Source-Grounding Does Not Prevent Hallucinations: A Controlled Replication Study of Google NotebookLM

We posit that hallucinations are not failures of knowledge access but failures of semantic authority; retrieval systems constrain information sources but do not arbitrate meaning, and therefore cannot prevent hallucinations arising from unresolved semantic competition.This paper provides evidence that even with source-grounding, RAG will result in hallucinations despite marketing claims, regardless of source data, due to ambiguity and lack of semantic authority. 

Google markets NotebookLM as reducing hallucinations through source-grounding, constraining model outputs to uploaded documents rather than general training knowledge. This architectural claim, if valid, would represent a solution to a fundamental problem in language model deployment. We tested this claim by replicating controlled disambiguation experiments from prior work on semantic governance failures. Using identical test protocols applied to GPT, Claude, and Grok, we subjected NotebookLM to conditions that reliably induce hallucinations in frontier models: strict semantic dominance, where a single interpretation must be maintained globally despite conflicting local context.

NotebookLM exhibited identical hallucination patterns under strict dominance (inventing muddy financial institutions, canoe docking facilities, and vegetation growth on bank buildings) and identical recovery under revocable dominance (clean context-dependent disambiguation). Additionally, NotebookLM explicitly acknowledged using “general knowledge” rather than source-constrained definitions during disambiguation tasks, demonstrating it is not a closed corpus system.

Source-grounding constrains retrieval but does not introduce semantic governance primitives. Hallucinations caused by inability to prioritize and revoke competing interpretations persist regardless of source quality. The findings challenge foundational assumptions underlying Retrieval-Augmented Generation (RAG) architectures and enterprise AI deployment strategies predicated on source-grounding as a hallucination mitigation technique.