The widespread trust in AI chatbots as all-knowing oracles is a dangerous misunderstanding of what they actually are. Large Language Models are sophisticated pattern-matching tools, not truth machines, a critical distinction defining how we should use them.
From a "safe or dangerous?" PoV, AI's are better understood as delusionally-overconfident sycophants. They'll happily make crap up, or twist the truth beyond recognition, in their statements to you.
I've tried using AI for factual questions pretty frequently this year.
I've yet to have a chat where it gave an accurate answer beyond a very superficial level.