What we are seeing in maths now is that AI ties up loose ends in the literature, using its advantage in memory capacity and knowledge and the ability to quicky go through many possible conjectures and examples. There is a lot of this unused potential in the scientific literature in other fields as well. It has gone unused simply because there is too much literature for any human to digest. This will in the short run lead to a lot of progress in many disciplines and scientists will complain that it's not "real science". The open question is whether there is anything about "real science" that is not just building on prior knowledge over and over. An example that people like bringing up is the question of whether AI could have done what Einstein did, given all the scientific literature prior to Einstein. I think the answer is "yes" if it had been prompted suitably, namely to look for inconsistency and resolving them. The reason is that Einstein lived at a time when the biggest pieces were already in place. Maxwell's equations, Lorentz transformation, Minkowski-space, the Riemann tensor. There's a reason these are all named after other people: Einstein did not invent them. Einstein's genius was to put the pieces together. I don't mean to belittle Einstein's achievement, but by way of deduction it was relatively (hah) straight-forward. Indeed I would say that Einstein's genius was mostly in taking his own conclusions seriously and moving forward, no matter how odd relativity seemed at first. The situation is considerably more murky for quantum mechanics imo because in this case it was really sorting through a mess of data and trying to make sense of it and inventing a mathematical apparatus that previously just didn't exist. All this happened parallel on different routes by multiple people until a kind-of consensus emerged. Personally I think that whether AI could reproduce quantum physics would be a much harder test.