For the past two years, the AI industry has repeated a reassuring story about hallucinations. When models get things wrong, the explanation goes, it’s because they were trained on imperfect data. If the model invents facts, misnames entities, or fabricates steps, the fix is assumed to be better sources. Cleaner data. Tighter grounding. Retrieval-augmented generation. Closed corpora. Upload the documents, constrain the model, and hallucinations should disappear.
This belief is so pervasive that even experienced practitioners catch themselves reverting to it. I do. Models do. Entire safety strategies are built around it. It is embedded in marketing copy, investor decks, procurement frameworks, and regulatory conversations. Hallucination, we are told, is a data problem.
It isn’t.
The recent surge of failures in agentic AI systems makes this impossible to ignore. These are not casual chatbots producing factual trivia errors thanks to outdated training.
These are systems with tools, memory, planning loops, and access to verified internal documents. They retrieve from approved sources. They cite. They reason step by step. And yet they still hallucinate — persistently, confidently, and often catastrophically.
The industry response has been to reach reflexively for the same explanation. Perhaps the agent retrieved the wrong document. Perhaps the embedding missed something. Perhaps the training data was outdated. Perhaps we need better grounding.
But what if the presupposition itself is wrong?
Source-grounding is supposed to prevent hallucinations by constraining what the model can draw from. The assumption is simple: hallucinations occur when models lack the right information, so supplying verified information should eliminate the problem. This logic feels intuitive, which is why it’s been so widely accepted. But intuition is not evidence, and when tested under controlled conditions, the explanation begins to unravel. Our new Zenodo paper, Source-Grounding Does Not Prevent Hallucinations: A Controlled Replication Study of Google NotebookLM, tests and validates this.
The critical mistake is treating hallucination as a failure of knowledge access rather than a failure of semantic control. In practice, models rarely hallucinate because they don’t “know” something. Far more often, they hallucinate because they know too many plausible things at once and lack a mechanism to decide which one is allowed to dominate.
This is not a training data issue. It is not a retrieval issue. It is an authority issue.
When a model is faced with competing interpretations — of a word, an entity, a goal, a plan, or a prior decision — it must assign authority to one and suppress the others. When that authority cannot be maintained or revoked cleanly, the system does not stop. It does not say “I can’t proceed.” Instead, it repairs coherence by inventing details that make the chosen interpretation seem plausible. That repair is what we call hallucination.
This is why hallucinations persist even when the correct information is present. This is why they appear in systems that have perfect access to documents. This is why retrieval does not solve them. Retrieval simply supplies more candidates for competition.
Agentic systems expose this failure mode more dramatically because they stretch it over time. A misinterpreted goal early in a plan is not corrected later; it is rationalized. A tool output that conflicts with an assumed interpretation is not allowed to override it; it is re-explained. The system preserves internal coherence at all costs, even when that coherence diverges from reality.
The same thing happens with proper nouns, technical terms, and domain-specific entities. The model does not hallucinate because the entity is missing from training data. It hallucinates because the entity fails to retain semantic authority across context. Nearby, more statistically common meanings bleed in. The system repairs the resulting conflict by substituting something that sounds right.
This is why the training-data explanation keeps failing under scrutiny. If hallucinations were caused by missing or unreliable data, then constraining models to verified sources would eliminate them. But when models hallucinate despite having the right data in front of them, the explanation collapses.
Source-grounding constrains retrieval. It does not constrain interpretation.
RAG systems retrieve passages. They do not decide which meaning must persist when passages disagree. They do not encode which entities are load-bearing and which are incidental. They do not provide mechanisms for semantic revocation when an earlier interpretation is invalidated. They assume that relevance scoring is sufficient. It is not.
The uncomfortable implication is that much of the industry’s confidence in RAG as a hallucination solution is misplaced. This does not mean retrieval is useless. It improves relevance. It reduces outright fabrication. It makes outputs easier to audit. But it does not address the core failure mode that produces hallucinations under ambiguity.
That matters enormously for enterprise AI.
If hallucinations are governance failures rather than data failures, then scaling data pipelines will not fix them. Improving embeddings will not fix them. Uploading more documents will not fix them. Worse, these interventions can actually make the failures more convincing by giving the model more material to rationalize against.
This is why agentic AI feels simultaneously powerful and unstable. The systems are capable. The knowledge is there. The failure is structural.
We need to be aggressive in dismantling the myth that hallucinations originate in training data because that myth is actively slowing progress. It sends researchers down the wrong path. It reassures buyers falsely. It frames governance as an afterthought instead of a core architectural requirement.
The right question is no longer “how do we ground models better?” It is “how do we give models the ability to assign, maintain, and revoke semantic authority?”
Until that question is addressed directly, hallucinations will persist as a fundamental property of how these systems preserve coherence under pressure.
Agentic systems did not create this problem. They revealed it.
And source-grounding does not solve it.
