Should Machine Learning be Sensitive to Copyright Infringement?
Authorship and Compensation
I believe authors of content should be compensated for the use of their data in training. Compensation is a difficult issue because of the magnitude of information that is used for training. Compensation can’t and shouldn’t occur at the output level because in most cases it is very difficult to identify what information the model is using. Unless you specify “in the style of” it is using a bunch of different content that it can’t direct you to. Compensation should occur at the training step because of this. Maybe creators should be compensated similarly to music artists on streaming services.
To be honest, I’m torn on whether the outputs should be considered derivative works or not. I think it partially depends on whether the creations are autonomous. I think if there is a significant amount of prompting involved, then you might be able to consider it fairly original. Every piece of artwork is based on something and many are inspired by other artwork. Many artists in history emulated each other’s styles and that doesn’t make it infringing. I do think though that if you are giving AI a general prompt and then claiming it is your art, it isn’t really. Art requires thinking, and AI can not think.
I think training with unlicensed works technically constitutes fair use under the current law. Since the Supreme Court has ruled that copyright protection should not depend on the merit of the artwork and only on whether it is factual information, I think AI generated art technically counts. Whether it should is a different issue altogether. Maybe the fair use laws need to be updated in light of new technology.
Final Thoughts
A question I would ask after this article is, “Should the current copyright law be updated to include AI?” Whether or not you believe AI is art, I think the current description of copyright is an insufficient framework to evaluate AI generated art from.