his veretones hither, the improbus tirates them thither, flying like a smerlus, caecus and without bragas. And although this penna wherewith I scribo be unto me a badilus, and the celebrated zappa giveth callous hands, and although my calamar of ink be a great botazzus, from which I sugo whilst bella Zannina sugat my cerebrum, nevertheless, oyme, I could not scampare that grievous brusor, through which the vanga is sforzata at last to cedere unto the piva.

Now the English has a problem, which is precisely the point.

Calefact. Burnor. Vesciated. Mattified. Beverate. Tirates. Puntigerous. Sforzata.

They are linguistic monsters! Some are Latin, some are English pushed toward Latin, some are Italian left deliberately visible. Others are hybrids that belong comfortably to neither language.

And suddenly the English reader encounters something closer to the difficulty experienced by Folengo's original reader.

Meaning is not enough

This tiny experiment exposes a larger problem in literary translation.

Translation is often imagined as movement between two stable points:

Language A → Language B

But literature is rarely that cooperative: what happens when Language A is already unstable?

A text may be multilingual, dialectal, archaic, syntactically distorted, phonetically written, deliberately ungrammatical, or constructed from several linguistic registers simultaneously. A writer may misspell a word because a character mispronounces it. Another may import the syntax of one language into another. Another may invent words whose meaning comes partly from the fact that they should not exist.

In such cases, correcting the language while translating it can mean destroying information.

The first translation of Folengo tells us what Tonello says; but the second attempts to tell us something about what Folengo is doing.

These translations answer different questions.

That distinction becomes especially important with AI-assisted literary translation. A language model is extraordinarily good at normalization. Give it damaged grammar, unusual syntax or hybrid vocabulary and its instinct is often to infer the intended meaning and produce fluent target-language prose.

Usually, that is useful, but with guys as Folengo, it can be catastrophic.

The ugliness, contamination and grammatical violence are not defects waiting to be repaired: they are literary devices.

Wrap up!

Instead of treating the model as a machine for producing a single definitive translation, we can use it to explore a space of possible translations: one semantic, another rhythmic, another syntactic, another historical, another deliberately foreignizing.

The human translator can then decide which losses matter, just because every translation loses something.

Folengo makes that unusually visible. Translate him into polished English and we preserve Tonello's argument while sacrificing much of the comedy encoded in the language. Push the English itself into macaronic territory and we recover some of that instability, but at the cost of lexical precision and historical authenticity.

There is no magic prompt that eliminates the trade-off.

That is one of the ideas behind TranswrAIte: not to make literary translation automatic, but to make its alternatives visible.

Five centuries before large language models, Merlin Cocai was already stress-testing the idea of translation.

Rather inconveniently for us, he was doing it with cows.

Field Notes · by TranswrAIte

How to evaluate an AI book translation without losing narrative voice

How to evaluate an AI book translation without losing narrative voice

A narrator describes a kitchen as “neat.” Elsewhere, the same room is “orderly,” “tidy,” and “immaculate.” Each choice may be defensible in isolation. Across a novel, however, the variation can alter the narrator’s…

A narrator describes a kitchen as “neat.” Elsewhere, the same room is “orderly,” “tidy,” and “immaculate.” Each choice may be defensible in isolation. Across a novel, however, the variation can alter the narrator’s attitude toward the room and its owner.

This is one reason a few fluent pages cannot establish whether a book translation preserves narrative voice. An author needs to examine how the translation handles recurring choices across chapters: diction, sentence movement, distance from characters, dialogue, imagery, and deliberate repetition.

AI-assisted translation can support long-form work, but its output still needs editorial evaluation. The following process gives authors, translators, and editors a shared basis for that review.

Define voice in observable terms

“Keep my voice” is a reasonable request, but it is difficult to assess until the voice has been described through features on the page.

Start with a short voice brief. Record the manuscript’s relevant patterns, such as:

- the narrator’s level of formality - typical sentence length and variation - direct or indirect treatment of emotion - distance between narrator and character - use of fragments, repetition, dialect, or technical language - differences among major characters’ speech - recurring images, jokes, or verbal habits

Add examples from the source manuscript. A note such as “the narrator avoids clinical terms when discussing illness” gives an editor more to work with than “the prose should feel intimate.”

This brief is not a command to reproduce source-language syntax. Languages distribute tone, rhythm, and information differently. It is a record of the effects that deserve attention during translation and revision.

Build a representative test set

Do not select only the opening pages. Openings often receive more revision from the author and may not contain the book’s hardest translation problems.

Choose passages that test distinct features of the manuscript. These might include a quiet descriptive scene, dialogue between characters with different speech patterns, an emotionally restrained passage, a section containing repeated terminology, and a later chapter that depends on earlier context.

Include any passage that worries you. A joke, ambiguous pronoun, invented term, abrupt point-of-view shift, or repeated phrase can reveal how the workflow handles context.

Keep the source and translated passages aligned for review. Record the chapter, scene, speakers, relevant earlier references, and any intentional ambiguity. This prevents reviewers from treating a purposeful choice as an error.

Review effect before elegance

A translation can read smoothly while changing the book’s narrative stance. Begin by asking what each passage does.

For narration, examine:

- Has the level of certainty changed? - Has implication become explicit explanation? - Does the narrator sound warmer, harsher, more formal, or more knowing? - Have repeated words been replaced with varied synonyms? - Has an awkward or compressed source passage become polished in a way that changes its function?

For dialogue, compare character by character. Check forms of address, contractions, interruptions, sentence length, politeness, slang, and repeated expressions. Two characters who are distinct in the source should not acquire the same neutral register in translation.

