Dynamic Derived Intentionality in Large Language Models: From Implicit Beliefs to Semantic Parasitism

Large Language Models instantiate a distinctive category of intentionality, which I call dynamic derived intentionality: a productive, context-sensitive form of derived intentionality that goes well beyond the static cases of words on a page or symbols on a map, yet falls short of original intentionality. What separates the two is neither consciousness nor introspection nor anything in the biological substrate, but normative standing. Original intentionality is the having of what I call ownership of normative status, the standing of a subject who occupies the space of reasons and can be held answerable for its commitments; LLMs lack it. The analogy I draw with human implicit beliefs is diagnostic rather than probative: it holds fixed the representational vehicle the two share, sub-symbolic, distributed, statistically acquired, opaque, so that the one property they do not share, answerability, stands out. This is the answerability gap. I develop the notion of semantic parasitism to capture how these systems operate through a meaning they do not own, and read Reinforcement Learning from Human Feedback as calibration to norms that reside elsewhere. The account is defended against neo-Searlean, anthropomorphist, and deflationary objections of both a Dennettian and a McDowellian cast. The gap it identifies is architectural, not metaphysical: an architecturally discontinuous system could in principle come to occupy the space of reasons.