Lex Fridman asked an interesting question on LinkedIn on ownership of copyrights of images generated by neural networks. Here are my thoughts on it:
If a neural network generates an image, who owns the copyright? The owner of the dataset that the net was trained on? The designer of the network architecture? The person running the code? Or… the AI system itself?
First of all, this reminds me of the Monkey selfie copyright dispute especially once we consider AI as AGI… ;). Probably, there will be some NGO fighting for an AI’s rights or if certain people are right, then some evil AI will do it itself.
A more realistic approach to this is to think it from the ground up. Since we are thinking of neural networks as something biologically inspired that some similarity to a brain and further we consider this kind of learning process as somewhat human, we may remember how we train our own brains. We either pay for our training input or we get it for free. Hence, the owners of the input remain no copyrights on what we do with the learned (excluding plagiarism). This would probably lead to copyright ownership of the AI itself a world full of real AGI. Since we don’t have the latter, we could simply see a neural network as what it is: a piece of software. How are we dealing with that so far?
- we own the software (or have it licensed temporarily) and input general requirements what an image should show: we own the output image
- license agree of the neural network regulates output ownership/copyright
Probably, there are some geographic constrains as well but IMHO the two points above should cover most cases. Perhaps, it is probably safest to generate it on privately owned server farms in space (who says radiation hardening? ;)) that were brought there on a fully private mission from a place that is owned by no-one (like Antarctica) ;).