Can ChatGPT Produce a Version of Proust Worth Reading?

Can ChatGPT Produce a Version of Proust Worth Reading? Bryan Charles on the Slow But Steady Encroachment of Artificial Intelligence Into Literary Translation “The part of ourselves that matters, when it matters, is outside time… Think of yourself simply as an instrument capable of whatever experiments in beauty or truth you wish to perform, and your gloom will evaporate.” –Marcel Proust * A while back I watched, with mounting perplexity, a webinar entitled Using CAT Tools in a Literary (Re-)Translation Project: A Case Study of Marcel’s Proust’s Novel ‘La Prisonniére’. The event was hosted by the Institute of Translation and Interpreting, a UK-based organization comprising professionals in those fields, and the featured speaker was Andrew Rothwell, whose translation of the fifth volume of À la recherche du temps perdu will be published later this year as part of Oxford World’s Classics new take on the novel. I confess to a degree of discomfort going in. I did not know what a “CAT tool” was—I leaned it meant “computer-assisted translation.” There was also mention of something called Wordscope, whose website blared Boost your productivity and increase your revenue! Not a sentiment I associate with Proust. Lower on the page were references to ChatGPT and “neural machine translation.” In what I believed was a dark harbinger, Rothwell evidently had used these in his translation of La Prisonnière. This is not the place for a sweeping review of my various antipathies and anxieties concerning the technologies that define our age, which date at least to the Friendster days; for now let it suffice to report that I have never owned a smartphone, and continue to bristle at—indeed, sometimes, impotently to rage against—the numberless horrors that device alone has unleashed. Nevertheless, despite my misgivings, I signed up for the webinar, anticipating a discussion of Proustian translation. That did not materialize, exactly. Rather, the webinar was mostly a technical affair whose function was to demonstrate Wordscope’s capabilities as “an all-in-one solution for professional translators.” Of course Proust was mentioned, along with his original English translator, C.K. Scott Moncrieff, and a couple of passages from La Prisonnière appeared, but these were incidental. The focus was the platform. Certainly this would go faster if I used a translation website or a chatbot. But I do not care about speed, or ease, and never have, really, with regard to reading or writing. Rothwell elucidated its functionality in a series of slides. In brief, Wordscope’s appeal for him was that it significantly streamlined the process by which texts, including in this case the original French and Moncrieff’s “legacy” translation, could be compared with each other, and the suggestions of multiple machine-translation engines like DeepL, Google Translate, and ChatGPT explored. Whereas earlier workstation-based CAT tools—and Rothwell had been using them for some years—necessitated cutting and pasting, tediously switching between screens, and were generally quite fiddly, Wordscope uncluttered the whole works, presenting source and target texts and CAT tool proposals in a single smooth interface. (Rothwell had in fact completed his La Prisonnière translation before learning about Wordscope, and subsequently used it to rework the first section of the novel.) The platform had, in his view, reduced the “cognitive load” of literary translation, resulting in a process akin to “collaborative drafting.” He referenced other thinkers on the subject, including Arle Lommel, who coined the term “augmented translator,” and Tong King Lee, whose paper on “posthumanist translation” argues for the concept of “distributed cognition” amid exponential recent advancements in artificial intelligence, conceiving of AI as “a prosthesis of translators’ minds.” The final slide offered webinar viewers a 40% discount on a Wordscope subscription. * The sequence of billboards that greets drivers entering San Francisco from the Bay Bridge along Interstate 80 is liable for a certain sort of person to elicit responses ranging from numb detachment at the state of things to heart-in-the-bowels, head-clutching despair. For as long as I have lived in the area, these billboards have advertised technology products, many of whose functions cannot be grasped from the copy, or whose vaunted innovations seem at best underwhelming, if not superfluous (we are making the office phone obsolete). And then one day, as if by city decree, they all featured AI products whose purpose, in keeping with the nebulous lexicon favored by technologists, frequently could not be discerned. There was at least one stark outlier—a company whose ads featured soullessly staring AI “employees” and, among other repulsive slogans, the message Stop Hiring Humans. Suddenly AI is everywhere, answering search-engine queries, composing one’s emails, helping one’s daughter’s friend’s stepfather explore spiritual matters in an eagerly shared ChatGPT-generated poem-thing (it is