Generative AI, Artist Royalties, and Intellectual Property Protection
An Economics Take
Generative AI has sparked a heated debate about intellectual property protection and artists’ royalties. In case you’ve been living under a rock, here’s a brief summary:
Generative AI like Midjourney can learn an artist’s style and imitate it surprisingly well.
These AI are trained on images scraped from the internet, often copyrighted.
Using copyrighted art for training AI falls into a legal gray area
Artists’ wages and employment are hurt, as many potential buyers are using GenAI instead
Artists’ advocates are pushing for retaliation or compensation. Royalties for original artists is one common proposal.
Most internet discourse on the topic devolves into armchair legal arguments. I think the law is not a useful way to resolve the debate. Laws are arbitrary; they are defined by society, meant to evolve over time, and corruptible by moneyed interests. Instead, I will examine it from an economic perspective on how to maximize social welfare to determine what the law should be. While we’re using art as the key example, this model applies generally to all intellectual property.
Tl;dr:
Original art needs stronger protection, since generative AI harms the original creator’s value capture
However, actually implementing an attribution and royalty policy seems intractable
Moreover, most intellectual property protection systems seem easily corruptible and abused by the powerful
—
Broadly speaking, intellectual property development works like this: there is a large upfront research and development cost to creating a new technology, design, or art style. Once that new idea is created, it can be cheaply reapplied to improve things and generate value. Society as a whole benefits the most when that idea can be copied and reused freely for maximum impact. However, the upside usually goes to the copier rather than the original creator. If the creator is unable to capture sufficient value from their innovation, then they will not be incentivized to develop that new idea in the first place. Intellectual property policy needs to balance these two goals; incentivizing creation while enabling broad utilization.
This balance varies by industry, depending on the R&D costs and value capture within that industry. For example, pharmaceutical R&D is extremely expensive upfront, but pays potentially huge dividends to society later. To properly incentivize creation, drug creators are given strong intellectual property protections, with 20 years of exclusive monopoly rights. On the flipside, computer technology protections tend to be pretty weak. Developers of a new computer protocol benefit from network effects, where their value captured is greater as more people adopt their new protocol, even if those adopters are not paying any royalties to the creator. Since their value capture benefits from free use, creation is sufficiently incentivized by weak protections. This is reflected in the open source software culture.
Fair use of art is an interesting case that tries to benefit society without harming the creator. Educational applications are protected under fair use; if art is being copied and distributed for educational and non-commercial purposes without harming the market for the original art piece, then it is usually fine. That last bit about “harming the market for the original piece” is key here. If an artist develops a new style, such as Studio Ghibli’s iconic films, they can profit from the striking uniqueness of their new style while it is fresh. By the time an art student can study, learn, and reproduce that style, much of the value has been captured by the original work. However, generative AI radically accelerates that process; supercomputers compress years of training into hours, and copycats can go to market within weeks. Arguably, fair use no longer applies.
Generative AI seems to be disincentivizing the creation of original art styles. Now that imitation is easier and faster, the original creator is able to capture far less value. Thus original creation is now disincentivized. To restore balance, we ought to provide stronger protections to original artists to encourage innovation rather than replication. Hence, there is much outcry for artists’ royalties.
—
However, enforcement is a totally different issue. So far I’ve talked in abstract terms of “IP protection” as some kind of sliding scale that is easily adjustable. But actually implementing some kind of attribution and royalty policy seems intractable to me. Let’s imagine what an attribution scheme could look like. There’s two ways to look at it; inputs (training data) and outputs.
Artists could impose royalties if their copyrighted art is used to train a model. There could be a variety of schemes; flat per-use cost, a % of revenue generated by the model owners, etc. This seems to make intuitive sense, as we’re basically treating the art’s use in training data to be akin to any other commercial application.
However, in practice, this seems difficult to enforce. Many of these GenAI models are trained behind closed doors, and it’s not transparent what training data they have used. Plenty of copyrighted art has copies floating around on the internet, sometimes low-res or watermarked, but usually good enough to work as training data. Original art could also be “laundered” by running it thru multiple models; a royalty-paying model trained on Studio Ghibli could produce stylistically similar images that are fed as training data into a second model. The original artist has no idea if a copy of their art is being used. Moreover, these neural network models are complex black boxes that are often iterations on previous neural networks. Even if a new model does not incorporate any additional training data from Studio Ghibli, it’s likely that the previous version of the model already has trained and learned that style. With billions of parameters, models are inscrutable black boxes that can easily hide copyright violations.
