This study evaluates the capability of Large Language Models (LLMs) to perform Peircean semiotic analysis on contemporary dance, using the choreography "Shadow Me" as a case study. Using a panel of six AI models—Claude 3 Opus, Claude 3.5 Sonnet, Gemini 2.5 Flash, Gemini 3-flash-preview, GPT-4o Mini, and GPT-5 Mini—we compared their ability to identify Iconic, Indexical, and Symbolic signs across 24 performance images under two conditions: with and without choreographer's context. Our results demonstrate that while LLMs show strong proficiency in Icon identification (~2.8/4.0), their performance in Indexicality and Symbolicity is highly dependent on contextual input. Notably, Gemini 3 Flash-Preview achieved the highest scores in Symbolicity (3.75/4.0) and Indexicality (3.38/4.0) with context. The study reveals a significant "context gap," particularly in symbolic interpretation, which improved by up to 22.5\% when artistic intent was provided. These findings suggest that while AI can systematize surface-level visual information, its interpretive depth remains tethered to human-provided context, highlighting a functional gap between pattern recognition and embodied artistic understanding.
Ji Hee Lee (Thu,) studied this question.