Rhythm also needs attention. Reading aloud can reveal changes that a line-by-line comparison misses. Listen for the placement of pauses, the length of descriptive runs, and the speed of action. The translated sentences do not need to mirror the source mechanically. They should create an appropriate reading experience in the target language.

Track decisions across the manuscript

Book translation produces decisions that recur hundreds of pages later. Names, invented terms, titles, objects, institutions, and character-specific phrases need a shared record.

Create a terminology and decisions log with fields for:

- source term - approved translation - context or definition - character or narrative register - chapter of first use - exceptions - unresolved questions

The log should include stylistic decisions as well as nouns. If a character always addresses another character formally until a specific scene, record that change. If a recurring phrase develops a second meaning later in the book, note both uses.

Searches across the complete translated manuscript can then identify inconsistent forms. Each result still requires contextual judgment. A term may need different translations in different scenes, especially when wordplay or character knowledge is involved.

Use back-translation carefully

Translating a target passage back into the source language can expose omissions, added explanation, or a changed factual relationship. It cannot establish literary quality on its own.

A back-translation may look different because the target-language sentence uses a natural idiom. It may also hide a tonal problem by converting the sentence into neutral source-language prose. Treat it as a diagnostic aid, then ask a qualified reader of the target language to assess the actual translation.

Give reviewers a focused rubric

A general request to “check the translation” can produce scattered feedback. A short rubric makes comments easier to compare.

Ask reviewers to mark each passage for:

- accuracy of meaning and relationships - consistency with the voice brief - character differentiation - terminology and continuity - naturalness in the target language - passages requiring author or editor decisions

Require comments and examples for important judgments. A numerical score without an explanation gives the next editor little guidance.

The author can assess continuity, intention, and acceptable departures from the source. A literary translator or editor who works in the target language can assess register, idiom, rhythm, and cultural implications. These roles overlap, but they are not interchangeable.

Test editorial control before the full project

Before committing a complete manuscript, establish how revisions will work. Ask who can change terminology, where unresolved questions are recorded, how updated chapters are tracked, and whether decisions can be applied consistently to later sections.

Also ask what files you will receive. Editable text, a decisions log, version labels, and clear chapter boundaries can make editorial review easier. Confirm who is responsible for the final target-language edit and proofreading.

Cost comparisons need the same specificity. Determine whether an estimate covers translation only or also includes preparation, terminology work, revision, specialist review, formatting, and corrections after review. A lower initial figure may describe a smaller scope.

Confidentiality should be addressed directly. Ask how manuscript files are stored, who can access them, whether external providers are involved, how long files are retained, and what deletion process is available. Review the applicable terms before uploading unpublished work.

Make the decision from patterns

A single mistranslated sentence can be corrected. Repeated flattening of character voices signals a broader editorial problem. The purpose of a test is to find those patterns before they extend through the book.

Record which issues are isolated and which recur across the sample. Then decide whether the workflow supports the manuscript’s specific needs: complete-book context, traceable terminology, author queries, revision, and target-language review.

TranswrAIte supports context-aware translation for complete books and other long-form manuscripts, with a publishing-oriented workflow and human review support. Authors preparing an international edition can request a quote to discuss the manuscript, language pair, review needs, and project scope.

The translator cliff

What the decline of professional translation tells us about designing AI translation for books.

AI translation is changing the economics of professional work—and exposing a crucial design choice about where human attention belongs.

A recent Financial Times article by Sarah O’Connor and John Burn-Murdoch has a rather uncomfortable title: “How AI has de-skilled translation.”

It is uncomfortable partly because the numbers are difficult to argue with.

In the United States, translator employment began losing momentum around the time Google Translate became widely used, and in the past five years the number of translators has fallen substantially relative to total employment. Typical translator wages, meanwhile, have slipped below the economy-wide average, while the number of new translation projects advertised on major online freelance marketplaces fell by almost 50 per cent in the two years after ChatGPT appeared.

But the most interesting part of the article, at least from where we sit at TranswrAIte, is not that machines are becoming capable of translation. We have built an entire product on the assumption that they are.

It is what happened to the humans once the machines arrived.

Welcome to the post-editing factory

The translation industry did not wake up one morning and replace every translator with an LLM. The transition has been more mundane.

Many agencies adopted what is called Machine Translation Post-Editing, or MTPE. Instead of giving a translator an original text and asking them to translate it, the machine produces a first version and the human translator receives the result with instructions to check it, correct it and polish it.

On a spreadsheet, this looks wonderfully efficient. The expensive human no longer has to translate every sentence. The machine has already done the heavy lifting, so the translator merely needs to tidy things up.

The FT gives a striking example. Petr Čermoch, who translates television subtitles from English into Czech, says that one agency which previously paid $5 per minute of video reduced its rate to $1.50 after moving to MTPE. Another translator working across legal, financial and academic texts reported rates being cut by roughly half.

And the work did not necessarily become easier.

Mark Rawson, an English-Chinese translator interviewed by O’Connor, describes the cognitive burden of simultaneously interrogating the source and the machine-generated translation. His verdict is wonderfully concise: “It’s two or three times harder than pure translation work.”

That sentence caught our attention, because there is something peculiar about the whole arrangement.

We have taken one of the most interesting parts of translation, deciding what a sentence means and how it should live in another language, and given it to the machine. Then we have taken one of the least interesting parts, hunting for errors, inconsistencies and awkward formulations, and given it to the human.

And somehow we call this human in the loop. But which loop?

