@mpatti@andrewscottbell@udiomusic@UMG@sonymusic@PaulMcCartney@yokoono@cinesamples Hi Mike, great description of how the NN approach works. Copyright re: training - in US, courts haven't made their position clear as yet, but may rule that this is 'fair use'. Copyright on outputs: unlikely to apply? in which case works could be essentially public domain.
@ednewtonrex@steph_palazzolo In the US, training on copyright works may be considered "fair use", depending on context, and the courts have not established a firm position on this as yet.
Is that widely understood in the creative community?
@ednewtonrex@Rahll@steph_palazzolo The more novel & diverse the outputs of Gen AI the more a synthetic data approach can be sustained.
Intuitively, I'm pessimistic - I envisage a drop of some kind of creative quality with repeated iterations and the eventual need for new human created works... my pure speculation!
@KevinSargentMus Or - they just focus on the big artists who are more likely to sue (and have a stronger chance of winning)
The rest wouldn't be much if a worry
@KevinSargentMus I imagine there is a database of artists they work from that is inherently limited... where does one find a database of all the music artists/copyright holders in the world?
(genuine question, maybe its doable in some way)
@ednewtonrex Training models on "synthetic data" is already common practice. It could be a way out of the copyright/TDM issues for Gen AI developers but there are some obvious potential problems with that
@ednewtonrex If enough AI music enters the market, the next generation of models will be (in part) trained on.... AI music ๐ณ
So where does that lead?
@ednewtonrex If enough AI music enters the market, the next generation of models will be (in part) trained on.... AI music ๐ณ
So where does that lead?
@RichCanavan@michael_cryne In my view this is unlikely to happen because there isn't a clear lawful basis for copyright in AI outputs.
Much more likely that customers will switch to new AI based services which generate unique, bespoke content, with no copyright claim.
I think?
@abstraktius I agree with the person who replied basically saying good/bad is a matter of opinion. Follow the joy in your art! (and take on feedback ๐)
@crispinhunt We are in a mad world where natural human language is the new language of software development so who knows!
I was wary of making this a semantic issue. My aim is "know thine enemy"* - the better we understand its working the better we can legislate โ
@crispinhunt re: collage - my understanding is that neural networks don't work like that. Instead, during the training process they gain a kind of subjective impression of the data. They then generate outputs based on this "impression". rather than a collage from chunks of the original data.
@crispinhunt Thanks for both of those references, Crispin. I had not heard of SOMMS AI. I am writing about GenAI as part of my legal studies and that is a useful reference point.
@BenLissen@crispinhunt Surely this is good news: if GenAI works are not copyrightable then I think that is likely to very much protect us copyright holders! But I would welcome other perspectives... I am a student of Law and not an expert ๐