possible this last one applies only to me)—whether one wanted it or not. Meanwhile, as ever, or at least since In Search of Lost Time became the life preserver I had clung to in the heaving sea of middle-aged failure, I was immersed in my Proust studies, comparing the different translations, reading the Oxford version as each volume came out. And while not a daily consideration, it is true that I saw the world I was lost in, the historical periods of the novel, of course, but also the very act of reading it, savoring sentences that demanded one’s full attention, that could not be raced through, and the encyclopedic information many of the dense paragraphs contained—I believed all of this to be, apropos of the novel’s themes, occurring outside of time somehow, untouched by contemporary grotesqueries, the imperatives toward automated summary and speed. I had interviewed Charlotte Mandell, whose translation of volume two of the Oxford Proust, In the Shadow of Girls in Blossom, moved me greatly, and asked if she felt threatened by machine translation. She did not. “No AI thing will ever be able to translate Proust,” she said. And now an AI thing—or, I guess, multiple AI things under the rubric of a frictionless, all-in-one “translator’s assistant”—had helped to translate Proust. And…was this cool? I felt queasy about it—should I? Surely Oxford World’s Classics would not entrust Proust to any random machine-curious translator with DeepL Pro? I had never so much as glanced at ChatGPT, or any similar product, but of course knew of their propensities for churning out trash and…well, there must be no place for these abominations in association with À la recherche du temps perdu. Right? * Prior to our conversation, Andrew Rothwell, maybe hoping to temper my dreamier instincts on the subject, told me that he was not a Proust specialist. But I learned that he had read La Recherche in French in his undergraduate days (I am predisposed to like anyone who finishes the whole novel, in any language). He had studied French and German at Oxford, and always had something of a side hustle as a translator, later specializing in the poet Bernard Nöel. Later still, as an educator at Leeds University, he cocreated a master’s program in professional translation, and did so again at Swansea University, where he has taught since 1999. And machine translation was always in the equation. “We were aware that the translation memory tools were beginning to come on the scene, in the early ’90s,” he said of his Leeds days when we talked over Zoom. “Our program started in 1995, and from the start we had a module on translation tools. And I’ve been involved with and interested in translation tools ever since.” I was mildly surprised; I had not realized the technology went that far back. Actually, Rothwell explained, the history of machine translation dates back much further. “In the post-Second World War period, there was great optimism that it would be possible to automate translation,” he said. “IBM and the American government got very interested in doing this, particularly with reference to Russian, for obvious reasons.” Research in the field went on for years, notably featuring 1954’s Georgetown-IBM experiment, a public demonstration of machine translation in which just over sixty sentences were converted from Russian into English (“Robot translates nimbly,” proclaimed the Christian Science Monitor), until funding was cut in 1966 in the aftermath of a government report issued by the Automatic Language Processing Advisory Committee (ALPAC). In subsequent decades, translation memory emerged, whereby a completed translation is stored in a database, and then used again when similar sentences or phrases appear later, so that a translator is never starting from scratch. TM has widespread commercial applications, in formulaic documents—financial reports or instruction manuals, for instance. It is meant to save time and achieve assembly-line consistency. “And of course that’s completely antithetical to literature,” Rothwell said. And yet. In his methodology, C.K. Scott Moncrieff’s celebrated rendering of a famously demanding text—Moncrieff’s version of La Prisonnière appeared in 1929 as The Captive, a title Rothwell keeps, whereas the Penguin Proust went with The Prisoner—is used as a translation memory, atomized into segments that appear in the translation tool alongside the original French. “With the Proust translation it struck me that it would be very interesting to see what the historical translator decided to do with each of the sentences I was translating. And the tool allows that—as you move through the job, sentence by sentence, you see Scott Moncrieff’s solution. Obviously his translation is a carefully considered reading of the text, which is rooted in its time, and the values of its time, to an interesting degree. You then can see why Scott Moncrieff doesn’t work now, why it’s not the voice I want to give it.” Of course, to the right of every Moncrieff-translated sentence were “proposals” generated by a suite of machine tools. “And you can just