Research Article
University LOTE Teachers’ Engagement with Generative AI: A Qualitative and Configurational Study
Synthesis: Wang, Ding and Liu ask how university teachers of languages other than English engage with generative AI when the underlying models are trained predominantly on English and a few high-resource languages. They interviewed 18 LOTE teachers at three universities in China's Yangtze River Delta — nine teaching high-resource languages such as French, German, Portuguese, Japanese and Russian, nine teaching low-resource languages such as Serbian, Greek, Swahili, Thai, Hebrew, Persian, Korean and Arabic — and analyzed the transcripts with procedural grounded theory before subjecting the same cases to fuzzy-set Qualitative Comparative Analysis. The grounded-theory stage yielded six conditions: belief in success, task value, relative cost, social support, perceived ease of use and perceived usefulness. The configurational stage produced four high-engagement pathways in two solution families, which the authors read as equifinality rather than a single adoption recipe. Where tools were mature, engagement was mainly behavioral; where they were deficient, engagement became sustained cognitive and affective effort, called adaptive labor. The sample is small, geographically concentrated and cross-sectional, and no single condition proved necessary for high engagement.
Key Findings
- Six conditions, one model. The grounded-theory analysis distilled 127 initial codes and 23 categories into six conditions shaping engagement: belief in success, task value, relative cost, social support, perceived ease of use and perceived usefulness.
- No condition was necessary for high engagement. Nothing reached the 0.90 consistency threshold; only low task value, social support and low perceived ease of use crossed it for low engagement, and social support was present in all 18 cases.
- Four pathways, two families. The intermediate solution combined four high-engagement configurations: an ease-of-use scaffold and three belief-anchored variants, illustrating equifinality — different combinations associated with the same outcome.
- Belief in success anchored the largest family. Three sub-pathways shared belief in success as their core condition, with task value, low relative cost, social support and perceived usefulness combining around it.
- Engagement had different profiles. The ease-of-use pathway showed routinized behavioral integration with modest personal investment; the belief-anchored pathway carried its engagement cognitively and affectively, even where live classroom use was deliberately limited.
- Adaptive labor came in four forms. Low-resource teachers adapted prompts, corrected language, contextualized culture and switched tools — off-stage work that institutions neither recognized nor rewarded.
- High-resource is relative. Even French, Portuguese and Japanese teachers described weaker tool support than English enjoys; hierarchies also ran between varieties of the same language, such as European and Brazilian Portuguese.
Design: interviews, artifacts and a configurational second stage
The study drew on three universities in the Yangtze River Delta and combined purposive sampling with theoretical sampling. Participants had taught at least one LOTE course in the previous three years, had practical experience with generative AI, and worked at institutions able to support integration if they chose it. Initial saturation came after 13 interviews; five further teachers of Persian, Japanese, Korean, Arabic and Russian were added, and final saturation was reached at 18. The sample covers 13 languages, split evenly between nine high-resource and nine low-resource languages, with 5 men and 13 women, 13 doctoral and 5 master's degrees, and ranks from teaching assistant to professor.
Data came from three sources that capture different levels of observability: transcribed one-to-one interviews (334,785 words, 50–70 min, in Mandarin), institutional policy documents (5,463 words), and voluntarily submitted teaching materials (2.535 GB of video, text and slides). Interview accounts were triangulated against artifacts and documents.
Two authors coded the first 13 transcripts, teaching materials and policy excerpts; initial inter-coder agreement was 78.3% and rose to 84.7% after three rounds of discussion, above the 80% threshold the authors set. Open coding generated 127 initial codes, axial coding 23 categories, and selective coding the six conditions. The same 18 cases then entered fsQCA 4.1 with a four-value calibration (0, 0.33, 0.67, 1); two researchers calibrated independently, with Cohen's Kappa of 0.78. Because the qualitative evidence and the calibration came from the same cases, the authors are explicit that no separate quantitative instrument or independently collected dataset was used. The dominant alternative in the literature — the technology acceptance model and its regression-based elaborations — assumes independent, uniform and linear effects, which the authors argue cannot represent conjunctural causation in linguistically unequal settings.
What the configurations show
Necessity analysis found no condition that had to be present for high engagement. For low engagement, three candidates crossed the 0.90 consistency threshold: low task value, social support and low perceived ease of use. Social support, however, was widespread — every case scored either 0.67 or 1 — so its low coverage limits its interpretation. The 18 cases occupied ten of the 64 logically possible combinations, leaving 54 logical remainders; the software reported a consistency cutoff of 0.914573.
