Helpful Yet Risky: Senior EFL Students' Perceptions of ChatGPT Feedback in Academic Writing at a Vietnamese University

Abstract

Generative artificial intelligence has entered English as a foreign language (EFL) writing classrooms faster than evidence about how learners construe it has accumulated. This study examined how senior English-major students at a Vietnamese public university perceive ChatGPT feedback on their academic writing. A cross-sectional descriptive survey with an embedded qualitative component combined a 15-item questionnaire – thirteen agreement items and two evaluation items – with written answers to three open-ended prompts from eight students in the same cohort. Thirty-six students responded and 30 complete cases were retained. No writing performance measure was collected, so no claim about learning gains is made. Benefit endorsement was highest for confidence in writing (M = 3.93) and for grammatical correction and academic vocabulary (both 76.7% agreement, M = 3.90), and lowest for organisation of ideas (60.0%, M = 3.67), though differences within this cluster are small relative to the sample precision. Risks were endorsed at least as strongly: 96.7% agreed that over-reliance might erode self-editing (M = 4.17), and 83.3% that ChatGPT revisions can make writing sound unlike their own (M = 4.07). Across the thirteen agreement items, means occupied a narrow range (3.67–4.17), indicating co-endorsement of benefits and risks rather than demonstrated statistical equivalence. The open-ended responses mirrored this pattern, portraying ChatGPT as a useful assistant, a threat to originality, and a tool requiring explicit rules. The pattern may reflect genuine ambivalence, acquiescent responding, or both; the present design cannot separate them. Implications are discussed for feedback literacy, classroom guidance, and instrument design.