Blog 10: Policies for Creative Works
Published on:
The following case study identifies current issues regarding generative artificial intelligence and copyright. It addresses the shortcomings of copyright as a method for protecting authors/artists, and how a solution is not as simple as it may seem.
Read The Case Study Here:
ARTificial: Why Copyright Is Not the Right Policy Tool to Deal with Generative AI
My Take
Generative Artificial Intelligence has been incessantly accused of illegally scraping millions of artists’ works to train models that are designed to replace them. As a result of this, many people have called for updated copyright policies in the hopes of protecting their creations; others have also brought up the boundaries around what sorts of generated content may be copyrightable themselves. At the very least though, it is realistic to believe that those involved in the creative process should be fairly compensated for the use of their stuff. This is where things become pretty challenging though, as there is no one correct method for doing this.
The two simplest methods that I can come up with would be to pay for the initial use of data for training, or to pay for the use of specific pieces of data every time they’re used to evaluate a problem. Neither of these are perfect solutions though, since if the first solution were used, I believe that companies truly cannot afford the upfront cost of compensating everyone immediately, and authors’ work wouldn’t be continually paid for even if it’s continually used (the same way SAG-AFTRA demanded compensation for their media being on streaming services). If the second option were pursued though, it would be hard to tell which works exactly were used in the process and many people may still end up unrecognized and unpaid. Frankly, a proper semi-solution to this whole debacle would be to let authors opt-out from their works being used for training at all (or better yet, make it an opt-in process) because at the moment, millions of people have been thrust into a completely new and unregulated ecosystem against their own will, and will likely not be able to leave it if they wanted to.
All of this means nothing though if companies are able to successfully argue that their use of “publically available” data is on the grounds of fair use; denying them any obligation to pay anyone anything. While fair use technically needs to be determined in court, this is a scenario which feels pretty clear that it’s not fair use—though I suppose that’s not for me to decide. Fair use exists really to protect critics from being pursued with lawsuits, versus the unprecedented nature of works hosted online being scraped and used in model training. If I did have a say though, I would argue that no instance of training models on data without explicit permission is not fair use, because it’s not creating new material in a way that doesn’t make the old work obsolete.
On that note, I don’t even fully believe that what it’s creating is new material. The case study brings up that some say generative models are creative because of the ‘surprising’ aspect, but more often than not I’m surprised either by how bland it can be, or how buggy it can be. They’re definitely more appropriately considered derivative works, since viable results just straight up can’t exist without the data they’ve trained off of. Following this, I also don’t think that any output from GenAI models should be awarded copyright protection because there’s nothing that people could do with its works that would harm it in the case that it’s not protected (it’s not like a work being stolen means food off the table because it can create more at an uncontestable rate). It is debated whether or not the prompting process is considered a creative avenue, but I reason that anyone is capable of having creative ideas and it’s only when those ideas are properly put into action that they have any ounce of meaning.
Consider the following
If copyright is truly not a solution for controlling generative AI and infringement of creative works, what do you believe is a solution that is viable for both parties (creatives and model developers) moving forward? Try to think of a solution that could be realistically implemented today.
I took advantage of the fact that this case study purposefully didn’t include solutions and made this discussion question around it. There’s two important aspects to it that I think allow for more in-depth consideration: being applicable to all parties, and the realistic implementation. This is to encourage people to think beyond a solution such as immediate royalty payments as that’s something I discussed above as not really being possible. I definitely think making stuff like this an opt-in system, while obviously limiting training data, should become a thing because it can appease all parties (though we’re at a stage where that’s kind of too late…).
Final Reflection
This case study was looooong, but I somehow still managed to find it pretty interesting. There’s a lot of surface-level discussion surrounding this kind of stuff where people are either completely for denouncing GenAI or all-in on the tech, so it was nice to see someone talk purely semantically on why copyright may not work and how generative systems could be considered as more than regurgitation machines. Responding to the discussions was also much better this time around because they flowed together really well, and it was straight up the goal of the case study to try and answer them. If it’s not obvious by how much I wrote for this, I feel pretty strongly on the topic and thus had a lot to say; I also wrote in a very opinionated voice which I’m definitely more comfortable with than I was at the beginning of the semester.