AI-mediated feedback in L2 writing: effects on performance, engagement and instructor perspectives in an Armenian university context

Abstract

Written corrective feedback (WCF) is a cornerstone of second language (L2) writing instruction, and the emergence of AI-based tools has substantially transformed its delivery. However, a critical conceptual distinction exists between AI tools that provide corrective feedback on linguistic form, such as Grammarly and Cambridge Write & Improve, and large language model (LLM)-based tools, such as ChatGPT, which generate broader discourse-level feedback encompassing ideas, organisation, and cohesion. This mixed-methods study investigates how a sequenced combination of these AI tools is associated with the writing performance, perceptions, and engagement of 80 intermediate ESL learners at the European University of Armenia, alongside the perspectives of 10 ESL instructors. A four-week classroom intervention incorporated in-class, timed writing tasks, a learner questionnaire, structured written reflections, and semi-structured instructor interviews. Inter-rater reliability between the two independent raters was high (ICC > .90). Quantitative analysis using paired-sample t-tests and a rubric adapted from Bitchener and Ferris (2012) revealed statistically significant pre-to-post improvements across grammatical accuracy, lexical variety, cohesion, and overall writing quality. In the absence of a control group and given the use of different essay prompts at pre- and post-test, these gains should be interpreted with caution rather than attributed causally to the AI intervention alone. Qualitative thematic analysis of reflections and interviews identified four learner themes: enhanced error awareness, structured revision support, difficulty evaluating feedback appropriateness, and emerging dependency, and five instructor themes centring on efficiency gains, risks of uncritical revision, uneven engagement, ethical concerns, and support for a hybrid pedagogical model. Drawing on feedback literacy (Carless & Boud, 2018), the Noticing Hypothesis (Schmidt, 1990), and Sociocultural Theory (Vygotsky, 1978), the study proposes a conceptual teacher-AI partnership framework in which automated tools complement, rather than replace, human pedagogical judgment, pending empirical validation through future research incorporating teacher-mediated designs. The study makes three contributions to the CALL literature: an empirical account of sequential multi-tool AWE integration examined across behavioural, cognitive, and affective engagement dimensions; evidence that feedback literacy development is proficiency-linked and requires explicit instructional scaffolding rather than tool exposure alone; and a theoretically grounded teacher-AI partnership framework positioned as a basis for future controlled experimental designs in EFL writing pedagogy.