Human-AI Collaboration in Language Assessment: EFL Learners' Trust, Anxiety, and AI Literacy in AI-Mediated Writing Feedback

الملخص

This study aimed to examine how EFL learners' trust in AI feedback, AI-related anxiety, and AI literacy relate to their engagement with, and uptake of, AI-mediated written corrective feedback. An explanatory sequential mixed-methods design combined survey data from 300 university EFL learners, analysed with partial least squares structural equation modelling (PLS-SEM), with follow-up semi-structured interviews (n = 16); the structural model explained 48% of the variance in feedback uptake. AI literacy positively predicted trust (β = .46) and negatively predicted anxiety (β = -.39); trust was the strongest positive predictor of both engagement (β = .41) and uptake (β = .44), while anxiety exerted smaller negative effects. A cluster analysis identified three learner profiles - Confident Collaborators, Anxious Verifiers, and Detached Low-Literacy Users. Because anxiety was measured at three points across the period of AI-mediated collaboration, change over time was also examined: a repeated-measures ANOVA showed AI anxiety declining significantly across the three waves (partial η² = .18), and a latent growth curve model showed that higher AI literacy and higher trust each predicted a steeper decline. The qualitative findings showed that learners did not accept AI feedback uncritically but practised calibrated trust, verifying machine feedback against teachers and peers, and that AI literacy functioned as a regulating resource that tempered both over-reliance and anxiety. On this basis the study proposes a Calibrated Trust Framework for AI-mediated assessment, in which AI literacy supports calibrated trust, calibrated trust drives engagement, and engagement converts into uptake. The study advances a learner-centred account of human-AI collaboration in language assessment and argues that AI literacy, rather than tool sophistication alone, is central to responsible, pedagogically meaningful, AI-mediated assessment. Implications for assessment design, teacher education, and ethical practice are discussed.