Grading grammar is a nuanced task, even for humans. Large language models (LLMs) like ChatGPT are surprisingly good at generating grammatically correct text, but how do we *measure* their accuracy, especially when there might be multiple valid ways to fix a sentence? Traditional metrics often fall short, struggling to capture the subtleties of language and sometimes penalizing LLMs for making perfectly acceptable edits. New research introduces DSGram, a clever framework that evaluates grammatical error correction (GEC) with a more human-like touch. Instead of relying on rigid scoring systems, DSGram dynamically weighs three key aspects: Semantic Coherence (does the meaning stay the same?), Edit Level (were the changes necessary?), and Fluency (does it sound natural?). By using the Analytic Hierarchy Process (AHP) and input from LLMs themselves, DSGram figures out how important each factor is for a specific sentence. For instance, a formal legal document needs different weighting than a casual conversation. This mimics how a human grader would intuitively prioritize different aspects depending on the context. Experiments on benchmark datasets like SEEDA show that DSGram correlates better with human judgment than traditional metrics. This not only helps us better evaluate current GEC systems but also provides valuable feedback for training even more sophisticated AI writing assistants in the future. While challenges remain, DSGram highlights how AI can move beyond simple rule-following to achieve more nuanced and human-like assessments of language. Further research will explore DSGram’s applicability with other LLMs and evaluate its use in practical GEC model training. This means future AI writing tools will likely be better equipped to produce not just correct, but truly effective and natural-sounding text, adapting to different writing styles and situations with ease.