Did AI Find Murakami’s Voice? Student’s $2 Experiment Divides China’s Literati

A 21-year-old Chinese university student used an AI-based tool called Wenyi to translate Haruki Murakami's new novel 'The Tale of KAHO' into Mandarin within four hours of its July 3 release in Tokyo, costing just 13 yuan ($1.90), and the unlicensed translation quickly spread on Chinese social media and platforms like Xianyu and Z-Library. The translation, which used DeepSeek's AI model, drew praise from readers on Douban, but publishing insiders and legal experts caution that AI cannot replace human translators and that copyright holders may suffer.

FEATURES /features/26167/features?source=normal article Did AI Find Murakami’s Voice? Student’s $2 Experiment Divides China’s Literati Wang Juyi /users/1017265/wang-juyi?source=normal article Chinese readers of Haruki Murakami used to wait over a year for a faithful Mandarin translation of his latest book after its official release in Japanese. Not this time. An unlicensed, high-fidelity digital version of the author’s new novel, “The Tale of KAHO,” was available to download in Chinese just hours after it went on sale in Tokyo on July 3. There were no lengthy negotiations over rights, no translators straining to make a deadline, no pondering on how to capture Murakami’s distinct voice. Instead, it was accomplished in a relative flash for less than $2 using an AI-based tool developed by a 21-year-old university student. “The Tale of KAHO,” a novel about a picture book illustrator who slips between dream and reality, was the first real test for Wenyi, a coding project that instructs large-language-model artificial intelligence on literary translation. Its designer, who asked to be identified by the pseudonym Ping Ming for privacy reasons, says he wanted to know whether AI could handle not only simple text but also glossary management, polishing, and rounds of spontaneous self-reviews. What he discovered could fundamentally alter how novels are translated. Wenyi’s unlicensed translation of the novel quickly began circulating on Chinese social media and the resale platform Xianyu, as well as the free online archive Z-Library. Anecdotal evidence suggests readers are impressed. On China’s interest-oriented networking platform Douban, the most-liked review of the novel reads, “If this is purely AI translation, the quality is astonishing … If AI can already do this much, what’s the point of having human translators?” Readers might be happy, but publishing industry insiders feel AI still has a long way to go before it can replace humans — while legal experts warn that copyright holders could lose out either way. Staying faithful Nanshan Youlu, a part-time Japanese-to-Chinese translator who asked to use his online alias, worked with a friend on a manual translation of “The Tale of KAHO,” which they also began sharing bit by bit on the lifestyle platform Xiaohongshu, or RedNote, on July 3. Their first post received more than 2,000 likes, with many users expressing gratitude and complimenting the quality of their translation. Yet shortly after completing the final chapter, the duo wiped their version from the internet, citing that Wenyi’s translation was just too good — virtually 100% faithful to the original Japanese text. “When human translators work on a novel, they will read the whole book first, and therefore will build their own interpretation and create their own ‘context.’ It’s basically impossible for them to achieve 100% wording fidelity,” Nanshan Youlu, a 30-something working at a tech company in Beijing, tells Sixth Tone. “They don’t translate mechanically — every sentence they translate serves the context they build.” Wenyi works differently. Ping explains that his program started by dividing the book into chunks of fewer than 2,000 words, which it then sent one by one to the Chinese AI model DeepSeek for translation. To maintain continuity, each chunk overlapped by six paragraphs. “If you feed AI tens of thousands of words all at once, there’s a risk of AI hallucination, which can result in lots of mistakes,” he says. The program also asked AI to extract important terms and proper nouns from the material so it can continuously update a glossary of keywords to ensure consistency. Once finished, the system reviews the entire book — again, chunk by chunk — to check for problems. Not a single human had read the Chinese translation before Ping uploaded it to Z-Library just four hours after the official version of the novel was released in Japan. The project had cost him just 13 yuan $1.90 in DeepSeek tokens. However, once online, he soon began noticing mistakes. Some nouns, such as the various nicknames for the protagonist Kaho, hadn’t been extracted for the glossary and were inconsistent, while some wording was unnatural, more machine-like. He decided to translate it again, this time refining his code to instruct the AI to be more meticulous about names, and requesting a complete polish of the text. This second version was better, he says, although it took longer. Ping, a literature major at Nanjing Normal University in the eastern Jiangsu province, taught himself to code using the Python programming language two years ago. To train Wenyi, he used English-language literature including works by the 19th-century American writer Edgar Allan Poe and Scottish novelist Ian McEwan. “When I was a child, I always found hackers in the movies so cool, and I wanted to have those skills,” Ping tells Sixth Tone, adding that in more recent years he has been attracted by the rapid developments in AI technology. The idea for Wenyi partially stemmed from his passion for literature. As a fan of Murakami, Ping always wanted to read the author’s books as soon as they were released, but he can’t read Japanese. He recalls attempting to digest an unlicensed Mandarin version of Murakami’s novel “The City and Its Uncertain Walls” produced using the free DeepL online translation tool in 2023, “but it was unacceptably bad — the grammar, pronouns, proper nouns, all were a mess.” He had to wait another year for the official Chinese version, translated by the prominent literary scholar Shi