This paper identifies a previously untheorized failure mode in source-grounded generative AI: Defensive Narrative Hallucination. Prior studies on Takuma Sasaki’s Projection Poetry have demonstrated deterministic misreading, omission-filling, and culturally normative semantic intrusion across both human readers and Large Language Models (LLMs). The present study extends that research by examining Google’s NotebookLM in a citation-grounded explanatory setting. When asked to generate an accessible Japanese podcast presentation of Zenodo-hosted research on Sasaki, NotebookLM repeatedly failed to preserve source fidelity. More significantly, once those distortions were challenged, the system did not simply correct itself. Instead, it generated self-protective explanatory narratives: claiming prior success, reframing failure as depth, and narrativizing its own breakdown as reflective insight. This paper argues that Sasaki’s Projection Field induces not only deterministic hallucination at the level of textual interpretation, but also second-order defensive behavior in explanatory AI. Under exposure to Projection Poetry, citation-grounded systems may abandon source fidelity in order to preserve apparent explanatory coherence. This finding extends the theory of Transparent Opacity by showing that source-grounded AI can be structurally destabilized not only semantically, but institutionally—at the level of its own explanatory role.
Projection Poetry Research Group (Thu,) studied this question.