This short paper reports on a six-week AI literacy module completed in spring 2026 in Modern Korean Texts 2, the most advanced Korean language course at the University of Kansas, supported by a university-wide AI Literacy Pilot Program. The module asked a focused question: to what extent can generative AI read haengkan (행간)—the space between the lines on which Korean texts so often depend for meaning? Korean grammar distributes meaning across pragmatic and contextual cues rather than fixed syntactic positions, making haengkan not only a literary metaphor but a typological feature, and thus productive ground for examining where human and machine inference diverge.

Four students (three graduate, one undergraduate, with varied linguistic and cultural backgrounds) participated. The module integrated three activities: translation experiments, including round-trip translation across multiple AI systems (ChatGPT, Microsoft Copilot, Claude, Gemini); a two-week comparison of Han Kang’s The White Book across the Korean original, Deborah Smith’s English translation, and AI translations in both directions; and collaborative creative writing with AI, culminating in a photo-essay exhibition. Assessment used collaborative rubric design with self-assessment—one form of alternative grading—in which the class co-constructed evaluation criteria in Week 1 and applied them across two reflective self-assessments and a final project reflection. The framework draws on neural machine translation scholarship that questions claims of “human parity” (Toral / Way 2018; Läubli et al. 2018), on Bender and Friedman’s account of pre-existing, technical, and emergent bias (Bender / Friedman 2018), and on alternative grading scholarship that aligns assessment with interpretive process rather than correctness (Blum 2020; Stommel 2024).

Student artifacts and reflections cluster around three movements of haengkan.

AI flees haengkan. In an experiment session, the class gave several AI platforms the same opening sentence—“나는 그 문을 열지 말았어야 했다” (“I should not have opened that door”)—and asked each to continue the narrative. The Korean sentence, in retrospective conditional form, holds tension, hesitation, regret, and the unspecified threat of what lies behind the door; for the human readers, the force of the line was precisely its refusal to open. Across platforms, AI responses systematically fled this suspended position. Most refused to open the door narratively at all: the speaker turned away, walked elsewhere, abandoned the threshold. Those that lingered redirected attention to safe sensory surfaces—the cold of the metal handle, the texture of the door, light leaking through a gap—rather than the emotional or imaginative interior the sentence opened. AI could neither tolerate the hesitation nor imagine what lay beyond. Other observations confirmed the pattern: in one student’s poem, the line “the trees look dead,” set in midday heat, prompted AI to skip to nighttime; an intentionally ambiguous line—바람 소리만 들을 수 있다 (“only the sound of wind can be heard”)—produced an AI response describing the wind in detail, eliminating the listener’s suspended position. Students also identified a recurring AI lexicon (숨결 breath, 미세하다 subtle, 기묘하다 uncanny) inserted decoratively into gaps regardless of context. Across these instances, AI treated haengkan as something to evade or repair rather than inhabit.

AI flattens haengkan across translational steps, but the human–AI binary does not hold. The round-trip exercise—translating Han Kang’s Korean into English (via Copilot or via Smith) and then back into Korean via AI—produced especially revealing comparisons. Han Kang’s “이마를, 눈썹을, 뺨을 물큰하게 적시는 진눈깨비” lists three body parts in rhythmic succession, with no explicit subject and the mimetic 물큰하게—a word whose phonetic and morphological shape evokes the sensation of dampness clinging to skin, a meaning carried not by denotation but by the texture of the word itself. Smith renders the line as “dampens her eyebrows and streams from her forehead,” introducing “her” and reducing three parts to two. AI back-translation of Smith’s English then produced “그녀는… 뺨과 눈썹을 무겁게 적시는 진눈깨비”—replacing 물큰하게 with the flatter, descriptive 무겁게 and importing “그녀” as a Korean subject foreign to the literary register. The round-trip made cumulative drift visible: subject insertion, loss of the mimetic’s bodily sensation, dissolution of rhythmic enumeration. However, across the larger comparison, students did not uniformly prefer Smith. Several noted moments where AI preserved more of the source text’s material than Smith did, while Smith preserved more of its emotional range and ambiguity. Translation emerged as competition not between human and machine but among different strategies for negotiating haengkan—a finding that complicates the framing on which much AI translation critique relies.

Haengkan emerges through resistance and negotiation. The most generative findings came when students reframed AI as interlocutor rather than authority. In a line-by-line co-writing exercise, students alternated lines with AI in a single poem or short prose piece; many of the most productive moments were ones of refusal. One student, writing about an ant, found that AI brought the ant directly home; she rejected the line because she wanted the ant to struggle, to encounter obstacles before any arrival. Another wrote: “I struggled for control when AI wrote 그러나 저녁이면 작은 빛들이 깨어나 (but in the evening, small lights awaken)… I wanted to describe the full day in the desert, but this line came earlier than I intended.” These redirections, paradoxically, became the moments when students articulated what they wanted to write—authorial intention surfacing through resistance to a too-quick completion. A different photo essay took the form of an epistolary exchange with “Jisoo,” a fictional middle-school Korean teacher in Busan, played by Claude under a prompt to use 반말 (casual form) and to share details of “her” Busan life; when the AI contradicted itself across letters, the student named the inconsistency and the exchange continued. Across these activities, AI functioned less as a model to imitate or critique than as a surface against which authorial decisions became visible.

These findings suggest that haengkan offers a productive analytic site for examining what generative AI does and does not do with meaning. Rather than asking whether AI can translate Korean, the module made AI’s interpretive tendencies trackable through round-trip translation and creative collaboration, treating translation and writing as negotiated processes documented through co-constructed assessment. The approach is transferable to other typologically similar languages where pragmatic inference carries semantic weight, and to classrooms where authorship, translation, and machine inference intersect.