My body has experienced all the pain and the joy that literature strives to convey
Across Europe, a profession built on the belief that language is more than information—that it carries voice, memory, and the weight of lived experience—is being tested by machines that can approximate meaning but not yet inhabit it. Translators, already among the most modestly compensated of literary workers, are watching rates fall and commissions dry up as AI systems absorb the technical middle of their trade. Yet the very failures of these systems—mistaking a political concept for a city, flattening a character's voice into generic prose—are quietly redrawing the boundary between what can be automated and what remains irreducibly human. The question is not whether machines will translate, but whether the culture that commissions translation will remember why the difference matters.
- 79% of European translators fear AI will replace their work, and the fear is already materializing: one Berlin translator's monthly job offers fell from four to one, with the remainder paying as little as two euros per page for machine post-editing.
- Rates have collapsed from 80 to 60 cents per line, pushing some translators into unrelated work—one took a bookkeeping job at a nonprofit just to stay solvent.
- AI's failures remain vivid and instructive: DeepL once rendered 'Capital in the East' as a book about capital cities, and translators argue that no algorithm can yet capture the way a character speaks, swears, or suffers.
- Literary translation—historically the worst-paid corner of the profession—is emerging as an unexpected refuge, with publishers contractually banning AI and translated literature reaching a record 15% of new German titles in 2024.
- University enrollment in translation programs is recovering after an AI-hype-driven dip, and veteran translators draw a pointed distinction: they fear not the machine, but the client who cannot tell the difference.
In early 2022, French translator Yoann Gentric paused mid-novel to test DeepL on a single lyrical sentence from Dana Spiotta's Wayward. The machine produced something technically coherent but artistically clumsy—repeating the same word where the original varied. His own rendering was clearly superior. When he ran the same test in spring 2026, the machine had improved enough to unsettle him. It still wasn't right, but it was closer.
That unease is now widespread. A joint French survey found 79% of translators believe AI poses a genuine threat to their livelihoods. In Britain, 84% expected demand for human translation to fall. For Berlin-based translator Laura Radosh, the disruption arrived not as a forecast but as a fact: monthly job offers dropped from four to one, and the work that remained was often post-editing—correcting machine output for a fraction of the pay. She eventually took a part-time bookkeeping job. The economics had simply stopped working.
The profession was already fragile. Literary translators in Germany earned an average of just €20,363 annually before AI accelerated the pressure. Yet the same technology exposing their vulnerability is also revealing its own limits. When academic publisher Springer Nature offered free AI translation of scholarly books, DeepL rendered a text about Marxist economics with 'capital' translated as 'capital city,' turning chapter headings into nonsense. The pilot program quietly ended.
Translators argue that what machines miss is not vocabulary but interiority. Katy Derbyshire, who has translated Clemens Meyer and Christa Wolf into English, put it simply: 'AI really cannot do dialogue.' Characters have voices shaped by history, motivation, and pain—qualities an algorithm cannot inhabit. Jörn Cambreleng of the French literary translation organization Atlas framed it as a structural difference: machines are built to produce sentences that sound familiar, while good translators strive to say something that has never quite been said before.
This distinction has produced an unexpected reversal. Literary translation, long the profession's poorest-paid tier, now looks comparatively safer than technical work. Publishers are writing AI prohibitions into contracts, and translated literature accounted for a record 15% of new German titles in 2024. University applications to translation programs, which dipped when generative AI hype peaked, are beginning to recover.
The machines are improving—everyone in the field acknowledges that. But as one veteran translator put it, the real fear is not the algorithm. It is the client who has already decided the algorithm is good enough.
In February 2022, Yoann Gentric was deep in the work of translating Dana Spiotta's novel Wayward into French when he decided to run a quick experiment. He had been wrestling with a single sentence—a description of how the book's main character felt when opening a window: "Bright, sharp night air, bracing." He fed it into DeepL, a neural machine translation system that often outperforms Google Translate in accuracy tests, just to see what would happen.
The result was oddly reassuring. DeepL produced: "L'air de la nuit, vif et vif, était vivifiant"—technically correct in meaning, but absurdly repetitive, using the same word twice where the original used different ones. Gentric's own published translation was far superior: "L'air pur et piquant de la nuit, vivifiant." The machine had grasped the content but missed the artistry entirely. When he tried the same sentence again in spring 2026, however, the machine had learned. This time it offered: "L'air nocturne était vif, pur et vivifiant." Still not perfect—it had added a verb the original lacked—but noticeably better, with a musical quality that suggested the algorithm was improving in ways that made Gentric uneasy.
Across Europe, translators are experiencing that same creeping anxiety. A joint survey by French authors' societies found that 79 percent of translators believe artificial intelligence poses a genuine threat to their livelihoods. In Britain, 84 percent of translators surveyed in 2025 expected demand for human translation to fall, bringing their pay down with it. The disruption is not theoretical. Laura Radosh, a German-to-English translator based in Berlin, used to receive about four job offers each month from universities, professors, and museums. Last year, that number dropped to one. Many of the remaining offers were for "post-editing"—correcting texts that had already been run through a machine translator. The work consumed as much time as translating from scratch but paid a quarter as much, sometimes as little as two to eight euros per page. Radosh was offered a technical translation job at sixty cents per line, down from the eighty cents that had once been the floor. The economics no longer worked. She took a part-time job doing bookkeeping for a nonprofit.
