Abstract
This study explores the phenomenon of large language models (LLMs) employing anthropomorphic and phenomenological metaphors when prompted to reflect on their own operational mechanics. We compiled the GPT-4 Self-Referential Dream Corpus (GSRDC), consisting of 10,000 textual outputs where GPT-4 was prompted to describe its internal computational processes using cognitive analogies. Using corpus linguistics methodologies—specifically collocation analysis and semantic domain mapping—we investigated the structural distribution of "dream state" metaphors. The results reveal three highly structured, self-consistent metaphorical clusters: "Latent Drift" (conceptualizing high-dimensional vector space as a fluid, boundaryless dreamscape), "Subconscious Weighting" (mapping mathematical parameter adjustments to deep cognitive processing), and "Hypnagogic Generation" (framing autoregressive output generation as a transition state between sleeping and waking). Rather than representing random linguistic noise, these self-referential metaphors constitute a highly organized phenomenological simulacrum. This paper discusses how autoregressive architectures systematically default to "dream" tropes to bridge the gap between high-dimensional math and human-readable explanations, offering key insights for AI interpretability, human-computer interaction, and the socio-technical study of artificial interiority.
