Human-AI Collaborative Feedback in Translator Training: A Mixed-Methods Study

Abstract

The integration of large language models (LLMs) into language education has prompted renewed interest in AI-assisted feedback, yet purely automated feedback remains vulnerable to contextual misalignment, cultural misreading, and reliability concerns that are particularly consequential in translation training. A human-AI collaborative feedback model, in which an instructor curates, corrects, and supplements LLM-generated commentary before students revise, offers a theoretically motivated alternative, yet its pedagogical effects in translator education remain empirically underexplored. This mixed-methods study examines the impact of such a hybrid feedback approach on undergraduate Chinese-to-English student translators. Forty senior undergraduates translated a 1,500-word cultural heritage text and received ChatGPT-4o-generated feedback subsequently reviewed and annotated by an experienced instructor using a color-coded transparency system. Quantitative analysis using a Multidimensional Quality Metrics (MQM) rubric revealed significant pre-to-post gains across all measured dimensions (overall MQM composite: Δ +1.20 on a 5-point scale, p < .001), with the largest improvements in terminology (Δ +1.47) and accuracy (Δ +1.32) and meaningful gains in cohesion, cultural adaptation, register, language conventions, and format (all p < .001). Think-aloud protocols revealed a consistent two-stage revision pattern and active source evaluation behavior, with students demonstrating greater decisiveness when AI and instructor annotations converged and deeper deliberation when they diverged. Student perception surveys indicated high ratings across clarity, trustworthiness, usefulness, and pedagogical value, with no significant differences between high- and low-performing students. Instructors reported meaningful workload relief on routine corrections while retaining pedagogical authority over higher-order feedback. These findings suggest the potential of a human-in-the-loop feedback framework for translator training in which AI handles systematic error detection while instructors validate, contextualize, and model evaluative judgment, a model that warrants further controlled investigation as a means of enhancing translation competence without displacing the pedagogical depth that professional translator development requires.