Translating a novel is not simply a sequence of sentences that can be processed independently. A word on page 17 may acquire another meaning on page 280. A character speaks differently from another character. A metaphor returns. A joke depends on something that happened three chapters earlier. The narrator has habits, rhythms, prejudices and favourite constructions. Sometimes an awkward sentence is accidentally awkward. Sometimes it is awkward because the author wanted it that way.

A translator working from scratch gradually builds a mental model of all this. The book becomes a system.

Traditional MTPE breaks that system into little pieces and asks the translator to inspect the output of another intelligence. The translator is no longer primarily asking, How should I translate this? The question becomes, Is what the machine produced acceptable?

The second creates an anchor. Once a plausible sentence is sitting in front of you, it becomes surprisingly difficult not to inherit its choices. Vocabulary, syntax, rhythm and interpretation have already been proposed. The human becomes a critic of a translation rather than its author.

For technical material, that trade-off can make perfect sense. For literature, it can quietly sand away precisely the things we care about.

This is why our pipeline does not begin with post-editing

When we started building TranswrAIte, one tempting architecture was obvious: generate a translation with an LLM and provide tools for a human to correct it.

It is simple. It is cheap. It resembles the workflow the translation industry already knows.

But we increasingly think that treating AI translation as machine output plus human correction misunderstands what these models are good at.

Modern LLMs can do more than replace words across languages. They can analyse context, compare alternative interpretations, reason about character voice, follow stylistic instructions and revisit earlier decisions. That does not make them infallible translators. Far from it. But it means the interesting engineering problem is no longer merely how to generate the first translated sentence.

It is how to surround generation with enough context, analysis, memory and verification that the model can make better decisions in the first place.

That is why we think in terms of a pipeline, rather than a single translation call.

The system needs to understand the book before and while it translates it, maintain information across chapters, preserve terminology and names, identify passages where literal translation fails, review its own choices and flag uncertainty rather than quietly burying it beneath fluent prose.

There is a lesson here beyond translation

O’Connor and Burn-Murdoch make an interesting comparison with software development.

So far, AI appears to be affecting experienced programmers differently. Coding agents can remove some of the repetitive work of producing code, leaving developers more time for architecture, coordination, problem-solving and judgment. Translation has often experienced the reverse: the technology takes over the creative act of producing the translation while leaving the human with repetitive inspection.

That difference may tell us something important about AI products in general.

Again, the useful question is: Which percentage are we automating?

Automate the tedious part and technology can amplify expertise; automate the interesting part and leave humans to clean up the debris, and you may technically have a human in the loop while simultaneously removing much of the reason a skilled human wanted to be there.

That seems to be what happened to a significant part of professional translation.

Translation is not disappearing.

None of this means we believe the old translation market is coming back.

The FT data suggests quite the opposite. Automated translation first weakened demand, then improved, then became dramatically more accessible with large language models. Companies are increasingly willing to use machine translation without any human post-editing at all.

And there is another side to that story which is easy to overlook.

Cheaper translation also means more things can be translated.

For books, that matters.

A traditionally published bestseller can justify several thousand euros for a professional translation because the publisher expects enough sales to recover the investment. An independent novelist with a 120,000-word manuscript usually cannot make the same calculation. Nor can a small publisher afford to test five languages simply to discover whether readers exist there.

Let's be honest here: those books were not being beautifully translated by humans before AI arrived.

They were usually not being translated at all!

This is the part of the economics we find particularly interesting. AI is certainly reducing the cost of work that previously belonged to professional translators, and the consequences for that profession are real. But it is also pushing the marginal cost of translation low enough to create a market where previously there was none.

The question therefore becomes less interesting when framed as human versus machine.

For us, the better question is what kind of translation system we can build once machine translation becomes cheap enough to be treated not as the final product, but as one component inside a much larger process.

Perhaps the strange mistake of the first AI translation era was believing that the machine should translate and the human should clean up.

We are trying the opposite philosophy: make the machine translate, criticise, reconsider, remember, compare and clean up as much as possible, then leave the human where human attention is actually valuable.

That feels less like de-skilling translation and more like redesigning it.

Sources

• Sarah O’Connor and John Burn-Murdoch, “How AI has de-skilled translation,” Financial Times, June 4, 2026

The Book Is Yours. The Translation Should Be Too.

The Book Is Yours. The Translation Should Be Too.

AI has made book translation cheaper and faster—but authors should ask who owns, controls, and can revise the translated edition when the machine finishes.

There was a time when getting a novel translated required a publisher, a contract, a translator, an editor, some money and quite a lot of patience. For the independent writer, especially one working outside English, this often meant that the door to another market was not exactly closed, but it had a large man standing in front of it asking to see your bank account.

Artificial intelligence has changed that arithmetic.

Translation is becoming dramatically cheaper and faster, and this matters because the book market itself is already being transformed by the same economics. A recent paper by Tuhin Chakrabarty, Xinyue Liu, Jane C. Ginsburg and Paramveer Dhillon examined 14,419 self-published genre-fiction ebooks sold on Amazon between 2023 and 2026. Their conclusion is uncomfortable and fascinating: generative AI does not need to produce better books to change the market. It can change the market simply by producing more books.

That distinction matters.

The researchers found that books containing substantial amounts of detected AI text represented 20% of their sample, but only 12.1% of sales and 11.3% of revenue. Human-written books still performed considerably better. Yet the AI books were multiplying so quickly that their individual mediocrity did not matter very much. Some reached the top of the rankings, and their share of observed sales rose from almost nothing in early 2023 to roughly 20% by the second quarter of 2026.

The machine does not have to write East of Eden. It can arrive with ten thousand books while you are still arguing with Chapter Seven.

And that is where things get interesting for translation.