The intermediate solution contained four high-engagement configurations organized into two families, with overall solution consistency of 0.921 and coverage of 0.816. The parsimonious solution was belief in success plus perceived ease of use. The first family, H1, pairs perceived ease of use as the core condition with social support; T15, a Japanese lecturer with four years of experience, illustrates it through routine lesson preparation and administrative use, alongside a judgment that generative AI's contribution to teaching quality was "not yet felt." The second family, H2, has three equifinal sub-pathways united by belief in success — a self-efficacy-like conviction — as the core condition. H2a adds task value, low relative cost and social support; H2b adds low relative cost, social support and perceived usefulness; H2c adds task value, social support and perceived usefulness. T7, a French professor, T4, a Greek program director, and T3, a Portuguese lecturer, exemplify the three variants. Task value, the authors note, functions as motivational value attached to the work itself.
The examined low-engagement combination did not meet the sufficiency criterion — its consistency fell short of the 0.80 threshold — so its coverage is not read as a sufficient pathway. A supplementary account of T18, a Russian teacher with 27 years of experience, recorded reservations about generative AI's pedagogical compatibility, including that it "cannot perceive students' expressions, nor can it longitudinally track their development," but the authors treat this as case-specific rather than a general limit of institutional support. Sensitivity analysis comparing frequency thresholds of 1 and 2 shifted complex-solution coverage from 0.746862 to 0.628312 while consistency rose from 0.913823 to 0.963636; the authors note this assesses sensitivity to the threshold, not the stability of the four baseline configurations.
Adaptive labor and linguistic hierarchy
The paper's central interpretive move is adaptive labor: the extra cognitive, linguistic and technical effort low-resource teachers expend to make tools designed for high-resource environments function in their teaching. The authors identify four types. Prompt adaptation means engineering queries to the target language level, as when T13 calibrated Hebrew prompts to his six-level proficiency scale. Language correction is the most pervasive: T9 re-processed AI feedback on homework and fact-checked extended responses, and T5 dismissed ChatGPT's Serbian grammar answers as superficial and cross-checked trusted sources. Cultural contextualization replaces or supplements output with authentic materials — T16 found a culturally loaded wine term rendered as "Tibetan wine." Tool switching moves across platforms when one fails, as T4 did when Greek-supportive tools were scarce.
This labor is articulation work and invisible work: unrecognized by institutions, unrewarded in workload accounting, and unmarked in its products. The authors distinguish it from adaptive teaching — adaptive teaching adapts instruction to learners, whereas adaptive labor adapts tools and their output to teaching practice. The labor is not a byproduct of inadequate digital competence; it is imposed by linguistic hierarchy. The equity argument follows from the training data: English accounts for over 40% of web pages while no other single language exceeds 7%, and the authors classify languages above a 1% share of the CommonCrawl corpus as high-resource. Korean illustrates the gap between speaker numbers and corpus share — it falls below the 1% threshold and is treated as low-resource here. The consequence is that linguistic equity, not only technology acceptance, becomes the frame for adoption, with implications for differentiated professional development, resource allocation and tool design. Multilingual learning contexts therefore cannot be assimilated into frameworks built for English-medium instruction.
What this means for practice
- Teachers. Expect differentiated support rather than one workshop: high-resource and low-resource teachers in this study occupied different configurations and faced different tool ecologies.
- Faculty developers. Design targeted professional development that respects distinct configurational profiles, including curated tool recommendations for low-resource languages.
- Administrators. Count adaptive labor in workload and resource decisions — time for correction, verification and tool assembly is a condition of engagement, not a fringe activity.
- Researchers. Treat the four configurations as exploratory patterns in a bounded sample; longitudinal work is needed to test their stability as tools and policies change.
- Developers. Move linguistic inclusivity from aspiration to technical reality: expand multilingual training data and co-design tools with LOTE educators.
Limitations
- A purposively selected sample of 18 teachers at three universities in one region; the authors treat the configurations as exploratory patterns within this bounded sample, not population-level estimates.
- A cross-sectional design captures one moment, while generative AI tools and institutional policies evolve rapidly; no longitudinal evidence is offered.
- The sample contains no EFL teachers, so the contrast with English rests on the computational and survey evidence the authors cite rather than on a comparison group.
- Low engagement is thinly evidenced: the examined low-engagement combination did not meet the sufficiency threshold, and the only illustrative account (T18) is case-specific.
- Necessity results are qualified — social support was present in all 18 cases, and the identical consistency values for social support and low perceived ease of use reflect identical case-level intersections.
- Adaptive labor is theorized from interviews and artifacts, not measured as time spent, workload or cost.
Citation
Wang, Y., Ding, W., & Liu, Y. (2026). University LOTE Teachers’ Engagement with Generative AI: A Qualitative and Configurational Study. Education Sciences, 16(11), 1994.