pick whichever one you think is closest to where you want to end up,” Rothwell said. I had appreciated hearing a bit about the history of machine translation, of which I knew nothing, and admit the perspective it granted was broadening. But Rothwell clearly was cognizant of the profound unease swirling around the subject now, certainly within his field. After all, he had opened his webinar with several slides referencing a 2024 survey by the European Council of Literary Translators’ Associations whose respondents held overwhelmingly negative views of machine translation and artificial intelligence. Rothwell, I gathered, was an early adopter: his first full-length translation, Zola’s Therèse Raquin, for Oxford World’s Classics in the late 1980s, was undertaken with another controversial tool. “Currently literary translators are suspicious of translation technologies,” he said. “But in those days, they were often suspicious of word processors. So there were literally people who seriously said, it cramps my style to use a computer, I’m much better with a pen.” Apart from the computer, his working method at the time was traditional. “I’ve got the book sitting on the desk and a ruler moving down the page to try and make sure I don’t miss anything. Typing line by line with dictionaries all around.” Decades later, however, working on Zola’s La joie de vivre—Rothwell’s title was The Bright Side of Life, after a Monty Python song—also for Oxford, the ruler and desk cluttered with dictionaries were a memory. The book, which was published in 2018, was translated using computerized tools, a fact not mentioned in the Translator’s Note, nor will such an acknowledgement appear in The Captive. Rothwell told me that while he has given research presentations on his use of translation tools, he never explicitly mentioned his methodology to Oxford. “I assume they’ll take the text that I send them and it’ll be reviewed academically to see whether it’s up to standard. But how it’s originated really doesn’t matter, I don’t think.” As far as Proust goes, there is a style guide to ensure consistency across the volumes. (Adam Watt and Brian Nelson, editors of the new translation—Nelson will also translate three of the seven volumes himself, probably the closest we will get to having a single person tackle the whole novel; Moncrieff only lived long enough to do six—wanted a more colloquial take, a lighter touch, one of whose key features appears to be contracted forms.) And in the end, Rothwell emphasized, it is he who determines the outcome. Throughout the drafting stage, he is in control—does not have to accept a particular machine proposal—and he always does final revisions outside of a CAT tool, in a two-column Word table with the French original and his own translation, the text now in flowing paragraphs down the screen rather than appearing as thousands of isolated sentences. In Rothwell’s reckoning, which is based on his years of research on the subject and thus reads differently from the deranged certainties spewed forth by anyone with a financial stake in AI—not to mention any number of casual users enchanted by the prospect of further removing the already rare act of cracking a book from the their lives—machine translation might best be understood as a collaborative tool that easily dispatches grunt work, allowing the literary translator to focus on higher-level decisions. “There’s a kind of eighteenth- or nineteenth-century romantic sense of the literary translator as someone working on the basis of inspiration,” he said, “that it comes somehow from the depths of your being, and the fertility of your creativity will produce something of value in its own right. And I just don’t see that there’s enough in my own mind, compared to absolute gigabytes of stuff that can be searched instantaneously to come up with interesting ideas and propose them.” * After the webinar I received a follow-up email from Wordscope imparting the company’s hope that the session had given me “food for thought regarding what AI and CAT tools can do for Literary Translation,” and again mentioning the 40% discount in the event I wanted to “explore Wordscope a bit further.” And I admit I was tempted. As it happened, I was hung up on several passages in Oxford’s The Guermantes Way, the third volume of the Search, translated by Peter Bush. To state just one quibble, I was bothered by the repeated use of the word “supper” in the narrator’s recounting of his ill-fated rendezvous with Madame de Stermaria—it sounded like something my stepfather might say (not long before, the latter had told me, seething, that I “ain’t worth a shit as a son” and never to “stick my fucking nose in [his] house again,” so I will allow there could have been other factors at play in my resistance to this single word than a desire for stricter fidelity in the translations of dîner and dîné). What would the machines say? I could drop this and other bits into Wordscope and— No. The drive fell apart on me. Another account? Another password? Clicking another box indicating that I had read another