VOL.99/2026.10

Topic

 

In Translation, Context Is Where the Difficulties Converge

 

 

2026.09

 

 

When we call something untranslatable, we usually don't mean a word that can't be rendered at all, but one that keeps demanding to be translated again. I'm not persuaded by explanations that locate the difficulty of translating Korean literature in the mystique of culturally loaded words like jeong or bap. What a translator has to carry across isn't some ultimate emotion sealed inside a word, but the connections that word forms within the work itself.

 

 

The narrator of Kim Ae-ran's short story “Ascending Scales” (Moonji Publishing, 2007) is the daughter of a dumpling shop owner. Her mother sends her to piano lessons, believing her child should get the same education as everyone else's. Then the family collapses after the father cosigns a loan that goes bad. The narrator moves the piano into her older sister's semi-basement room in Seoul. She types up manuscripts on a computer to save for tuition while her sister works and studies English. The piano, which once seemed like it would raise the family's standing in life, sits silent, shut inside the lowest place a person can live. One day, during a downpour, water starts seeping into the semi-basement. As it rises, the father calls. All he asks is whether she's eaten (“Bap-eun meogeonnya”). But packed into that question is the shabbiness and exhaustion of their lives, and family ties she cannot escape.

 

Translator Jamie Chang renders the father's question as “Have you eaten?” and the follow-up as “What'd you have?” The specific time of day, dinner, opens out into a meal in general, but the image of food doesn't disappear. The second question survives intact as well. Here, a translation doesn't have to compress every layer of meaning into a single line, because the surrounding narrative context carries much of the weight. The reader has already passed through the dumpling shop and the piano, the tuition and the semi-basement room, the water rising.

 

How well does generative AI pick up on context and the connections between words? While preparing this piece, I ran “Bap-eun meokgo daninya?” through a generative AI model, varying the conditions one at a time. Given the line with no context at all, it came back with “Have you been eating properly?” When I supplied a situation where a mother is asking an adult child she hasn't seen in a long time, it shifted to “Have you been eating okay?” In a context where a suspect in a criminal case is being questioned but it isn't clear whether he's the culprit, it gave me “How can you even eat after what you've done?” When I instructed it to preserve the ambiguity of the original, it offered “Do you still manage to eat?”, and when I told it the image of food had to remain, “Do you still eat?”

 

As the context shifted, the outputs moved consistently along with it—toward care, accusation, or checking on someone's survival. The problem with AI is that its response to context is far too decisive from a literary perspective. When a human translator leaves the strangeness of a literal rendering intact, that isn't a failure of translation. It's a choice to preserve room for the reader to inhabit the context. That's a choice machines simply aren't equipped to make. Feed a machine context, and it digs a channel through which only that particular context can flow. When several channels converge at once into a bottleneck, the machine still struggles to recognize that moment and make a choice. What it lacks is the capacity for poetic compression. A machine's literal rendering and a human's may look alike on the surface, but they operate on entirely different levels.

 

 

If we collaborate with AI in translation, we should have it untangle the various layers of context embedded within a sentence, whether a factual query, an act of care, an ingrained family habit, an accusation, uncertainty, or the materiality of food. From there, it can generate candidate translations that bring each specific layer to the fore, detailing what gets sacrificed with every choice. As a researcher and curator, AI proves genuinely useful. It becomes a dramaturg of translation, bringing unchosen possibilities onto the stage.

 

Yet the final choice still belongs to the human translator. This is not because a human always understands Korean better than an AI, but because someone must decide which layer of meaning to preserve across the work as a whole and take full accountability for what is lost in that choice. In literary translation, context is not a clue that hands us an answer. It is a force field that makes a single sentence even more complex. The true authority of a human translator lies in briefly putting the brakes on that rapid comprehension, recalling the meanings displaced while the machine locks onto a single option. In the age of AI, the translator's most vital virtue is not the speed of finding the right answer, but the capacity to defer resolution.

 

 

 

---

Written by Oh, YoungJin

Since 2015, he has created the courses “Software and Humanities Criticism” and “Machine Criticism” at Hanyang University ERICA. His research focuses on the aesthetics and politics of digital culture as represented by computer games, webtoons, and social networks. He is a co-developer of “Under the Sunlight”, a web-responsive interactive story about Syrian refugees. He also works as a producer and writer, having directed “Erangel: Dark Tour”, a dark tourism piece that lets users experience sites of tragedy in a virtual world, along with “The Voyage Without End”, a metaverse project tied to the academic conference “SF and Geopolitical Aesthetics,” as well as “AI Horror Radio Show”, “ChatGPT WAR Part 1”, and “Creative Boundaries: AI Sentence Miner”. He is the author of The Artificial Intelligence Art Business and serves as general director of the Art Korea Lab 2026 AI Open Prompt Community.