There's a huge difference between these AI programs and BiaB, sample libraries, and MIDI tracks: the AI companies don't have the rights to use the voices and instruments that are in their songs.

They make of point of not telling where they get their materials from, but here's a project in 2020 that described how it gathered its source material (emphasis added):

Quote
To train this model, we crawled the web to curate a new dataset of 1.2 million songs (600,000 of which are in English), paired with the corresponding lyrics and metadata from LyricWiki.

See: https://openai.com/research/jukebox

That is, these programs trawl through millions of copyrighted songs to get their training material. No artist has granted them rights to use their materials, and no artist is compensated.

The companies creating the AI programs do a number of things to make it difficult to determine where the source materials come from.

One of the most obvious is omitting artist names in tags. So although the AI is capable of rendering a song with the voice/instrument/arrangement of a particular artist, there's no way to request the AI do so.

And since a specific voice can't be requested, you'll end up getting a voice that's a combines the attributes of similar voices - enough so that the original singer can't be identified.

However, here's something Udio produced when prompted to generate something in the style of the Beatles:

https://twitter.com/i/status/1778900566917165512

Udio didn't pay the rights holders of The Beatles songs to use their songs, and Paul McCartney didn't authorize Udio to use his voice.

In my mind, this is theft (well, technically massive copyright infringement), disguised the same way that money laundering hides the source of illegal profits.

-- David Cuny
My virtual singer development blog

Vocal control, you say. Never heard of it. Is that some kind of ProTools thing?