set of “terms” I had actually ignored? Mucking about with another technology product whose highest goal was to boost efficiency and increase revenue? Guess I was not that curious. And while I had enjoyed talking to Andrew Rothwell, it was not at a personal thing, our conversation continued to discomfit me for reasons I could not define. Why was I still so bugged out about translation tools? Maybe there was someone else I could bring into the conversation—whom? It was funny, in alluding to the notion of a literary translator as a romantic figure toiling away, digging deep to produce a work of enduring value, Rothwell had come close to crystalizing my own (possibly naive) view of the craft, and my attitude toward making art generally, I suppose. It is in fact a key tenet of the Search, whose seven volumes are both a prelude to and the consequence of the narrator realizing his life’s calling: producing the “essential book, the only true book, [which is] not something the writer needs to invent, in the usual sense of the word, so much as translate, because it already exists within each of us.” In the midst of these reflections I read The Seventy-Five Folios, a collection of early drafts from In Search of Lost Time. The manuscripts had been discovered only in 2018, and an English translation published in 2023. One day, recumbent in bed like Proust himself, books and notebooks scattered over the comforter, postponing the day when I would begin to write this essay, assuming my inchoate thoughts on the subject of machine translation could even achieve such a form, my eye landed on the cover of The Seventy-Five Folios, which featured a striking image of Proust’s head bursting into a pink hawthorn blossom, and drifted toward the names at the bottom. Translated by Sam Taylor. And I knew what I would do next. * Sam Taylor told me that he had read Swann’s Way in the Moncrieff translation some years ago but none of the other books—”Because I didn’t like it that much”—until in preparation for translating The Seventy-Five Folios he went back and read the first three. His path to that point had been improbable. In London he fell in love with a French woman, and they moved to southwest France, where they raised their children and Taylor, who is also a fiction writer, left a career in journalism behind and wrote and published novels. He had never studied French before—picked it up simply living there—or received instruction in translation. Eventually, needing income, he got a gig writing book reports of French novels for English publishers. The first of these happened to be for Laurent Binet’s HHhH, a metatextual novel about the assassination of Nazi official Reinhard Heydrich. It was the best book Taylor had read in any language in about a decade. He wrote a rave report for Faber & Faber, who passed. But another publisher bought the novel, at which point Taylor asked if he could do the translation. He submitted a sample. And the publisher said yes. “That was the first time I had translated anything in my life,” Taylor said of Binet’s breakout. Since then—and it is not all that long a period, really; HHhH was published in English only in 2012, although Taylor had translated it two years prior—he has worked more or less full time as a translator, producing seventy-odd books in a range of styles and genres, while continuing to produce his own fiction (the most recent of his five novels, The Two Loves of Sophie Strom, was published in 2024). In a minor bit of symmetry I had just read another Taylor translation, Leïla Slimani’s propulsive domestic thriller The Perfect Nanny, using it and the original—whose title is Chanson douce—as teaching tools, since the French in the novel is within my grasp. “Some writers have a very French style, for want of a better word. And you have to almost take their sentences apart and put them together in a completely different order to make it work in English,” Taylor said. “Leïla Slimani, certainly in that book and I think her first book as well, which was translated as Adèle, it’s almost like it was written in English syntax in French. The sentences are so short and so direct. A bit like Camus, almost. It doesn’t read like a French book in lots of ways.” Nor does—and even I was able to grasp this, within two minutes of starting to read the novel in French—La Vérité sur l’Affaire Harry Quebert, Joël Dicker’s blockbuster thriller, known to English readers, also in Sam Taylor’s rendering, as The Truth About the Harry Quebert Affair. However, it also does not read like an American novel, despite being set here, opening in an alternate-universe Manhattan, in the depths of the Great Recession, where a wealthy and globally famous writer cannot leave his building without being accosted by zealous fans: “Le Tout-New York se passionnait pour mon livre ; il y avait deux semaines qu’il était paru et il promettait déjà d’être la meilleure vente de l’année sur le continent américain.” (I am all for escapism, but someone at Penguin, Dicker’s US publisher, should have had a quick word with him about the realities of