AI can help us predict properties of matter within the "bubble of similar systems". That means that if you want to predict these properties and plenty of similar systems, let them be molecules or condensed matter, have already been simulated, you could skip the computationally demanding QC calculation and let AI predict the properties by analogy conclusion.

On the other hand, if there are little to no references with sufficient similarity, the AI algorithm doesn't have enough (or anything) to "machine learn" from. In that case you will probably encounter the same thing you get when you ask ChatGPT about something about which it doesn't have any backup learning: it will autofill the voids with random stuff that sounds right, but is quite often blatant nonsense.

Complementing Jürgen Weippert's answer, there is also the possibility of using machine learning and similar algorithms to create shortcuts for quantum chemistry (QC) calculations. By this, I mean developing improved functionals and approximations through the generalization of benchmarked results. Such approaches can—and likely will—enhance computational performance and possibly even accuracy, but at the expense of a fundamental understanding of the underlying physical nature of the system.

Is it worth sacrificing knowledge to create black-box models that yield good results? This is a question we have been grappling with for decades.