Beyond Algorithmic Reliance: Mitigating Cognitive Offloading in AI-Assisted Translation Training through Multidimensional Scaffolding

AI-assisted translation gives students fluent target-language drafts almost instantly, but this convenience may weaken their engagement with the source text. When students accept AI output before forming their own interpretation, they risk offloading translation judgment along with linguistic search. This study addresses this risk by developing a multidimensional scaffolding framework for AI-assisted translation training. The framework was put into practice in a process-oriented English-to-Chinese translation assignment requiring students to identify difficulties, evaluate AI drafts, revise prompts, annotate decisions, and reflect on their own contribution. A qualitative analysis of students’ diagnostic notes, prompt records, annotations, and reflections shows that external pedagogical scaffolds can make human–AI decision-making more visible and push students toward more targeted prompt revision, supporting translator agency development. The paper argues that AI-assisted translation in educational settings should be understood as a structured process of interpretation and responsibility-taking that leads students beyond the pursuit of a fluent product.