the American literary landscape.) I asked Taylor, “How do you do your translations?” He replied, “This is like my least favorite question. Usually the question is, what’s your translation process? And it’s like, I don’t know—I just sit at the computer, read this book in French, and it comes out in English.” (This reminded me of a video I had watched in which Jodie Foster, whose facility with the language is renowned, was asked how she speaks in French: “How do I do it? I open up my mouth, and then it comes out.”) Fair enough. But what I wanted to know is what is Taylor looking at as he translates? How are the texts arranged? He told me that he works on a desktop computer with the French in a PDF on the left side of the screen and the Word document of his translation on the right. And he explicated his process after all: “I like to translate quickly, and intuitively. If I slow down too much I kind of lose my balance. So I usually translate in short bursts. I do like forty-five minutes or an hour in one sitting. I’ll translate a thousand or twelve hundred words in that time. And then I’ll go do something else for an hour. And then I’ll sit down again. I’ll do three or four sessions like that in a day. Most of my translation career, I’ve been translating four thousand words a day.” I prefer to read slowly, carefully. Puzzling over the language. Turning the pages. My fingers on paper. Outside of time. This pace is a carry-over from his time as a commercial translator, a period of approximately a year and a half, after HHhH but before he picked up other literary jobs. Taylor had to translate a huge amount of material then simply to meet deadlines, and he now finds it difficult to work any other way. His desktop computer is not connected to the Internet—but a laptop to the right is, and Taylor switches over to it for basic research, Wikipedia and the like, as well as deeper dives with ChatGPT. The latter has been helpful in researching his own novels and has eased the workload of recent translations—Rachid Benzine’s The Man Who Read Books, for instance, in which there are references to many Arabic books translated into French. Taylor needed to find out whether they had also been translated into English, or maybe only certain poems had, and to differentiate between various titles and editions. A complicated task that would have taken a lot of time before but was speedy with ChatGPT. Although as everyone knows, or should know by now, the information retrieved can be less than rock solid. Take his current project: Joseph Incardona’s La Soustraction des possibles. A brilliant novel but dense, again with a ton of references. There was a quote from a novel that Taylor had to look up. He asked ChatGPT, which produced the title immediately. But when Taylor searched for it, nothing appeared. Turned out the machine made it up. It explained that it was just trying to be quick. And apologized and said it would not happen again. “It’s like working with a brilliant but lazy child,” Taylor said. “You can’t leave it unsupervised, basically. But it’s still a useful tool.” Also useful are Google Translate and DeepL, which Taylor uses sparingly, finding they just slow him down. Still, if there is a thorny sentence, something really difficult to get one’s head round, or three or four words or expressions he does not understand, he will drop it into the translation websites, skim the results, and usually that will give him a better understanding of what it means. Usually. Sometimes the language is so… literary, so poetic, and the translation apps, while much better than they used to be, in Taylor’s estimation, still do it word for word, hewing closely to the French syntax. “I would never use the sentences that those websites give me as translations, but it does help in terms of understanding it more quickly than I used to be able to.” At this stage of the conversation I was feeling…better?—I guess? Lighter of spirit. I had been so distraught that my beloved Proust, whose sentences are among the most gorgeously written in any language (although I do not otherwise see myself as one of those “sentence” freaks catapulted into ecstasies strictly by seamless literary craft; if anything, Proust’s work as a whole defies almost everything…no, make that everything I was taught both as an undergraduate creative writing major and in my MFA courses)—that Proust had been subjected to what I perceived to be the indecencies of computer-assisted translation among other technological nightmares marketed as productivity- and revenue-boosting marvels. Yet here was Sam Taylor, himself a Proust translator—winner of the Scott Moncrieff Prize, no less!