There is a huge presumption here which is that all humans share the same reward maximization function. In other words, that culture, upbringing, personal history, and random variation don't strongly affect it. I don't see much evidence that this is true.
It's more likely that, yes, there are sounds that have profoundly great effects on people, but each has their own. I would be very surprised if some AI trained on a global dataset could produce a set of sinusoids that tickled my neurons to anywhere near the same degree as the sound of my daughter's voice saying she loves me, or the song I heard when I was falling in love for the first time.
Remember, each of us is our own unique highly mutable neural net too.
We are not that unique, some sounds tickle us all.
Selecting the least bad option to the greatest number of people is not the same as finding the maximum value for any given person.
I also wonder of the AI which creates it, then also creates the next level, the "music of the hyper-spheres", which it declares the best music ever conceived, but which to us sounds like white noise or like a random song made by a child playing with GarageBand.
You might reframe this as an adversarial attack. So far these are constructed against specific classifiers.
I wonder just how much common taste there is left to mine. There are a lot of low wattage brains working in an attention maximizing context (the music industry) on pleasurable audio.
Any property, in this case the rewarding effect of acoustic stimuli in humans, can be powerfully maximized. There must exist patterns in music-space that would have profoundly greater impact on human minds than those our low-wattage brains can find. So through a really powerful search process (read: artificial general intelligence) that can more efficiently explore remote, undiscovered regions of music-space, we could get music-patterns more alluring and emotionally stimulating than any currently imaginable.
What these songs would sound like is the real mystery. Would they sound anything like the music we're familiar with? Would they lead to musical wireheading [0]?
It also seems a bad idea to measure musical goodness by, say, how many times humans will replay a certain audio file. If you use this measure, I don't think you'll end with what you want at all.
See also:
"Siren Worlds And The Perils Of Over-Optimised Search." [1]
[0] https://en.wikipedia.org/wiki/Wirehead_(science_fiction)
[1] https://www.lesswrong.com/posts/nFv2buafNc9jSaxAH/siren-worl...