AI models hallucinate because they are probabilistic engines designed to predict the next likely word. They don't know what "truth" is, they just say what is "plausible-looking" based on training data. Some numerical inaccuracies in the model may also contribute to "hallucinations".

Humans hallucinate under sensory deprivation because, I think, our brains never evolved to respond reasonably to full sensory deprivation (it basically almost never happens in real life to healthy individuals who leave offspring), i.e. this kind of scenario was never "tested" by evolution and we get results that make no sense (random garbage).