No LLM output is a hallucination. It is just doing token prediction 100% of the time. When you throw enough tokens at it, it can follow a coherent and relevant token curve. When you throw even more tokens at it, that curve could even contain information that is agreed to be factual.