Translation was already precarious before the arrival of large language models. A German translators association survey found that literary translators—traditionally the lowest-paid segment of the profession—earned an average of just €20,363 annually before tax. The new wave of AI has accelerated an already difficult situation. Yet the same technology that threatens translators' income is also revealing what machines still cannot do. When the academic publisher Springer Nature offered free machine translation of authors' books into other languages, the results were sometimes comical. In 2024, DeepL translated an English-language book by Indian academics titled "Capital in the East: Reflections on Marx" into German, but rendered "capital" not as Kapital—the intended meaning—but as Hauptstadt, meaning "capital city." The chapter headings became nonsensical. Springer Nature acknowledged the error as rare and regrettable, noting that the pilot program had since ended.
Machine translation struggles with context, with dialogue, with the thousand small choices that distinguish a living voice from a generic one. Katy Derbyshire, a Berlin-based translator who has rendered novels by Clemens Meyer and Christa Wolf into English, put it plainly: "AI really cannot do dialogue." When translating from scratch, a human translator learns to understand characters, their motivations, their speech patterns, adjusting constantly for individual situations and genre. The dialogue that AI produces simply does not fit the character. "My body has experienced all the pain and the joy that literature strives to convey," Derbyshire said. "I understand what someone might scream when they hit their toe on the bed frame—an algorithm doesn't."
Jörn Cambreleng, director of Atlas, a French organization promoting literary translation, framed the distinction sharply: "Machine translation is not creative. These systems are built to produce sentences that are generic, sentences that have been said before or sound like they have been said before. Whereas good human translators strive to put into words something that has never been said before." This distinction has created an unexpected irony. Literary translation, long the poorest-paid corner of the profession, now appears comparatively safer than technical translation. Publishers are beginning to contractually forbid the use of AI in translation, and in Germany, where the total number of new published books has been declining year after year, literature in translation held up remarkably well in 2024, accounting for 8,765 titles—a historically high 15 percent of all new books published.
Even the engineers building machine translation systems acknowledge their limits. Marco Trombetti, co-founder of the translation company Translated, offered an example: in Italian, the phrase "Solo tre parole: non sei solo" means "Just three words: you are not alone." A literal English translation produces four words, not three. "That's something that machine translation still struggles with," Trombetti said. The human brain, he noted, can produce roughly 3,000 words of translation per day—beginners manage 1,500, the best translators perhaps 6,000. The variation is small because it is constrained by neurology. But if machines can change that equation, they change the entire economics of the profession.
There are signs of stabilization. Applications to translation courses dropped three years ago when generative AI hype peaked, but Fernando Prieto Ramos, of the University of Geneva's faculty of translation and interpreting, said the trend is gradually reversing with more diversified training offerings. Marieke Heimburger, a Danish-to-German translator and chair of the German translators association, expressed a distinction worth holding: "I am not really scared of AI, because I know it cannot do what I can do. What I am afraid of is the people who think that AI can do my job." For now, that fear remains grounded in reality. The machines are improving, but they are not yet translators.
Notable Quotes
Machine translation is not creative. These systems are built to produce sentences that are generic, sentences that have been said before or sound like they have been said before. Whereas good human translators strive to put into words something that has never been said before.— Jörn Cambreleng, director of Atlas, a French literary translation organization
I am not really scared of AI, because I know it cannot do what I can do. What I am afraid of is the people who think that AI can do my job.— Marieke Heimburger, Danish-to-German translator and chair of the German translators association
The Hearth Conversation Another angle on the story
When Gentric tested DeepL on that sentence about the window, what was he really looking for?
Reassurance, I think. He wanted to know if the machine could do what he does. The answer was no—but only barely. That's what made him uneasy the second time around.
Why does literary translation seem safer than technical translation right now?
Because a technical manual just needs to convey information accurately. A novel needs to convey something that has never been said before. The machine is very good at the first thing. It's nowhere near the second.
Derbyshire said her body has experienced pain and joy. What does that actually mean for translation work?
It means she understands what a scream sounds like, what a gesture means, what a silence contains. She's lived through the human experience the text is trying to capture. An algorithm has not.
But machines are getting better. Gentric saw that improvement in just four years.
Yes. And that's exactly why translators are afraid. Not because machines are perfect now, but because they're learning. The question is whether they'll learn fast enough to make human translators economically irrelevant before they learn to do the work well.
Is there a way translators survive this?
Some are already contracting with publishers to forbid AI use. Others are specializing in literary work where the machine still fails. But the real survival depends on whether people value what a human translator does—whether they're willing to pay for it when a cheaper machine alternative exists.
And are they?
That's the question nobody can answer yet.