With the wide application of large language models in language education, their capability boundaries in academic-level error tagging urgently need examination. This study takes 30 Chinese compositions by Korean learners from the HSK Dynamic Composition Corpus, strictly follows the corpus tagging code, and compares ChatGPT (GPT-5.3) with professional teachers in error tagging. Quantitatively, precision, recall, and F1-score evaluate tagging quality. Qualitatively, four typical scenarios are analyzed to explore human-machine differences. Three findings emerge. First, teachers significantly outperform ChatGPT in precision (98.82% vs. 21.61%), recall (93.31% vs. 18.96%), and F1-score (0.96 vs. 0.20). Over half of ChatGPT's errors are format violations, not content misjudgments. Second, ChatGPT's performance declines stepwise from HSK levels A to C, while teachers maintain high precision across levels. In A-level texts, ChatGPT detects subtle errors teachers overlook. Third, both agree on rule-governed explicit errors, but differ in semantic differentiation, cultural judgment, and tagging granularity. This study clarifies the capability boundaries of large language models in academic corpus tagging and provides empirical evidence for human-machine collaborative interlanguage tagging systems.