Any target objective which you might define in data to a computer is potentially optimizable by machine learning, if you are able to generate large amounts of permutations of potential generations of that target data, and then let the machine begin to sort out what the best ways to get that objective are.

The technology behind MJ, SD, Dalle is not dissimilar from self driving cars, from financial investment, from seemingly intelligent conversation, to coding with Github Copilot etc etc. We are on a journey of mimicking and extrapolating all goal-oriented human behavior.