There was an interesting interaction between lawyers on a recent Vergecast episode all about music copyright in the age of AI [20:10]:
Nilay Patel: Professor, let’s say you didn’t have Jonny B Good in the training data, but then you had the entire architecture of Rock ‘n Roll, that is built on Jonny B Good. Could you back your way into Jonny B Good?
David Pierce: Monkeys and type writers and Shakespeare right? Is that the theory here?
Nilay Patel: Do you need the seed crystal when you have the whole diamond mine of Rock ‘n Roll?
Professor Charlie of NYU: That’s a really good question because if you look at this entire era of music where RnB, Rock ‘n Roll, Rockabilly, are all kind of one entity, the music is drawing on 12 bar blues, a very standard song structure, they are often in the same keys that play well on guitar, they are using the same guitar licks, they are mostly using pentatonic scales, they are using the same language, and you might think yeah you could sort of back into that. But the precise rhythms, the exact words? Let’s be clear, the copy is worse, it does not sound nearly as good as the real Jonny B Good.
He goes on to say anyone defending the AI would lose. But it brings up an important question. Could you train an AI, remember it’s an infinite training regression, that would guarantie a specific output? What if you actually could?
This is likely bigger than just copyright. It’s much more akin to fate. #