Machine learning often works remarkably well on problems to which it can be applied. However, for most problems of interest, there is no dataset and there is no simulator.
The ever-accumulating successes of machine learning notwithstanding, I think the technology is much less generally applicable (and hence less generally useful) than many others seem to think.
For example, I have heard several people say that the William Lowell Putnam Mathematical Competition is one of the next grand challenges for AI, and express optimism that machine learning will be able to solve it. I agree it is a grand challenge, but I don't even see how to apply machine learning to it, let alone how to solve it with machine learning. There is no dataset. There is no simulator.