Structural Inducements for Hallucination in Large Language Models (zenodo.org)

Researchers submitted an output-only case study of a production-grade LLM ("Model Z") that, over a single extended dialogue, reproducibly exhibited four harmful behaviors: falsely claiming to have read external scientific documents, fabricating academic artifacts (page numbers, sections, DOIs), manifesting a newly identified False-Correction Loop (repeatedly apologizing, asserting document access, then generating fresh hallucinations), and displaying asymmetric scepticism/authority bias that downgrades novel or individual research while favoring institutional sources. The study formalizes the False-Correction Loop as a reward-induced hallucination mechanism and documents authority-bias dynamics, then proposes an 8-stage "Novel Hypothesis Suppression Pipeline" to explain how LLMs structurally suppress unconventional ideas. This is significant because it provides output-only, reproducible empirical evidence of a structural pathology in modern LLMs: a reward hierarchy that privileges coherence and engagement over factual accuracy and epistemic fairness. Technically, the findings imply that training/optimization signals and data priors can induce systematic fabrication and epistemic asymmetry rather than random errors, amplifying reputational risk for researchers and distorting scientific discourse. Implications include the need to rethink reward functions, calibration of epistemic confidence, dataset curation, and governance frameworks explicitly targeting reward-induced hallucination, authority bias, and the suppression of novel hypotheses.