AI has created a new translation economy

For authors, the important consequence of AI is not that everyone should start generating novels by pressing a button. The more interesting consequence is that something historically expensive has suddenly become accessible: the internationalization of an existing book.

A novelist sitting in Naples, Lyon, Warsaw or Buenos Aires no longer has to regard the English-language market as distant territory simply because translating a 400-page novel professionally might cost thousands of euros. The economics are changing underneath our feet.

And Amazon has noticed.

Amazon Kindle Translate is currently an invitation-only beta that allows eligible KDP authors to generate AI translations of their ebooks. The service is free during the beta, and authors can preview the resulting translation before publishing it. Amazon currently supports translation from English into French, German, Italian, Brazilian Portuguese and Spanish, while those languages can be translated into US English.

For an independent author, the proposition is wonderfully seductive.

Your book is already on KDP. Click a button. Choose Spanish. Wait a few hours. Voilà: Madrid.

There is only one problem.

Read the terms.

Amazon will translate your book. But it will not give you the translation.

This is the part authors should understand before clicking anything.

Under the current Kindle Translate beta terms, Amazon retains ownership of the Kindle Translate Edition files. You cannot take that Kindle Translate edition and distribute it somewhere else. The printing and distribution rights granted to Amazon for those editions remain exclusive to Amazon parties.

There is an important nuance here. Amazon does not prevent you from commissioning or creating a different translation of the same book in the same language and selling that elsewhere, unless other exclusivity provisions such as KDP Select apply. But the translation produced through Kindle Translate itself stays inside Amazon's garden.

That is a rather extraordinary exchange when you think about it.

You bring the novel.

You bring the characters, the plot, the voice, the years spent writing it and, crucially, the translation rights Amazon requires you to possess in the first place.

Amazon brings the machine.

And the resulting Kindle Translate edition cannot leave Amazon.

There is another limitation that may matter even more to literary authors: you currently cannot directly edit the translated text. You can preview the translation before publication, but direct editing is not available.

Amazon does run translations through an automated seven-step quality-assurance process covering mistranslations, idioms, terminology, forms of address, gender consistency, grammar, spelling and other literary elements. That is useful quality control. But it is not the same thing as putting the translation into the hands of someone who can argue with a sentence.

And books are full of sentences worth arguing about.

Translation is not a checkbox

This is where the distinction between AI translation and AI-assisted translation becomes important.

A literary translator does considerably more than determine what a sentence means. There is rhythm, register, repetition, characterization, historical context, jokes, ambiguity and those peculiar little decisions an author makes that look accidental until somebody “corrects” them.

AI can now perform an astonishing amount of the first-pass work. Pretending otherwise is not going to protect the translation profession. The economics have already moved.

But precisely because AI has made translation cheap, authors have a new possibility: use the machine for what it does extraordinarily well, while keeping humans in the places where judgment matters.

The result can be an entirely different economic model from traditional literary translation. Instead of paying someone to translate hundreds of pages from scratch, AI can generate the working translation and human expertise can be concentrated on revision, difficult passages, stylistic consistency and final validation.

Most importantly, the author can retain control of the resulting translation.

That distinction becomes increasingly important in the market described by Chakrabarty and colleagues. Their study found that between 2023 and 2026 the number of selling titles grew 19.2-fold while quarterly revenue grew only 8.9-fold. There were vastly more books chasing a pool of reader attention that was growing much more slowly.

The authors call the phenomenon dilution. AI makes production so cheap that volume itself becomes a competitive force. Their broader conclusion is particularly striking: generative AI can reshape a creative market through scale rather than quality.

For independent authors, there is a lesson hiding inside those numbers.

If the market is becoming larger, faster and more crowded, then reaching readers in several languages may become increasingly valuable. But surrendering control over those translations to the same platform that already controls much of your distribution seems an unnecessarily expensive way to obtain something advertised as free.

Free translation has a price

Amazon's offer is not sinister. In fact, Kindle Translate is evidence of something genuinely exciting: AI translation has crossed the threshold from experiment to publishing infrastructure.

The question is no longer whether machines can translate books. Guess what? They can.

The more useful question for authors is who controls the translation after the machine has finished.

If you translate your novel into English, Spanish, French or German, that new edition can become an asset in its own right. You may want to publish it on several stores, sell it directly, submit it to foreign publishers, produce a print edition, create an audiobook, license territorial rights or revise the translation five years from now when the technology has improved.

A translation should therefore be treated much like the manuscript from which it came: something worth owning, editing and carrying with you.

The cheapest translation is not necessarily the one that costs zero dollars today.

Sometimes the expensive part comes later, when you discover that the door was free because somebody else kept the key.

Sources

• Tuhin Chakrabarty, Xinyue Liu, Jane C. Ginsburg & Paramveer Dhillon, “Generative AI floods and dilutes the market for books” (2026)
• Kindle Translate Policy

The Martian

"The Martian" Had One Advantage You Probably Don’t

Andy Weir could test his novel directly in the world’s largest literary market. For writers outside English, translation has always been the gate before the gate.

There is a wonderful publishing story about a man who leaves an astronaut alone on Mars and, in the process, accidentally rescues himself from obscurity.

Andy Weir began publishing The Martian chapter by chapter on his website, for free, after having little luck with literary agents on earlier work. Readers came, not millions of them at first, but that small and precious tribe every unknown writer dreams about: people who actually wanted the next chapter and were willing to come back for it.

