Neural Networks, which are the base technology of what nowadays gets called AI, are just great automated pattern detection systems, which in the last couple of years with the invention of things like adversarial training can also be made to output content that match those patterns.

The simpler stuff that just does pattern recognition without the fancy outputting stuff that matches the pattern was already, way back 3 decades ago, recognized at being able to process large datasets and spot patterns which humans hadn't been able to spot: for example there was this NN trained to find tumors in photos which seemed to work perfectly in testing but didn't work at all in practice, and it turned out that the NN had been trained with pictures were all those with tumors had a ruler next to it showing its size and those without tumors did not, so the pattern derived in training by the NN for "tumor present" was actually the presence of the ruler.

Anyways, it's mainly this simpler and older stuff that can be used to help with scientific discovery by spotting in large datasets patterns which we humans have not, mainly because they can much faster and more easily trawl through an entire haystack to find the needles than we humans can, but like in the "tumor detection NN" example above, sometimes the patterns aren't in the data but in the way the data was obtained.

The fancy stuff that actually outputs content that matches patterns detected in the data, such as LLMs and image generation, and which is fueling the current AI bubble, is totally irrelevant for this kind of use.