Machine translation can already process large volumes of text quickly and often at good quality, but literary translation demands more: beyond conveying word meanings, it must transmit cultural, emotional and stylistic layers. Translator Patat Bence made this point on the podcast A nagy AI-sztori; the show’s regular expert guest is Aczél Petra, communication researcher and professor at Széchenyi István Egyetem.
Professional translation versus literary translation
At the start of the episode Patat drew a clear distinction between professional (technical) translation and literary translation. In technical work, accuracy and clarity are paramount. Literary translation, by contrast, requires the translator to recognize and render the author’s voice, the text’s nuances and its cultural context in another language. Patat emphasized that every new translation carries some element of the translator’s own reading — their personal experience and cultural background — elements that current AI lacks.
Practical uses: AI as an aid
Patat reported encountering a Hungarian-published book where no traditional translator was credited and the text had been produced with the assistance of a generative tool. Abroad, it is already common for a machine to produce a rapid draft that is then edited by a human — a workflow Patat considers realistic for parts of the trade. This approach helps publishers get anticipated titles to market faster, which is a strong economic incentive.
Patat compared the difference between a fast AI-based solution and a human-crafted translation to the contrast between an instant soup package and a Michelin-starred restaurant: the quicker, cheaper option may be perfectly adequate for some purposes, but human contribution still delivers added value.
Where AI still falls short
The discussion highlighted that large language models are built primarily on past texts, while literature often experiments with new styles and perspectives. AI also lacks implicit knowledge and personal experience: it has no lived life, personality or subjective memory to bring into interpretation.
This limitation is particularly visible with smaller languages. Patat, who currently also works from Faroese, notes that the fewer the textual examples available in a language, the less reliable the model becomes. The technology also remains prone to confidently presenting fabricated information (so-called hallucinations), which poses specific risks for historical, legal or scientific texts.
How Patat uses AI himself
Patat does not reject AI outright: he uses generative tools for research and idea generation, and expects they will appear in many more tasks in the future. His central question is what will remain valuable when machine translation becomes widespread.
"Human translated" as a quality label
Patat argues that as AI-generated translations proliferate, the presence of a clearly credited human translator on a book cover could grow in importance. In that sense, an explicit “human translated” credit could become a sign of quality in the AI era.
Topics and timestamps from the episode
- What is the difference between translation and literary translation? (00:30)
- How does AI appear in book translation? (06:47)
- What can humans do that AI cannot? (12:30)
- Fast AI vs. human quality? (17:43)
- Small languages, errors and fabricated answers (20:29)
- Growing publisher pressure and reader expectations (28:27)
- Is there a future for human literary translation? (31:59)
The full episode is available on the podcast channel; the conversation sheds light on how AI is reshaping the translation profession and what role human creativity and experience may continue to play.
