The claim that the AI has been trained on copyright material has been verified by a number of people. Although Udio and other companies have made it difficult to ask for a song in the style of a specific artist, people have been able to write prompts to get around that limitation and get ABBA's harmonies, or Paul McCartney's voice.

It's the same sort of thing that happened when AI image programs didn't let you write "Mario", so people used "video game plumber" and got... Mario.

When an AI program is "trained" on a song, that song is stored in the neural network. Even if it's a fairly coarse copy, it's still enough of a copy that, if given the rights prompts, the original song could be reproduced.

AI programs aren't trained on just a handful of songs. They're fed millions of songs. That's why they are able to generate such a variety of material.

But you can feed in the works of The Eagles, Tom Petty, Chicago, CSNY, Supertramp and The Beatles and generate songs all day long, and you're still not going to get Michael McDonald's throaty vocals, or the chugging guitar on "Long Train Runnin'".

-- David Cuny
My virtual singer development blog

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