AI is Always Wrong, in Matters of Taste

Large language models produce technically competent but recognisably hollow output. We argue this reflects a structural limitation, not insufficient scale. We distinguish craft (competence within a domain, evaluable from inside) from taste (evaluation from outside a domain, requiring a metalanguage the domain cannot provide) and develop a three-level framework: the domain (formal properties of the artifact), human dimensions (body, social context, personal identity), and a philosophical foundation (terminal values governing tradeoffs between dimensions). Taste requires attending to contextual features no fixed feature set can enumerate in advance. This is the frame problem (McCarthy & Hayes, 1969) applied to evaluation. A supporting information-theoretic argument shows that text is a low-dimensional projection of high-dimensional embodied experience, systematically losing the sensory and situational dimensions along which evaluative judgment operates. Generative large language models are confined to domain-level operation: they process text, and text is a lossy compression of the context that determines quality. This confinement is architectural in current systems, though not necessarily permanent for all conceivable architectures. It manifests as hollow AI-generated content (craft without taste), hallucination (truth requires a metalanguage text cannot provide), and absent epistemic calibration (the model cannot sense the boundary of its competence). The majority of professional judgment consists of applying codified frameworks, which language models handle effectively. The irreducible human premium concentrates at the frontier, where the relevant contextual feature falls outside any predetermined framework.