Translated Korean literature sold around 1.2 million copies abroad in 2024, more than double the previous year. But the widely circulated story of a Korean Translators Association campaign against publishers using DeepL and Papago on manuscripts doesn’t hold up: the KTA is a training and certification body, not a literary translators’ union. The actual bottleneck is elsewhere, and machine translation doesn’t fix it.
In 2024 translated Korean literature sold roughly 1.2 million copies abroad — more than double the 520,000 of the year before. At home, fiction sales grew by almost 29% year on year, and the Seoul International Book Fair drew a record 150,000 visitors. Han Kang, the first Korean writer to win the Nobel Prize in Literature, detonated a curve that had already been climbing for years. The problem with that success is mundane and structural: Korean is a “small” language, and there simply aren’t enough literary translators capable of carrying an entire novel into English, French or Italian. Industry observers point to the shortage of translation capacity as the choke point of Korea’s publishing export boom.
When a market has more demand than craftspeople, the technological temptation shows up on schedule. And the technology exists: neural machine translation systems — networks trained on vast parallel corpora, meaning texts already translated by humans, which learn to map word sequences from one language to another. Papago is the multilingual service run by Naver, Korea’s dominant internet group; DeepL, a Cologne-based company founded in 2009 and operating since 2017, didn’t even support Korean for years, and only in August 2023 launched its paid version in Korea with 31 languages, in direct competition with Papago. These tools produce plausible output in seconds. The idea that someone might want to use them as a base layer for literary translation, leaving humans to clean up, is anything but far-fetched.
The campaign nobody can find
A sharper version of this story circulates, though: that the Korean Translators Association launched a campaign against publishers handing translators machine-pre-translated drafts from DeepL or Papago to fix. It’s a perfect news item — it has a villain, a technology, a trade under siege. We couldn’t find it.
The Korean Translators Association (사단법인 한국번역가협회) does exist. It’s a non-profit founded in 1971, headquartered in Seoul’s Jongno-gu district. But its business is something else entirely: it trains translators and certifies their skills, running the TCT examination three times a year, in March, July and November. It is not a literary translators’ union, it doesn’t negotiate with publishers, and nothing in its documented activity touches artificial intelligence or post-editing. No statement, no petition, no code of ethics. If such a mobilisation existed, the plausible actors would be others — LTI Korea, the publishers’ association, the translator collectives that have emerged in recent years. No source we reached attests to it.
This is worth saying plainly, because the pattern recurs: when a technology enters a craft, stories of organised resistance appear that sound true before they are true. Telling them badly helps nobody, least of all the people who actually do the work.
What a translation model actually does to a novel
The technical point is more interesting than the controversy. A neural machine translation system essentially optimises the probability of the target sentence given the source. It is superb at producing the most predictable segment. Literature, by definition, lives on choices that are not predictable: registers that break, deliberate repetitions, ambiguities that must stay ambiguous, a rhythm belonging to one narrator rather than to the language at large.
Korean makes this sharper. Politeness levels are grammaticalised into verb endings: the distance between two characters is encoded in the suffix, not in the content. Subjects are dropped with a freedom English won’t tolerate, and reinstating them means deciding, every single time, who is speaking and about whom. Korean ideophones and onomatopoeia form a dense expressive system with no European equivalent. A statistical model resolves these things by averaging: it picks the most frequent pronoun, flattens the register, normalises the anomaly. The result isn’t wrong. It’s worse than wrong — it’s acceptable.
Post-editing, for that matter, is no academic taboo in Korea. It is an established object of research and teaching: there are published studies on post-editing guidelines for the English-Korean pair applied to literary style, and experimental research conducted with translation students. The Korean scholarly community already has the problem in hand — which makes inventing a holy war about it even less necessary.
The craft isn’t typing time
The case for post-editing is that the machine does the first pass and the human does the second, saving time. But a literary translator’s time doesn’t go into the first pass. It goes into building a coherent voice that holds for three hundred pages, into decisions made on page 40 and paid for on page 280, into the reread that overturns a choice made three months earlier.
Anyone who has worked on pre-translated text knows this: correcting is cognitively different from translating. Faced with a sentence already written and not entirely wrong, the brain economises. It’s called anchoring, and it’s one of the sturdiest regularities in decision psychology. The cost isn’t effort — it’s that the mediocre solution that is present almost always beats the optimal solution that is absent.
A machine translation model doesn’t make crude errors on easy text. It makes average choices on difficult text. That is precisely the opposite of what literature needs.
Who opened the door, and how
That the value sits in the person rather than in throughput is proved by the very trajectory that made this boom possible. Han Kang’s The Vegetarian appeared in English in 2015, translated by Deborah Smith, and won the Man Booker International the following year, opening a market that had not previously seen Korean fiction. Smith went on to found Tilted Axis Press, an independent house devoted to translation, which won the International Booker in 2022 with Geetanjali Shree’s Tomb of Sand, translated from Hindi by Daisy Rockwell.
None of that came from optimising delivery times. It came from translators who chose what to translate, built the editorial context to publish it in, and imposed a voice. Since the mid-2010s, selection power in Korea has shifted from state agencies to publishers and foreign agents, who now push more than two hundred exported titles a year; the growth, Han Kang aside, concentrates on so-called healing novels and genre fiction. It is a market that asks for volume — and volume is exactly what a machine knows how to promise.
The thing to watch isn’t the software
Here’s what’s worth taking away. The risk for Korean literary fiction is not that some publisher presses a button and mails a DeepL output to a translator. It’s that a language which entered the global market thanks to a handful of people who interpreted it gets treated as a stream to be processed, at a moment when demand is growing faster than translators can be trained.
And there’s a detail the debate almost always ignores. Machine translation systems learn from parallel corpora: texts already translated by humans. For a language with such a limited pool of literary translators, that corpus is thin and reflects a small number of hands. The more the machine is used to replace those hands, the less new and different material enters the corpus the machine learns from. Literary machine translation, in a “small” language, is not a renewable resource: it consumes the very thing it is supposed to produce.