In recent years, generative AI has made large progress, making it possible to generate contents with greater naturalness than ever before. However, the appearance of unnatural (globally inconsistent or unstructured) parts to the content, such as faking, plagiarism, or hallucination, has become a problem. In this study, we constructed a probabilistic model of song development to identify unnatural melody. In order to learn natural song structure, we built an n-gram model of the bars. We characterized the bars with the results of analysis by a cognitive music theory: the Implication-Realization theory. Furthermore, using the proposed model, we analyzed melodies in the dataset and melodies that were partially regenerated by AI, and compared the results. The results suggested that bar entropy transitions reflect phrase repetition or structure and useful for discriminating between human-composed and AI-regenerated parts.