This study provides a qualitative evaluation of Seqtagger, a state-of-the-art machine learning-based sequence-tagging model developed for grammatical error detection (GED) and correction (GEC). The model's performance is evaluated on error detection against human benchmarks, with academic texts written by Japanese university students. Through human annotation and subsequent thematic analysis on failures in error detection, this study reveals that Seqtagger performs well in detecting errors related to simpler grammatical rules such as adverb position and prepositions in fixed collocations, with poorer performance in errors possibly influenced by the Japanese language, macro-structure errors and errors where human judgment is required. The underlying reasons for failures in detection are identified to be a narrow context window that fails to capture broader textual information, insufficient training data, particularly data that fully represents the linguistic characteristics of the Japanese students, and overgeneralization of patterns from the training data. These findings highlight the need for sequence-tagging GED and GEC tools to enhance their context window, be more adaptable to the diverse linguistic features of global learners and to enhance the ability to understand the linguistic complexities of the English language.