Xiaowei. Machine translation has been around a while, but conventional tools have never been considered good enough for literary works. “Someone needs to step forward and explore how good AI translation can be,” says Ping. “If it were not me, it would be someone else. … The original AI models set the lower limit of translation , while Wenyi sets the upper limit.” However, not everyone is impressed. Lin Shaohua, who translated Murakami’s earlier works, including the best-sellers “Kafka on the Shore” and “Norwegian Wood,” told https://mp.weixin.qq.com/s/iR9oYomFhvyzlCAuL1xMKQ China Newsweek that while Wenyi’s translation was technically fluent, it had failed to capture the author’s true voice. “AI translation has achieved ‘grammatically coherent and reads smoothly’ — but that’s all,” he’s quoted as saying. “AI is still not quite capable of conveying subtle nuances of language and stylistic nuances. … The story itself is not literature. Only its particular tone makes it literature. Literature is the art of language.” However, over the years, Lin has faced criticism from readers over alleged mistakes in his translations that might have changed what the original author intended to convey. Nanshan Youlu believes this is hard to avoid and sees translation as a continuation of the creative process. While some translators aim to follow an author’s style to the letter, for him, the core of the job is to convey the feelings in a text, rather than reproduce the exact wording. Sometimes, after working for hours on a draft, he will review it and find areas that aren’t exactly faithful to the source novel — though revising them might disrupt the structure and context that he’s already built. “If you read the AI translation closely, you’ll find it’s quite fragmented when it comes to context,” he says. “There’s a certain disharmony to it, a sense that the whole thing was pieced together.” However, for Ping, fidelity to the original is essential. “If a translation doesn’t match, I’d simply call it a mistake — that’s non-negotiable,” he says. “No matter how good your education and cultural literacy, or how beautiful and fluent your translations are, if you’re not faithful to the original, there’s nothing to talk about.” There be pirates Of course, without a licensing agreement, both Ping Ming and Nanshan Youlu risk accusations of copyright infringement for releasing their translations online, regardless of whether they profited financially. Under China’s Copyright Law, using published works for personal study, research, or appreciation is considered fair use. However, this doesn’t apply if the use is seen as harmful to the copyright holder’s rights and interests — for example, negatively affecting sales of Murakami’s books in China. “The nonprofit nature of their conduct does not constitute grounds for exemption from liability,” Zhang Jie, an attorney from Shanghai-based law firm Zhihe Partners with experience handling intellectual property IP disputes, tells Sixth Tone. “The law’s control over acts such as reproduction, adaptation, translation, and dissemination via information networks naturally extends to the use of AI. Whether infringement has occurred has little to do with whether a work was created by a human or with AI assistance.” She adds that China legally protects the IP rights of overseas authors under the Berne Convention for the Protection of Literary and Artistic Works https://www.wipo.int/en/web/treaties/ip/berne/summary berne and other international treaties. While this initial breakthrough in AI literary translation may have skirted the law, it does demonstrate the technology’s potential legitimate use by publishing houses. However, industry insiders suggest the overwhelming preference is still for human translators, especially when handling more complex projects. Liu Chengxu, an editor at Guangdong People’s Publishing House, says, “If the original book is straightforward to understand, I’d choose AI translation. Frankly speaking, in those cases, there isn’t much difference between the work of a seasoned human translator and that of an AI translator. But if the original is more complex, like if it’s hard sci-fi or a thought experiment, I would choose a human translator.” Compared with the lightning speed of DeepSeek, a professional translator will need about six months to complete a 350-page book like “The Tale of KAHO.” Then come editorial reviews, proofreading, publication, and distribution, with the book finally hitting shelves after about a year. Meanwhile, translators earn 100 to 120 yuan per 1,000 characters — not a high rate, but still far more than the current cost of AI. Despite the obvious benefits, Puzhao, an editor at the Shanghai Translation Publishing House who goes by a pen name, feels that using AI to translate fiction still isn’t a viable option. “I prefer collaborating with humans,” he says. “When it comes to literary books, we need humans to translate them in their own words.” He cites Emily Wilson’s acclaimed 2017 version of Homer’s epic “The Odyssey” as an example that a new translation of a classical or even modern text has the power to spark widespread discussion. This, he says, demonstrates that translation is not merely a mechanical transfer of words but a way to use contemporary language to introduce stories to new audiences. On an ethical level, Puzhao concedes that he dislikes AI translation because it “deprives humans of job opportunities,” adding that people “put so much time and effort into creating original content, only for some tech monopoly to feed it into AI training. All we can do is make some small acts of resistance. My position is non-cooperation.” In the summer of 2023, China introduced the Interim Measures for the Management of Generative AI Services, requiring tech companies to carry out training activities only with data from lawful sources and to guard against IP infringement. Whether Ping’s experiment will fundamentally change the future of literary translation is yet to be seen — but it’s certainly got people reading into the consequences. Editor: Hao Qibao. Header image: Visuals from Westend61, Shijue, and iStock/VCG, reedited by Ding Yining/Sixth Tone