This study investigates the application of large language models (LLMs) in translation teaching, focusing on their effectiveness and limitations in assessing student translations. Using established standards for human translation quality evaluation, a two-tier analytical framework combining quantitative and qualitative analyses was developed, incorporating human scores, LLM-generated scores, and evaluative comments. Quantitative results indicated that the LLM performed reliably in structural dimensions of Chinese-to-English tasks, but a marked decline was observed in semantic and cultural dimensions in English-to-Chinese tasks, exposing weaknesses in deep semantic understanding and cultural adaptation. Qualitative analysis further revealed issues such as templated feedback, misattribution of errors, and rejection of creative translation, thereby corroborating the quantitative findings. Based on these results, the study proposes a human-AI collaborative pathway for teaching practice, highlighting the role of LLMs as auxiliary tools for structural checking while reserving semantic and cultural evaluation for teachers. The findings provide theoretical grounding and pedagogical implications for the integration of intelligent assessment in translation education, and suggest a shift in LLMs’ role from instrumental support to cognitive collaboration.