Abstract
This study analyzed the cognitive hierarchy of learners' judgments when creating Japanese-to-English subtitles using AI tools within the Translation in Language Teaching (TILT) framework. While AI fluency is a strength, it can induce an "objectivity bias," causing learners to overlook basic grammatical and syntactic errors (Layer 1). A qualitative case study of 27 Japanese university students applied a three-layer cognitive model integrating linguistic accuracy, pragmatic function based on Relevance and Skopos Theory, and cultural adaptation from Venuti's framework. The analysis revealed a "cascade effect," where lower-layer failures negatively impact higher-layer judgments. Key findings include learners' difficulty detecting pragmatic implicatures (Layer 2) and the occurrence of "cultural flattening," where cultural nuances are neutralized (Layer 3). Learners often struggled to shift from evaluating surface fluency to deeper recontextualization. These findings highlight the model's utility as a diagnostic framework for AI-specific cognitive challenges and hold significant implications for fostering critical AI literacy.