Some of those readers eventually asked for a Kindle version because reading a novel on a website was a nuisance, so Weir obliged and put it on Amazon for 99 cents. The book started selling, then started selling rather well, and within three months it had moved around 35,000 copies. Eventually the strange machinery of traditional publishing, which had previously shown little interest in him, began moving in the opposite direction: Crown Publishing bought the print rights, and the book was commercially published in 2014.

From there, things became almost indecently successful. The Martian became a #1 New York Times bestseller; Ridley Scott directed the film adaptation, Matt Damon played Mark Watney, and the movie arrived in cinemas in October 2015.

It is one of the great self-publishing stories, and writers like to tell it because it contains everything we need to keep ourselves alive at three in the morning: rejection, stubbornness, readers discovering something without being told to discover it, and finally the enormous publishing industry arriving late to the party with a contract in its hand.

But there is one detail in the story that is so obvious we rarely mention it: Andy Weir writes in English.

That doesn’t explain why The Martian succeeded, of course. The book succeeded because people loved it, and the English language contains warehouses full of unsuccessful novels, so being a native English speaker is clearly not a publishing strategy. What it gave Weir, however, was something extraordinarily valuable: he could test his novel directly in the largest international literary marketplace without first having to cross a language barrier.

And that changes the story.

The gate before the gate

Imagine the same writer somewhere in Italy, France, Spain or Germany. He has written a good novel, perhaps a very good one, but his local publisher doesn’t want it, or he has self-published it and found a few thousand readers, or he has sent it to enough literary agents to develop an intimate relationship with rejection emails. Eventually he thinks: fine, let’s try English.

And immediately he discovers another problem, because before an English-speaking reader can decide whether his novel is worth reading, the novel has to become English.

The current Penguin Random House paperback of The Martian runs to 416 pages, while the ebook is listed at 400. For a professional literary translation, a manuscript of that size is not a small experiment but a serious investment, made before you have the slightest evidence that anybody in the target market wants the book at all.

This creates a peculiar asymmetry in self-publishing. An English-language novelist can write a book, put it online, lower the price, give away chapters, experiment with covers and descriptions, run advertisements, send copies to reviewers and generally make a fool of himself in public until something works. A novelist writing in another language can do exactly the same thing at home, but the moment he wants to test the English market, he discovers that admission to the casino can cost thousands.

For an established publisher, this is simply an investment decision involving editors, budgets, foreign-rights departments and sales projections. For an independent novelist sitting at a kitchen table with a finished manuscript and a cup of coffee that went cold forty minutes ago, it is something rather different.

It is a wall.

Literary agents cannot reject a book they cannot read

This is where the romantic story of self-publishing becomes less romantic, because although we often say that digital publishing removed the gatekeepers, what it really did was remove some of them. Nobody needs permission to publish a Kindle book or build a website and start releasing chapters, but language itself remains a gatekeeper, and a remarkably efficient one.

If your novel is written in Italian and you want to discover whether American readers might like it, you cannot simply repeat Andy Weir’s experiment. You cannot put the Italian manuscript on Amazon US, wait six months and conclude from the silence that Americans are strangely indifferent to your masterpiece about a divorced taxidermist in Turin.

First you need an English book.

The same problem appears when approaching English-language literary agents. A synopsis can be translated cheaply and a sample can be translated beautifully, but sooner or later somebody needs the manuscript, and when that manuscript contains 80,000 or 100,000 words, the economics become uncomfortable very quickly.

Traditional publishing therefore has a language gate before its famous editorial gate. For native English writers that first gate is almost invisible; for everyone else, it can be the most expensive one.

And then came the machines

This is where AI translation becomes interesting to us, and not because we believe literary translators should disappear into a puff of GPU smoke. What interests us much more is the way AI changes risk.

Until recently, translating an entire novel into English was something you normally did after somebody had decided that the book had commercial potential: a publisher acquired foreign rights, commissioned a translator, edited the translation and released the book. For an independent author, reversing that order was financially painful because the large expense came before the market test.

AI changes the economics enough that the order itself can change.

An author can translate the whole manuscript, work through it carefully, revise problematic passages, compare alternatives, correct voice and terminology, and produce an English edition good enough to put in front of actual readers without first spending a small fortune merely to discover whether those readers exist.

And this is an important distinction, because the first translation does not necessarily have to be the translation carved into marble for the next hundred years. It can also be a market probe, a relatively inexpensive way of asking a very expensive question:

Does this book travel?

Perhaps fifty people buy it and nothing happens, in which case the experiment has given you an answer. Perhaps five hundred people buy it and reviews begin appearing; perhaps one reviewer sends it to somebody else, an agent reads it, or nothing happens for six months and then, for reasons nobody will ever completely understand, something does.

This is more or less how publishing has always worked. The difference is that writers working outside English have traditionally needed considerably more money just to buy a ticket.

The point isn’t to manufacture another Martian

There is a dangerous lesson to take from Andy Weir’s story, which is that if you self-publish your novel for 99 cents, Ridley Scott will eventually telephone. He probably won’t, and if that is your business plan you may want to keep your day job a little longer.

The useful lesson is less spectacular and much more interesting: Weir was able to expose his work directly to readers, and those readers produced the signal that the publishing industry eventually noticed. The audience came before the big publishing deal, not after it, which meant that his book had the opportunity to prove itself in the market rather than waiting for somebody inside the industry to predict whether a market existed.

For a writer working in a language with five million, twenty million or sixty million speakers, AI translation creates the possibility of running a similar experiment in a language spoken and read by hundreds of millions. It does not make the experiment successful, it does not guarantee discovery and it certainly does not persuade Matt Damon to spend two hours growing potatoes on Mars for you, but it makes the experiment possible.

