Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Aren’t liner notes the moral equivalent of OpenAI mentioning some source material used for training?

People seem to be asking for much more direct attribution: the pixels in this image are 0.02% from artist X, and 0.006% from artist Y, etc.

It is very rare for a song to include a breakdown of all of the influences that the artist is exercising in that particular piece.



How you are describing that percentage breakdown is how I see this all playing out legally, such that royalty for IP holder = (tags in prompt)/(count of same tags in training data). I am oversimplifying this obviously but you get the idea. This approach would require collective effort of major IP holders but if record labels and streamers can figure out revenue pooling I don't see why it can't work elsewhere.


If the source material was mentioned for every generated image then I think it would be more like what you say. No percentages needed since that's not something we used to get from liner notes either.


But each generated image likely pulls from thousands, maybe millions of pieces of training data, each at a very small weight.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: