Learner agency in GenAI-supported EFL academic writing under conditions of teacher trust

Learner agency in L2 academic writing has taken on new significance with the rapid uptake of generative artificial intelligence (GenAI) tools, raising questions about how agency is exercised in interaction with these tools and what pedagogical conditions support agentic and learning-oriented use. This study examined how 27 Japanese undergraduates in a 14-week English academic writing course exercised agency across three AI-permitted out-of-class essays, and how that agency related to the development of their unassisted writing. The course was designed around a pedagogical stance of teacher trust with two named elements: discretion granted to learners over their AI use, and a structure of explicit guidelines, AI literacy instruction, articulated principles, and weekly group discussions. Post-writing reflections (n = 81) were coded for the function of tool use and the quality of agency across seven facets, with substantial inter-rater reliability for both. Computational indices of lexical diversity, lexical sophistication, syntactic complexity, and accuracy were extracted from in-class AI-free pre- and post-essays, and Turnitin AI detection scores were recorded as research data, not as a policing instrument. Individual facets of agency were associated with different aspects of unassisted writing development: metacognitive awareness with differences in length, learning orientation with gains in lexical diversity, and critical evaluation with gains in clause density. Composite agency was not associated with development on any measure, supporting a dimensional view of agency in which facets carry different consequences for learning. Unassisted writing also shifted toward shorter, less subordinated clauses across the semester, a direction conventional measures of L2 syntactic development were not designed to detect. The findings support a view in which trust as a pedagogical condition shapes the forms of agency available to learners.