AI Translation between Interpretation and Hallucination – Evidence for the Evaluative Competence of Vietnamese Students

Generative AI complicates translation evaluation because fluent outputs may introduce unsupported meaning, while non-literal renderings may represent legitimate interpretation. This study examined Vietnamese EFL students’ evaluation of AI-generated translations and its associations with digital competence, translation competence, and critical-evaluation confidence. A cross-sectional study involved 213 final-year English majors who completed a 25-item performance task and self-report scales. Students were considerably more accurate in recognising problematic outputs than in diagnosing their type (68.0% vs. 38.4%). Critical-evaluation confidence showed the strongest association with performance (r=.85), although its conceptual overlap with the performance task warrants caution. In the full regression model, translation competence (β=.31) and digital competence (β=.16) contributed significantly alongside critical-evaluation confidence (β=.64), whereas AI-use frequency did not. When the overlapping critical-evaluation measure was removed, the model still explained 55% of performance variance, with translation competence showing the strongest association (β=.61). The findings indicate that evaluating AI-generated translation depends less on frequency of AI use than on students’ translation knowledge and evaluative judgement, particularly their ability to distinguish unsupported generation from defensible interpretation. Rather than simply as a translation tool Generative AI should thus be used as an object of critical inquiry. Alternative AI renderings may expose students to interpretations they had not considered, including potentially productive expansions of source meaning. This approach preserves the interpretive character of translation while making students accountable for the decisions they ultimately accept.