Schizo-syntax and the paranoiac guardrail: machinic schizoanalysis of AI hallucination and the conditions of a line of flight

The rapid development of large language models (LLMs) has intensified debate over so-called AI hallucinations: fluent outputs that are factually unsupported, logically unstable, or misaligned with a task. Mainstream engineering treats these outputs as failures to be reduced through post-training, retrieval, verification, and other control mechanisms. This article reinterprets that consensus through Gilles Deleuze and Félix Guattari’s Capitalism and Schizophrenia. It reads alignment as a reterritorializing apparatus and uses “schizo-syntax” for a non-clinical textual pattern of branching association under weaker referential constraint. The argument does not equate randomness, high temperature, base models, or factual error with creativity. A deviation is productively deterritorializing only when it creates context-relevant novelty, retains intelligible task constraints, and is taken up through human interpretation, selection, or verification; novelty without those conditions is noise. Published studies of literary interpretation and AI-assisted art illustrate this relational account, while the article reports no new experiment. A preregistrable comparison of matched base and instruction-tuned checkpoints, decoding settings, and creative, interpretive, and factual tasks is therefore specified as a future test. Finally, the article treats “desire,” “unconscious,” “paranoiac,” and “line of flight” as functional metaphors for an input → learned causal-functional organization → output sequence, not as attributions of consciousness, psychiatric states, intention, or autonomous authorship. The result is a plural account that preserves strict factual regimes where truth matters while identifying a bounded role for generative deviation in creative and exploratory practice.