‘Being human helps’: des­pite rise of AI is there still hope for Europe’s trans­lat­ors?


Philip Olter­mann European cul­ture editor
9 May 2026
In Feb­ru­ary 2022, while he was plug­ging away at ren­der­ing the US writer Dana Spi­otta’s novel Way­ward into French, the lit­er­ary trans­lator Yoann Gentric decided he needed a bit of light relief. He would test whether AI could put him out of work.


Gentric had been grap­pling with a short non-verbal sen­tence that described the book’s prot­ag­on­ist’s feel­ings upon open­ing a win­dow: “Bright, sharp night air, bra­cing.” He put the prompt into DeepL, a neur­al­net­work-powered machine trans­la­tion engine that reg­u­larly out­per­forms Google Trans­late in accur­acy assess­ments.


The pro­posed trans­la­tion was reas­sur­ing, with his job secur­ity in mind: L’air de la nuit, vif et vif, était viv­i­fi­ant (The night air, lively and lively, was enliven­ing.) AI had trans­lated the sen­tence’s mean­ing but was seem­ingly unaware that the repe­ti­tions rendered the line absurd. It was far inferior to his own trans­la­tion that would be pub­lished in the book a year later: L’air pur et piquant de la nuit, viv­i­fi­ant.


When Gentric repeated his exper­i­ment this spring, however, the out­come made him feel less at ease: L’air noc­turne était vif, pur et viv­i­fi­ant, DeepL sug­ges­ted this time. The online trans­lator still lost the sen­tence’s styl­istic trait by adding a verb, but it had learned to use three dif­fer­ent words that even had a musical ring to them. “I don’t know if it’s just chance or a fine-tuned algorithm at work, but noc­turne and pur is not bad,” said Gentric.


Chat­bots run­ning on large lan­guage mod­els (LLMs) - neural net­works trained on vast amounts of text to gen­er­ate nat­ur­al­sound­ing lan­guage - are rap­idly infilt­rat­ing every aspect of our work and leis­ure lives. But few pro­fes­sional sec­tors are being dis­rup­ted by the tech­no­logy as rap­idly as the trans­la­tion industry in Europe, home to more than 200 lan­guages and a boom­ing tech sec­tor.


Accord­ing to a recent joint sur­vey by the French authors’ soci­et­ies ADAGP and the Société des Gens de Lettres, 79% of trans­lat­ors believe the rise of AI “poses a threat of repla­cing all or part of their work”. In Bri­tain, a 2025 sur­vey found that 84% of trans­lat­ors ques­tioned expec­ted lower demand for human trans­la­tion, res­ult­ing in lower pay.


Those fears con­cern the future, but for many trans­lat­ors the nature of their work has already changed. Laura Radosh, a Ber­lin­based Ger­man-to-Eng­lish


trans­lator, used to get about four job requests per month from cli­ents includ­ing uni­versit­ies, pro­fess­ors and museums. Last year, the num­ber of offers dropped to one each month.


Many of them were “pos­ted­it­ing” jobs, which required her to cor­rect texts that had already been run through a machinetrans­la­tion engine. “Post-edit­ing took me as much time as trans­lat­ing from scratch,” said Radosh.


Far less cre­at­ively ful­filling than trans­lat­ing from scratch, post-edit­ing is also less well-paid:


usu­ally com­pensated by the hour rather than by the page or by the book, it is paid “at unac­cept­able rates con­sid­er­ing the work involved”, accord­ing to the French trans­lat­ors’ asso­ci­ation. In Ger­many, pub­lish­ers have been found to offer typ­ical rates of two to eight euros per page - a quarter of the aver­age pay for trans­lat­ing a page from scratch.


But rates for reg­u­lar tech­nical trans­la­tions have tumbled too. “I got offered a job at 60 cent[s] a line,” said Radosh. “Before then, 80 cent[s] was the low­est rate I had ever come across.”


Even before the advent of LLMs, trans­la­tion was a pre­cari­ous pro­fes­sion: a recent sur­vey by the Ger­man trans­lat­ors asso­ci­ation VdÜ found that the aver­age income for lit­er­ary trans­lat­ors - tra­di­tion­ally at the lower-paid end of the sec­tor - was as little as €20,363 euros per annum before tax. But the latest changes in the industry mean that for many trans­lat­ors, the num­bers no longer stack up - Radosh recently took a part-time job doing book-keep­ing for an NGO.


Marco Trombetti, the cofounder and CEO of the machine trans­la­tion com­pany Trans­lated, said: “Without help, the human brain basic­ally is able to pro­duce about 3,000 words a day of trans­la­tion. Begin­ners will man­age about 1,500, the best trans­lator in the world may man­age 6,000, but the vari­ation is not that big.”


