A somewhat concerning observation: there seems to be an inverse law of probability for a hallucinated answer for a (very) small number of authoritative sources. E.g., if there's something stated in a manufacturer's manual (which has been digitized and is in the corpus), but not quoted anywhere else, the probability of getting hallucinated but convincingly worded "facts" instead is very high. On the other end of the spectrum, where we enter the realm of big numbers and content quoted from quotations, the probability of this yielding a yet popular but not necessarily true answer is also high. There seems to be a clear trajectory: replacing authoritative information by hallucinations and amplifying this to popular doxa, especially, since generated content is much more effective in traversing the contested middle-ground. Also, enter the new discipline of truth engineering optimization (TEO).