I disagree.The tech for music, language, and images is fundamentally the same. It requires consuming a huge body of other works and imitating that with variation.
Just as it doesn’t just randomly assign pixels, which given infinities would produce a great work of art but in a world with limits would make colored visual static, randomly creating sound waveforms from the basis of pure tones (which is how all sound is fundamentally made) will also in practice only ever create noise. Literally radio static fuzz. Given infinities, music will emerge, but there are not infinities. It’s not even AI if it’s just doing it randomly, anyway.
All AI music is training on real music and then fitting to stylistic convergences exactly the same as visual art. It is equally theft or not theft.
You could tell AI to use random sounds as starting points and then proceed in specific patterns that would produce more or less palatable sequences.
This would be much harder to do with text, but still somewhat possible to randomly cram random words into predefined positions for "verbs" and "nouns" separately.
And this would be impossible to do with pixels without feeding AI any actual ready picture drafts, which unlike music are NOT made via mathematic sequences.
This is the crucial difference:
Drafts for music and text are "rules" that only define TYPES of input data (closeness of sounds to each other, and word categories).
But pictures DO NOT follow any such "rules" that could be defined on the PIXEL level - if you want anything more complex than simple geometry shapes (which ARE math based).
This still doesn't mean that it's easy to do any of the first types - most of the results would still be garbage, but it's possible to randomly come upon something that SOMEONE would like.
Whereas in the case of pixels, it's nearly impossible to program a pattern that WOULDN'T already "steal" basic shapes (as "elements") or more.
Yet that would still only account for "kindergarten" level of "skill" - simple lines in schematic patterns.
But if you expect AI to be capable of painting even a simple apple in any "realistic" shape and color tone from scratch - yeah, nope.
Even be allowed to use "basic geometric shapes" would NOT help it to do so, lol.