—shrugging, saying, yeah, he uses these tools (not, as Andrew Rothwell does, as a framework for the job, but still)—why wouldn’t you? It is possible that I had been a bit of a hothouse flower regarding the incursion of any technology more sophisticated than Microsoft Word (and of course even that has become infested with unasked-for AI “assistance”) into the field of translation. The question, though—one Ian Curtis bellowed decades ago—is: where will it end? “When I was a commercial translator,” Taylor said, “there were already a lot of agencies at the time that would only employ translators who used translation tools.” He suspects that many have now lost their jobs. “Some of what I was translating was ingredients on processed food packets. It was ridiculously repetitive. And really there’s no reason why a human needs to translate that. So that kind of stuff I am sure has been taken over by machines.” He alluded to the interview I had done with Charlotte Mandell and her quote about AI never being up to the task of literary translation. “I don’t know if that’s true or not. I mean, we’re like two years into the AI revolution, right? It’s very, very new. It’s improved massively in the past few years. I don’t think anybody has any idea what it’s going to be like in five or ten years.” Taylor paused, then went on, “Certainly what it lacks now is any kind of discernment. Right? It’s got every language on earth at its fingertips, but it has no aesthetic sense. The main part of a translation, in a way, is giving it the music in English—not the same music—but whatever flow, whatever music it had in the original language, to give it the equivalent flow and music in the target language. And to do that you’ve got to be able to write well in the target language. I honestly don’t know if machines will ever do that. But there are plenty of humans who aren’t that good at it either. I’ve read books that have won translation prizes that read to me like they were translated by machines.” He had a point. I had read such a translation a few months ago, another novel I read in both French and English. The translation had not won prizes but the book had been successful, although much more so in France. And the English version was utterly pedestrian. (To be fair, the French did not exude high style.) Nor does Taylor foresee an eventuality as reductive as machines translating popular fiction and humans doing highbrow literature, mostly because dialogue is hard to get right. “A lot of people can’t do it well, and I am sure that machines can’t,” he said. What the rise of artificial intelligence might do, Taylor allowed—assuming the technology evolves to the point where it can produce a full-length literary translation in the range of sufficient or good enough, basically—is thin the field of “lazy, cowardly translators.” Because if publishers can use a machine to churn out one of those dull translations that sticks to the word order and syntax of the original language, even if it reads poorly in the target language, and is entirely lacking the aforementioned flow and music—and as Taylor pointed out, plenty of people make those already—why pay someone to do it? AI may be fine for simple prose, he said, but it struggles with heavily allusive, referential, literary texts, also dialogue and humor, “because those are based much more on being human and alive, and interacting with other alive humans!” So when—if—you have AI knocking out those literal, stilted translations, the people who also work in that manner will no longer have a niche. Which will leave the job to “translators who have their own style, their own personality, and who are bold enough to do something original.” * I continue to receive promotional emails from Wordscope. In April the company hosted a series of webinars in conjunction with translation associations around the globe, on ChatGPT, workflows, and “customizable AI.” Earlier this week I received a fundraising appeal from Words Without Borders, the magazine devoted to literature in translation, whose first paragraph mentioned a 2025 study that found translators are “the group whose profession is most threatened by the rise of AI. Writers are not far behind, coming in third place.” And later today I will make a cup of coffee and withdraw for thirty minutes or so of solitary French study. At the moment I am reading Mon vrai nom est Elisabeth, Adèle Yon’s innovative recounting of her investigation into the disturbing treatment, in the guise of mental health “care,” of her great-grandmother. The French in the book is harder for me, but I enjoy the challenge. I read it slowly, at times aloud, jotting down words and phrases, frequently reaching for my French-English dictionary and Barron’s 501 French Verbs. Sometimes I will translate a sentence or two. Certainly this would go faster if I used a translation website or a chatbot. But I do not care about speed, or ease, and never have, really, with regard to reading or writing. I guess that is why I am not in the least curious about the many supposedly world-changing innovations I am told I that I must—actually am forced by certain corporations—to accept. No, I prefer to read slowly, carefully. Puzzling over the language. Turning the pages. My fingers on paper. Outside of time. Bryan Alistair Charles Bryan Alistair Charles is the author of a novel, a memoir, and a book about Pavement written for Bloomsbury's 33 1/3 series. He lives in California. Find him at bryanalistaircharles.com and Bluesky.