Alternatively, we could try to attribute credit to original artists based on the output. Perhaps some algorithm could examine the result’s similarity to original art and attribute royalties accordingly. This seems difficult to do at scale. In traditional copyright law, the jury determines whether the “heart” of the original artwork was copied in a derivative work. How do you determine this? “I know it when I see it.” This process relies on human intuition and is prone to variance and subjectivity. It is impossible to scale human review up to AI-level output. Some kind of algorithmic attribution could be scalable, but it feels prone to abuse. Algorithms monitoring algorithms… is it algorithms all the way down?
—
Even if a royalty policy was feasible and enforceable, I fear that it would be corrupted and abused by the powerful. In econ jargon, this is regulatory capture. Regulatory capture has come for every major form of IP protection, from drugs to media. Even the most well-intentioned policies often become twisted and corrupted.
Pharmaceuticals are supposed to have a 20 year patent window, as mentioned before. However, those patents are extendable if some additional innovation is patented on top of the original drug, which can extend the monopoly period they have on the original. One notorious case is with Teva, who produce asthma inhalers. The pharmaceutical chemical is half a century old at this point, but Teva extended their monopoly control by filing a new patent for a new inhaler cap. That’s that little plastic thing that covers the mouthpiece of the inhaler. The cap has nothing to do with the inhaler medication itself. The patent extension is clearly against the spirit of the law. But the FDA let this through anyway. Who knows what backroom deals were involved. Only recently have these patent abuses come under scrutiny by the Federal Trade Commission, which recently filed a lawsuit against this exact issue.
Closer to the art world, Disney’s copyright extensions protecting Mickey Mouse are another infamous case of regulatory capture. Copyright protection originally spanned 28 years, with the option to renew once for a total of 56 years. This would cover the creator for their lifetime and allow them to reap the rewards for whatever new worlds or characters they invented, but shift them to public domain after they died. Creation incentives and free utilization are both satisfied. However, Disney didn’t want Mickey Mouse to enter the public domain. Whenever Mickey Mouse (1928) came close to expiring from copyright, Disney lobbied to increase this copyright protection window; first up to 75 years with the 1976 Copyright Act, and then again up to 95 years in the 1998 Copyright Term Extension Act. Last year in 2024, Steamboat Willie Mickey finally entered public domain, and if you were on Youtube at the time you likely saw some raunchy Mickey Mouse parodies, but importantly for Disney, the colored version of Mickey Mouse with white gloves and red shorts remains copyrighted. 95 years seems ludicrously long. The irony of it all is that Disney built their media empire off of public domain IP like Cinderella and Sleeping Beauty, but are now refusing to allow their own works to enter public domain.
Major media entities are taking advantage of these long copyrights by leaning into copyrighted IP. The Marvel cinematic universe is a clear example of this. Industry observers have noted that audiences are now attached to the characters of Spiderman and Ironman rather than superstar actors; if a big name actor gets uppity and tries to demand greater pay, Marvel can simply cycle them out for a cheaper young actor. Superhero movies have been compared to the western cowboy movie craze of the 50s, where the same characters were recycled over and over. But there is a key difference; western characters like Jesse James or Billy the Kid were public domain, and any movie director could reuse and remix them however they liked. Batman and Spiderman are privately owned IPs and can only be used by their owners. This has arguably resulted in less innovation and creativity from those characters.
I have to imagine that if some kind of art royalty system was introduced for generative AI, the system would be taken over and controlled by entities, either media or tech, that have the resources to lobby for self-serving policies. The spirit of these protections is usually to help out the small independent artist, but in practice they often help the rich get richer.
—
In a nutshell: society would benefit from original art having stronger copyright protection against generative AI, but I can’t see a workable policy. I think the path forward is to search for a feasible and robust attribution mechanism, and in the meantime leave things the way they are. I expect that we’ll see less and lower-quality original art, while gen AI recycles and fast-follows any new trends.
I think this issue is a microcosm of my political views in general. I am sympathetic to the goals of the far left, but I think their policies are poorly designed or impractical. Without any feasible policy, I would rather err towards laissez-faire policies and end up siding with the center-left.