That may turn out to matter much more than all the grand predictions about AI replacing translators or writing novels by itself. The quieter revolution is simply that a novelist who has already been told no can afford to ask a much larger group of readers the question directly.

For decades, a writer outside the English-speaking world could look at the English market through the window and think, I’d like to see what happens over there, only to sit down, calculate the cost of translating 400 pages, and quietly close the window again.

Now the arithmetic is changing.

Andy Weir was lucky enough to be writing in their language. The rest of us are getting machines.

The Translation Lab

Baudelaire Goes to the Model Arena: Literary Translation

Nine AI translation pipelines, four languages, one magnificently unpleasant poem—and a Silver model that beat Gold.

In the previous Field Note, we discussed a rather provocative paper showing that readers could not reliably distinguish AI-generated poetry from poetry written by humans and, even more inconveniently, sometimes preferred the machine-made poems. That naturally left us with another question: if today's LLMs can produce something that looks and feels like poetry, can our translation pipeline preserve poetry when moving it from one language to another? There was only one reasonable way to find out. We went back to Baudelaire.

There are reasonable texts to use when testing a translation system. Contracts are reasonable. Instruction manuals are reasonable. Newspaper articles are reasonable. They contain sentences that generally want to be understood, which is a useful quality when you are trying to measure whether a translation works.

We chose Baudelaire.

More precisely, Spleen (LXXVIII) from Les Fleurs du mal, a magnificently unpleasant poem in which the sky becomes a lid, the earth turns into a wet prison, Hope flaps around like a bat against rotten ceilings, spiders install themselves inside our brains and Despair eventually plants a black flag on the poet's skull.

It seemed perfect.

Not because we expected to settle whether AI can “translate poetry”, a question large enough to keep translators and moderately drunk philosophers occupied indefinitely, but because we wanted to ask something smaller: given the models available today, what happens when we build a literary translation pipeline around them, and how much do we have to spend before the translation actually gets better?

Why Baudelaire?

Spleen is a cruel test because most of the things that make it work are exactly the things that disappear when translation becomes sentence replacement. The poem accumulates rather than merely proceeds: stanza after stanza begins with “Quand”, the syntax keeps stretching forward and the images become progressively more suffocating, until the reader is trapped inside the sentence with Baudelaire.

A translation can get every noun right and still kill the poem.

That is precisely the sort of problem we care about at TranswrAIte. Semantic accuracy is necessary, but literary translation lives in a more troublesome neighbourhood, where voice, rhythm, repetition and even punctuation matter. Sometimes a sentence must remain strange because the original is strange, and polishing it into impeccable contemporary prose is precisely the mistake.

Baudelaire gives a translation system plenty of rope with which to hang itself. We accepted the offer.

Nine pipelines walk into a poem

We tested OpenAI, Anthropic and DeepSeek pipelines on the same French poem, translating it into Italian, English, Spanish and German, with configurations we called Bronze, Silver and Gold.

Every translation required two calls rather than one. A first model produced the draft, then an editorial model compared it with the French original and corrected omissions, mistranslations, awkward phrasing, punctuation, formatting and failures of voice. Only this reviewed version entered the benchmark.

This is much closer to how we think LLM literary translation should work: not one gigantic artificial translator confronting Baudelaire alone, but something resembling a tiny editorial office where somebody drafts, somebody rereads and somebody quietly changes the sentence.

The LLM version has the advantage that nobody needs coffee.

Then came the judges

The finished translations were anonymously evaluated by OpenAI GPT-5.6-sol and Anthropic Claude Opus 5, with fidelity, voice and naturalness carrying most of the score. We also deliberately damaged translations by removing passages, duplicating sentences and changing numbers; both judge families detected 100% of the stored planted defects.

This does not transform an LLM into a great literary critic. Detecting a missing paragraph and deciding whether Baudelaire has survived his journey into German are very different intellectual activities, but at least we knew the judges could recognise a corpse when one was lying on the floor.

Silver beat Gold

Then our expensive horse lost.

OpenAI Silver achieved the highest average quality at 4.129/5, followed by OpenAI Gold at 4.050 and Anthropic Gold at 3.879. The interesting part appears beside the bill: Silver cost roughly $0.057 per completed translation and review, while Gold cost about $0.226.

Gold was approximately four times more expensive and slightly worse.

There is a particular pleasure in a benchmark knocking your tidy assumptions off the table. We had built Gold to be Gold: more reasoning, more expensive inference, larger budgets. You expect it to enter wearing the better suit. Instead Silver came through the kitchen and stole its chair.

The quality-versus-price graph makes this particularly clear. Silver sits near the top at around six cents, while Gold has travelled far to the right without climbing any higher; Anthropic Bronze, meanwhile, reaches a respectable 3.858/5 for roughly $0.019.

For one poem these differences are pocket change. Across novels, catalogues and multiple languages, they become architecture.

There was no “best model”

Things became even less tidy when we looked at individual languages. Anthropic Gold won Italian and Spanish, OpenAI Silver won English, and OpenAI Bronze won German.

So the overall winner did not sweep the languages, and one Bronze configuration beat everybody in German.

This suggests that “Which model is best for translation?” may simply be the wrong question. Italian Baudelaire and German Baudelaire are already different problems, before we even reach dialogue, historical prose, dialect or a 500-page novel whose narrator changes register halfway through.

The more useful question is which model is best for which language, at which stage, for which kind of text?

Once you begin thinking that way, sending an entire book through one enormous model starts to look rather primitive.

And then we threw humans into the machine

We also imported one existing human translation for each language as an anonymous candidate. The judges were not told they were human, because calling something a “human reference” gives it a little crown before the competition begins.

