Designing AI for learning, not just correction: A study of ISAR-based interventions on learners' writing outcomes, frustration tolerance, error detection, and error correction

Despite the growing use of AI in writing instruction, limited research has examined how different AI-integration designs simultaneously influence EFL learners’ writing performance, frustration tolerance, and error detection/correction. Addressing this gap, this mixed-methods study investigated four AI-based interventions mapped onto the ISAR model: inversion (ChatGPT), substitution (NoRedInk), augmentation ge(Ecree), and redefinition (Socratic Playground), compared with a control group. Instruments included pre/post writing tasks, the Frustration Discomfort Scale, an adapted Error Orientation Questionnaire, and semi-structured interviews. Using a sequential explanatory design with 125 intermediate learners, repeated-measures ANOVA revealed significant between-group differences: augmentation produced the strongest gains in error correction, substitution yielded the highest writing improvement, and redefinition was associated with the largest increase in frustration intolerance, indicating reduced frustration tolerance. Qualitative findings highlighted autonomy, cognitive challenge, and varied emotional responses across groups. The study recommends pedagogically aligned, scaffolded AI integration, favoring augmentation and redefinition for deeper learning, while providing training to prevent over-reliance and cognitive overload.