[mjr00]

> The human will be acting at least as a curator, because not everything created by AI is going to be pleasing, so some selection and catering to human taste will be required.

Yes, and keep in mind humans are already doing this! It's very common to do tweaking of knobs on a synth/VST while recording and create a 10-20 minute audio file, commonly called a bass jam or mud pie, then select the best bits to use in a song. And of course, people use randomization tools to tweak the knobs for them. IMO use of AI to support this type of workflow is far more promising than going directly to the finished product.