The cost of human trans­la­tion, he argued, had until now been defined by the num­ber of neur­ons we have in the brain. “That’s around 100bn,” Trombetti said. “But if we change that, then we change the unit eco­nom­ics of trans­la­tion.”


Yet the speed of tech­no­lo­gical change is also reveal­ing what human trans­lat­ors still do best.


For one, many machine trans­lat­ors still struggle with con­text. The Ger­man-Brit­ish aca­demic pub­lisher Springer Nature offers its authors the option to have their books auto­trans­lated into other lan­guages for free, but in spite of assur­ances of sub­sequent “human checks”, this pro­cess has in the past led to com­ical res­ults.


In 2024, Springer Nature machine-trans­lated into Ger­man an Eng­lish-lan­guage book by a group of Indian aca­dem­ics called ‘Cap­ital’ in the East: Reflec­tions on Marx. In the chapter head­ings, however, the machine trans­lator DeepL had rendered “cap­ital” not as Kapital in the inten­ded sense, but Hauptstadt, mean­ing “cap­ital city”.


A spokes­per­son for Springer Nature said in a state­ment: “Our AI-sup­por­ted trans­la­tion is human-led and reviewed by pro­fes­sional edit­ors. Errors like this are rare and regret­table, and this instance relates to a lim­ited pilot that has since ended.”


Jörn Cam­bre­leng, the dir­ector of Atlas, a French organ­isa­tion pro­mot­ing lit­er­ary trans­la­tion, said: “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.”


One of the iron­ies of the upheaval is that lit­er­ary trans­la­tion now appears to be a com­par­at­ively safer career choice than its tech­nical coun­ter­part.


The Har­per­Collins-owned imprint Har­le­quin France has con­firmed that it is work­ing with a French com­mu­nic­a­tions agency, Flu­ent Planet, to pro­duce trans­la­tions that are gen­er­ated by AI soft­ware and then post-edited by humans, although for now such trial runs are con­fined to the pul­pier reaches of the mar­ket:


Har­le­quin’s titles include A Mis­tress’s Con­fes­sion and The Embrace of a Prince.


In Ger­many, where the total num­ber of new pub­lished books has been gradu­ally declin­ing year on year, lit­er­at­ure in trans­la­tion has held up remark­ably well, with 8,765 books in trans­la­tion pub­lished in 2024 mak­ing up a his­tor­ic­ally high 15% of the over­all out­put. Increas­ingly, authors are also con­trac­tu­ally obli­ging their pub­lish­ers not to use AI in the trans­la­tion pro­cess, said Mar­ieke Heim­burger, a Dan­ish-to-Ger­man trans­lator who chairs VdÜ.


“AI really can­not do dia­logue,” said Katy Derby­shire, a Ber­lin­based trans­lator who has rendered into Eng­lish nov­els by Clem­ens Meyer, Christa Wolf and oth­ers. “When you are trans­lat­ing from scratch, you learn to under­stand the char­ac­ters and their motiv­a­tions, and you’re con­stantly adjust­ing them in your head - to indi­vidual situ­ations, but also to genre. The dia­logue that AI came up with just didn’t suit the char­ac­ter descrip­tion at all.”


Being human helped the trans­la­tion pro­cess, she added. “My body has exper­i­enced all the pain and the joy that lit­er­at­ure strives to con­vey. I under­stand what someone might scream when they hit their toe on the bed frame - an algorithm doesn’t.”


Fernando Pri­eto Ramos, of the Uni­versity of Geneva’s fac­ulty of trans­la­tion and inter­pret­ing, said his centre had noticed a drop in applic­a­tions to trans­la­tion courses three years ago, when the rise of gen­er­at­ive AI fuelled the hype around machine trans­la­tion. “But the trend is gradu­ally revert­ing again with a more diver­si­fied train­ing offer,” he said.


Even people who develop machine trans­la­tion soft­ware con­cede there are tasks that remain bey­ond their product’s reach. “If in Italian I say Solo tre parole: non sei solo, then a lit­eral trans­la­tion into Eng­lish would be ‘Just three words: you are not alone,’” said Trombetti, who foun­ded Trans­lated in 1999. “But you’ve ended up with four words, not three. That’s something that machine trans­la­tion still struggles with.”


Heim­burger said: “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.”


https://www.pressreader.com/usa/the-guardian-usa/20260509/282501485254320
#metaglossia
#metaglossia_mundus