This study investigates the pedagogical effectiveness of explainable artificial intelligence (XAI) as a novel feedback model for foreign language learners, responding to the rapid advancement of generative artificial intelligence (Generative AI) and t...
This study investigates the pedagogical effectiveness of explainable artificial intelligence (XAI) as a novel feedback model for foreign language learners, responding to the rapid advancement of generative artificial intelligence (Generative AI) and the resulting paradigm shift in academic writing instruction. Traditional AI-based feedback has been criticized for its answer-driven, one-directional corrective approach, which may limit learners’ active cognitive engagement. To address this limitation, the present study focuses on the explanatory capacity of XAI to provide not only revision outcomes but also the reasons and logical rationales behind those revisions, and examines its effects on learners’ cognitive uptake and self-revision processes.
Participants were foreign learners of Korean enrolled in undergraduate programs. Using a within-subject crossover design, this study compared the effects of ChatGPT-based XAI feedback and human teacher feedback through both quantitative and qualitative analyses. The main findings are summarized as follows.
First, XAI-based feedback produced significantly greater gains in academic writing performance than human teacher feedback. In particular, in higher-order discourse domains requiring persuasive content development and structural coherence, the logical scaffolding provided by XAI enabled learners to identify structural weaknesses in their texts and independently reorganize them. By contrast, no significant differences emerged between the two feedback conditions at the micro level (e.g., linguistic accuracy), indicating that XAI’s primary strength lies in supporting macro-level writing strategies rather than surface-level corrections.
Second, time-series analyses revealed not only a general practice effect, reflected in improvements in learners’ overall writing proficiency across repeated tasks, but also a clear development of feedback literacy. Learners increasingly evaluated generative AI suggestions critically and adapted them to align with their own communicative intentions. This finding suggests that XAI functions as a sustainable instructional tool that fosters learners’ internalized writing competence.
Third, learner perception data and qualitative analyses indicated high levels of satisfaction with and trust in the immediacy and explanatory depth of XAI feedback. Learners reconceptualized generative AI from a simple tool to a “cognitive partner” and demonstrated a shift toward greater authorial agency, moving beyond uncritical acceptance to actively selecting and validating feedback. Nonetheless, the findings also confirm that human teachers’ professional mediation remains essential for ensuring the reliability of final decisions and guiding contextually appropriate revisions.
Based on these findings, this study proposes a complementary two-stage feedback model in which XAI provides initial structural diagnosis and logical scaffolding, while teachers offer advanced, expert-level guidance. By empirically integrating XAI into Korean academic writing instruction, this study highlights its pedagogical potential and provides foundational insights for designing future-oriented writing pedagogy in which generative AI and human teachers coexist.