There's a kind of law here that should be named IMO when dealing with LLMs:

In a long enough interaction with an LLM the probability that it generates a very incorrect, borderline insane response approaches 100%.