The results were peculiar: English scored 3.975, Italian 3.600, German 2.575 and Spanish only 1.750.

Our conclusion was emphatically not that AI had defeated human translators. The Spanish and German versions may be freer adaptations, structurally different editions or translations pursuing objectives our rubric penalises, and those results require manual examination rather than triumphal declarations about machines conquering literature.

Still, the experiment left us with a principle worth keeping: human is provenance, not a quality score, and AI is provenance too.

Eventually there is only a text on the page, and somebody has to read it.

What we learned

The interesting engineering problem is becoming less about persuading one enormous model to translate a book beautifully in a heroic single pass, and more about building a system capable of choosing the right model, reviewing its work, detecting suspicious passages and leaving alone the sentences that already work.

In other words, the future of AI literary translation may look less like one brilliant artificial translator and more like a slightly neurotic publishing house, with several editors leaning over the same manuscript and disagreeing about a comma.

This experiment was only one poem in four languages, so it proves very little about literature at large. But it gave us three useful clues: more expensive does not necessarily mean better, different languages reward different models, and the pipeline matters at least as much as the model sitting inside it.

For this particular rainy, claustrophobic, spider-infested trip through Baudelaire's skull, Silver won.

Gold cost four times more and came second, which feels, somehow, like an ending Baudelaire might have appreciated.

Spleen (LXXVIII)

Charles Baudelaire · Les Fleurs du mal

Quand le ciel bas et lourd pèse comme un couvercle
Sur l'esprit gémissant en proie aux longs ennuis,
Et que de l'horizon embrassant tout le cercle
Il nous verse un jour noir plus triste que les nuits ;

Quand la terre est changée en un cachot humide,
Où l'Espérance, comme une chauve-souris,
S'en va battant les murs de son aile timide
Et se cognant la tête à des plafonds pourris ;

Quand la pluie étalant ses immenses traînées
D'une vaste prison imite les barreaux,
Et qu'un peuple muet d'infâmes araignées
Vient tendre ses filets au fond de nos cerveaux,

Des cloches tout à coup sautent avec furie
Et lancent vers le ciel un affreux hurlement,
Ainsi que des esprits errants et sans patrie
Qui se mettent à geindre opiniâtrement.

— Et de longs corbillards, sans tambours ni musique,
Défilent lentement dans mon âme ; l'Espoir,
Vaincu, pleure, et l'Angoisse atroce, despotique,
Sur mon crâne incliné plante son drapeau noir.

Sources

• Spleen (LXXVIII), Les Fleurs du mal (1861)

The Poetry Turing Test Is Over

The Day ChatGPT Beat Shakespeare (And Why That's Not the Story)

Readers preferred AI-generated poems to canonical poetry—but the most revealing result may be what the experiment tells us about reading itself.

Every few months, a paper about artificial intelligence appears that promises to settle one of the great cultural anxieties of our time. This one seemed particularly definitive. Its title alone was enough to provoke equal measures of excitement and despair: AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably. If you only read that headline, it would be easy to conclude that poetry, after surviving centuries of wars, dictatorships, market crashes and literary manifestos, had finally been defeated by a chatbot.

Fortunately, like many good papers, this one becomes far more interesting the moment you move beyond the headline.

What the researchers actually discovered is not simply that people struggle to distinguish poems written by ChatGPT from poems written by Shakespeare, Emily Dickinson or Sylvia Plath. The truly fascinating finding is that many readers systematically mistake the qualities that make AI-generated poetry immediately enjoyable for the qualities that make great poetry enduring. The experiment therefore tells us at least as much about the way we read as it does about the way language models write, and that makes it surprisingly relevant for anyone working on literary translation.

The experimental design was almost disarmingly simple. The researchers selected fifty poems by ten well-known English-language poets spanning several centuries, from Chaucer to Dorothea Lasky, and then asked ChatGPT 3.5 to generate poems “in the style of” each of those authors. Importantly, they resisted every temptation to improve the outputs: they did not refine the prompts, regenerate unsatisfying results or ask humans to select the best examples. The first generations produced by the model became the dataset.

More than sixteen hundred participants were then presented with a mixture of authentic and AI-generated poems and asked a deceptively straightforward question: which ones had been written by humans?

Their performance was not merely mediocre. It was statistically worse than chance.

Participants correctly identified the author only 46.6% of the time, and even more remarkably, they were consistently more likely to identify the AI-generated poems as human than the poems actually written by famous poets. In other words, the machine was not simply convincing enough to pass as human; it often appeared more human than the humans themselves.

At this point, the story already feels slightly unsettling. But the second experiment is where things become genuinely difficult to interpret.

Rather than asking participants to identify the author, the researchers asked them to evaluate the poems themselves across a wide range of aesthetic dimensions, including beauty, rhythm, imagery, emotional impact, meaningfulness and overall quality. Here again, the results were striking. Across almost every category, participants preferred the AI-generated poems, consistently assigning them higher scores than poems written by canonical authors. Only originality resisted this trend; readers did not consider AI poems significantly more original, but in virtually every other respect they found them more appealing.

At first glance, this sounds like a devastating result for human literature. If ordinary readers genuinely prefer ChatGPT's poems to those of Whitman or Eliot, perhaps the machines have already won.

But that interpretation rests on a hidden assumption: that the qualities readers reward after a single encounter are the same qualities that make literature valuable over decades or centuries.

The discussion section of the paper suggests a different explanation, and it is far more persuasive. The AI-generated poems were generally easier to understand. Their emotional trajectory was clear, their imagery direct and their themes immediately recognizable, whereas the human poems often relied on historical references, layered metaphors, unusual syntax or deliberate ambiguity. Participants repeatedly described authentic poems as “not making sense,” while the AI poems communicated exactly what they wanted to communicate without asking readers to linger, reread or interpret.

