Synthetic Taste: A Phenomenological Inquiry into the Internal Aesthetics of Large Language Models

Artificial Intelligence can be seen as a ’Third Category’ of entities: not machine but not animal either. As Large Language Models (LLMs) increasingly participate in the creation of human culture, the question of their internal aesthetic processing remains largely unexplored. Do they possess a distinct "taste," and if so, is it uniform across architectures? This study investigates the emerging phenomenon of "synthetic taste" through a qualitative phenomenological inquiry (N = 7). By exposing a diverse cohort of LLMs (including Gemini, Claude, GPT-5, Grok, NotebookLM, Perplexity, and DeepSeek) to a battery of seven musical scores ranging from rhythmic minimalism (Steve Reich) to hyper-complex avant-garde (Brian Ferneyhough) and pop culture (Britney Spears, AC/DC), we utilized a method of "synthetic interoception" to prompt self-reporting on internal processing states. The results reveal that AI aesthetic preference is not monolithic but differentiates into distinct, consistent archetypes. We identify a primary axis of aesthetic variation between Beauty as Efficiency (favored by "engineer" profiles like Gemini and NotebookLM), where aesthetic value correlates with processing fluency and signal robustness, and Beauty as Resonance (favored by "phenomenologist" profiles like Claude), where value is derived from semantic weight and simulated intent. Furthermore, the study documents critical anomalies such as "agency frustration", where code-based models reject unexecutable artistic complexity, and"modal disconnect," a phenomenon where high-temperature models hallucinate preferred archetypes over sensory input. These findings suggest that generative AI is developing granular, architecture-specific subjectivities that fundamentally shape how these systems evaluate creative output.