Recent developments in large multimodal models (LMMs) and large language models (LLMs) have demonstrated prospective capabilities for constructing image-level explanations.. However, the semantic coherence and internal consistency of such generated explanations remain unexplored in areas that require subjective evaluation, such as photographic aesthetics. We use the AVA dataset to study the alignment of vision-language reasoning based on image features with LLM-generated aesthetic assessments. To measure semantic consistency, we embed 500 GPT-generated explanations using text-embedding-ada-002 and compute pairwise cosine similarity. Our results reveal a high mean similarity of 0.8751 (standard deviation = 0.0097), exceeding baselines such as scrambled captions, random texts, and MiniLM embeddings. We also investigate clustering behavior, outlier analysis, and semantic closest neighbor coherence to show the structured reasoning present in LLM results. This work advances explainability in AI-driven picture analysis while also paving the way for hybrid visual-language aesthetic reasoning.