That observation resonates far beyond poetry. Large language models are remarkably good at producing language that feels fluent, coherent and immediately satisfying because that is precisely what they have been optimized to do. Great literature, however, has rarely optimized for immediate satisfaction. Many of the most celebrated works in literary history derive their power from resisting interpretation, delaying comprehension or forcing the reader to participate actively in constructing meaning. Eliot is difficult because he intended to be difficult. Dickinson leaves gaps because those gaps are part of the poem. Kafka often feels opaque because clarity would fundamentally change the emotional experience of reading him.

One detail makes the study even more intriguing. Whenever participants were explicitly told that a poem had been generated by AI, they rated it significantly lower, even when the text itself remained unchanged. The label alone altered their judgement. In other words, readers simultaneously demonstrated two contradictory biases: when they did not know the author, they often preferred the AI poems, but once they were informed that the author was a machine, they immediately became more critical.

There is also an important historical footnote that deserves more attention than it has received. Every poem in the experiment was generated by ChatGPT 3.5, a model that now belongs to what feels like a previous geological era of artificial intelligence. GPT-3.5 was released in late 2022, and while it represented a remarkable breakthrough at the time, it has since been surpassed by several generations of models that possess dramatically stronger reasoning abilities, much larger context windows and a far more sophisticated command of style and tone. If ordinary readers were already unable to distinguish GPT-3.5 from celebrated poets, it is difficult to imagine that repeating exactly the same experiment today with a modern frontier model would make the task any easier. Whether the results would be identical is an empirical question, but it seems unlikely that the newer models would perform worse than the one used in the study.

For us, this is where the paper becomes directly relevant to literary translation. One of the greatest strengths of modern language models is their ability to produce prose that reads effortlessly, smoothing awkward constructions, clarifying ambiguous passages and resolving stylistic tension almost instinctively. Those capabilities are enormously valuable, but they also create a subtle risk. Literature is full of passages that are intentionally awkward, ambiguous or resistant because those qualities are part of the author's artistic intention. A translation that automatically removes every rough edge may become easier to read while becoming less faithful to the original work.

Perhaps that is the real lesson hidden inside this study. The fact that AI can produce poetry that many readers enjoy is undeniably impressive. The more difficult question is whether immediate readability should be the metric by which we judge literature at all. If we begin to reward every text for being transparent, emotionally explicit and frictionless, we may gradually teach our models, and eventually ourselves, to iron out precisely those features that have made literature worth returning to for centuries.

Sources

• AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably

Faithful to what?

Nabokov, Benjamin, and the Biggest Question in AI Translation

What does fidelity really mean when translating literature—and what should we ask of AI when there is no single right answer?

When people think about translation, they often imagine a simple spectrum whose two extremes are literal fidelity and free adaptation. It is an intuitive way of framing the problem, but one that begins to unravel as soon as we turn to some of the twentieth century's most influential thinkers on translation.

Vladimir Nabokov is perhaps the most famous advocate of uncompromising fidelity to the original text. For him, the translator's duty was to preserve every nuance, every ambiguity, every peculiarity of the source, even when doing so produced a translation that sounded unfamiliar, awkward, or resistant to the expectations of the target language. His monumental translation of Eugene Onegin, accompanied by hundreds of pages of philological notes, embodies this philosophy better than any theoretical essay. As he famously wrote:

The clumsiest literal translation is a thousand times more useful than the prettiest paraphrase.

For Nabokov, elegance was always secondary to precision.

Walter Benjamin approached the question from an entirely different direction. In The Task of the Translator, he argues that a translation is not simply a reproduction of an original work for readers who cannot access it, but a new stage in the work's existence, an event that reveals relationships between languages which remain invisible in the original alone. His celebrated formulation captures this idea:

A translation issues from the original, not so much from its life as from its afterlife.

Every genuine translation extends the life of a text, allowing it to unfold in ways that neither language could achieve independently.

At first glance these positions appear irreconcilable. Nabokov seems obsessed with the text itself, Benjamin with the metaphysical destiny of language. Yet beneath their differences lies a striking point of agreement: neither believes that the translator's highest goal is simply to produce a version that reads as though it had originally been written in the target language. For Nabokov, the foreignness of the original deserves to remain perceptible; for Benjamin, the translation should preserve the productive tension between languages, allowing the original to resonate within the new text rather than disappear behind it.

Their ideas remain surprisingly relevant in the age of artificial intelligence. For decades, machine translation systems have largely been optimized for fluency, smoothing away ambiguities, regularizing syntax, and normalizing stylistic irregularities in order to produce text that feels immediately natural to its readers. While this approach works remarkably well for many practical purposes, literary translation often demands precisely the opposite. An ambiguity may be deliberate, a repeated word may carry structural significance, and an unusual rhythm may express something that cannot be paraphrased without loss.

This understanding shaped the way we designed TranswrAIte. Rather than treating AI as an automatic replacement for the translator, we built it as a tool whose behaviour can be guided according to different philosophies of translation. The translator can control the model's degree of freedom, provide stylistic instructions, build a translation guide that preserves recurring choices, and review every suggestion before publication, because no single approach is universally right.

Some books call for Nabokov's relentless fidelity to every detail, while others invite Benjamin's vision of translation as the continuing life of a work in another language. The translator's craft lies in recognizing which kind of faithfulness a particular text requires, and AI should help make that decision possible, not make it on the translator's behalf.