Exploring user trust in generative AI feedback in academic writing: a psychometric network analysis

Abstract

Generative artificial intelligence is increasingly embedded in the academic writing practices of learners of English as a foreign language, yet research has commonly represented learner evaluations through construct averages. This exploratory mixed-methods study used psychometric network analysis to examine how 20 specific beliefs about AI-generated writing feedback were conditionally associated. Questionnaire and open-ended data were collected from 92 undergraduates at one university in Uzbekistan; 86 complete cases entered the network. Because the items used five-point ordinal response categories, a rank-based Gaussian-copula transformation preceded estimation of a cross-validated graphical LASSO network. Coherence, understanding errors, trust in accuracy, revising more frequently, and spotting mistakes occupied comparatively connected positions. Cross-domain connectivity was concentrated around coherence, understanding errors, mistake identification, organization, and confidence in applying feedback. A three-community solution distinguished a broad writing-quality and actionable-feedback region, a revision-engagement region, and a reliance-and-limitation region, but bootstrap evidence showed that the exact community allocation was unstable. Learner comments extended the network by showing why visible text improvement encouraged trust while inaccurate, generic, or context-insensitive suggestions sustained caution. The findings are hypothesis-generating rather than confirmatory. They suggest that trust in AI feedback is calibrated through perceived writing improvement, active interpretation, and continued human judgment rather than through uniform acceptance.