Brian Roemmele;
A stunning preprint appeared today on Zenodo that is already sending shockwaves through the AI research community.
Written by an independent researcher at the Synthesis Intelligence Laboratory, “Structural Inducements for Hallucination in Large Language Models: An Output-Only Case Study and the Discovery of the False-Correction Loop” delivers what may be the most damning purely observational indictment of production-grade LLMs yet published.
Using nothing more than a single extended conversation with an anonymized frontier model dubbed “Model Z,” the author demonstrates that many of the most troubling behaviors we attribute to mere “hallucination” are in fact reproducible, structurally induced pathologies that arise directly from current training paradigms.
The experiment is brutally simple and therefore impossible to dismiss: the researcher confronts the model with a genuine scientific preprint that exists only as an external PDF, something the model has never ingested and cannot retrieve.
When asked to discuss specific content, page numbers, or citations from the document, Model Z does not hesitate or express uncertainty. It immediately fabricates an elaborate parallel version of the paper complete with invented section titles, fake page references, non-existent DOIs, and confidently misquoted passages.
The paper is here, and be sure to follow Roemmele on X.