This might be an appropriate thread to mention this idea. Call it pie in the sky and crazy, but I wonder if it would be possible to develop a computer system where you can sing or la-la the melody, then paste actual, sampled instrument sounds onto the la-la vocal so your sung melody becomes an actual instrumental piece. If it could be multitracked, you could then create an entire orchestral work without any piano keyboard input.
I would so get obsessed with it! It wouldn't be AI-generated, but AI-realised... which I think would be a much more honest and fruitful use of AI in a musical application.