Somewhat tailored response per your username: Anything the ministry of truth decides should not be part of the training set.

But more generally, even in things like neurodivergence, lots of questions you could ask that would be in the bounds of "guesstimation" by professionals. The problem is these LLM will spout pure utter highly-believably-confident snake oil with the tiniest little legal microscopic fineprint on the side of the html webpage that you can never scroll to saying "AI responses can be wrong".

For me, the mere fact that you have a nicely worded source of information that, depending if you're lucky with your question, is pure scientific fact, or utter maddening bullshit, makes me terrified to use it.