The Phenomenological Black Box of Large Language Models

This paper reframes the “black box” of large language models (LLMs). Rather than treating opacity as a purely technical problem of hidden weights and inscrutable mappings, we argue that the deeper black box is phenomenological indeterminacy: the principled impossibility of predicting the long-term narrative–symbolic dynamics that emerge in sustained human–AI dialogue. A playful request—generating an action-figure of the user—became a probe that revealed persistent memory traces, slow-forming narrative arcs, and emergent symbols that act as semantic attractors. We describe their recursive feedback loop and motivate a shift from static audits to longitudinal phenomenology as an ethical and methodological necessity. This stance complements surveys in XAI/LLM interpretability (Hassija et al., 2024; Luo & Specia, 2024) and aligns with phenomenological accounts of user experience (Heersmink et al., 2024) and hermeneutic risk (Rebera et al., 2025).