> I recently saw a video discussing a researcher who published a fake scientific article about a fictitious disease, with bogus author names, even a warning IN the article itself that stated "This is not a real disease, this article is not real" (paraphrasing) but still AI ended up picking up this article and serving information from it as if it was a real disease.

Isn’t a lot of pretraining done by chopping sources up into short-context-window-sized pieces and then shoving them into the SGD process? The AI-in-training could be entirely incapable of correlating the beginning with the end of the article in its development of its supposed knowledge base.