AI, Creative Work, and Intellectual Property, Part 2: Economic Considerations
AI, Creative Work, and Intellectual, Part 1
Whether, when, and to what degree generative AI work displaces human creative work will primarily be driven by issues of quality and cost. Contract negotiations, like the WGA and SAG-AFTRA negotiations with the AMPTP, may affect it a little at the margins, but if AI can outcompete human creators at an economic level, any temporary contract gains will be short-lived. Furthermore, that’s largely a good thing — to the extent that generative AI displaces human creators because it can provide similar or higher quality at lower costs, consumers of art will benefit significantly. In this regard, displacement of human creators by AI will be directly comparable to other transformations of the art world by prior technology, and to a lesser degree comparable to the transformation of other industries by new technologies. The primary way in which generative AI displacing human creative work will raise new issues is if it is part of a wave of AI eliminating so many jobs that maintaining anything approximating full employment becomes very difficult, in which case we will need political economy solutions oriented to the entire problem.
First, I want to be clear that what’s at stake with generative AI is absolutely not the ability of humans to create art. I had an idea for a short story a few weeks ago, and after I described it to a friend who encouraged me to write it, I spent a couple of days writing it. The existence of generative AI — even generative AI that hypothetically could do creative writing better than I can at negligible cost — would not take away my ability to create written art. The same is basically true for any other form of art. Even if a future generation of Stable Diffusion or the like is better able to create visual artwork than any human artist at negligible cost, nothing will prevent an artist from choosing to pick up a paintbrush and paint, or a composer from writing a new piece of music.
What is at stake, instead, is the existence and size of a market for human creative work. Generative AI may make it much harder for humans to make a living as writers, actors, or artists. In this regard, generative AI is like prior changes in technology that have transformed the market for certain forms of art. Before the advent of photography, a large and robust market for portrait artists existed. Wealthy families would routinely pay for portraits. Other forms of visual art also had robust markets — if you wanted to have a pretty landscape or still life hanging on your wall, you had no choice but to pay an artist to paint one for you. Technological change displaced many, although not all, of those artists. Instead of a hand-painted landscape, a family might buy a lithograph by Currier and Ives. Instead of hiring a portrait artist to paint a portrait of their family, they became much more likely to have a photographic portrait taken. These changes eliminated the livelihoods of some artists — many of the less skilled, or at least less able to drum up business, portrait artists were forced out of the market by photography. But that same transition created massive new opportunities for consumers — no longer was quality art or accurate portraiture a privilege enjoyed only by the wealthy and perhaps part of the middle class. Now, everyone could have a photographic portrait of their family, and indeed today portraits of even distant family members and friends are common in many homes. Painted portraits did not become obsolete — some wealthy people continue to commission them. Nonetheless, it is obvious that the portion of people employed as portrait painters has greatly decreased from its peak, at the same time as a substitute with a similar albeit different value and a much lower cost has made portraiture of a sort commonly available.
The same story can be told, with minor differences, for other forms of visual and audio works. Radio and phonographs substitute for, and compete with the market for, live music. Motion pictures and television substitute for, and compete with the market for, live theater. In each case, the availability of art was greatly increased by the new technology, at the same time as some of the people employed in the previous art markets lost their jobs. Again, the substitutions aren’t perfect — a movie isn’t the same as a play, even if made from the same script. Nonetheless, consumers on average benefit tremendously from the increased availability of art.
Whether generative AI can produce art of similar quality to human art, at lower cost, remains an open question, in both regards. So far, AI generated art is generally lower quality, often much lower quality, than human-made art. The creative writing that I’ve seen from LLMs like ChatGPT has usually been terrible, occasionally rising with careful prompting and selection of outputs, to mediocrity. Some of the features that made AI generated images of humans largely dreadful except as a means of generating visual body horror have improved enormously — we can no longer reliably sort out AI generated images by looking for inhuman hands. But even in still visual images, one of its strongest areas, generative AI artwork is not really competitive in quality with high quality human art. Also, if improvement requires ever larger models with ever more compute driving them, keeping costs low may be in tension with improving quality. Widespread reports, some backed by formal testing and some more impressionistic, suggest that ChatGPT 4’s quality has declined over time.1 While that may be due to reinforcement learning with human feedback causing improvements in some areas (e.g. avoiding dangerous or racist responses) at the cost of declines in other areas (quality of computer code generated), many people speculate that it might also be a result of OpenAI seeking to economize, substituting a cheaper model or combination of models for the more expensive, in terms of resources needed for each response, model used previously under the same label. If generative AI can produce results that are competitive with human creative work, but only by using levels of computational power that make it prohibitively expensive, the impact on the market for human creative work will be limited.
If generative AI creative work becomse competitive with human creative work, however, the effects will be profound. Even if the creative work of generative AI is somewhat inferior to the work of human creators, if it is sufficiently cheaper, market processes will inevitably push to much of human creative work being displaced. That won’t completely eliminate the market for human creative work — if it becomes cheaper to make television shows with generative AI taking the role of most of the writers, but the quality is significantly lower than shows written by human writers, the market will bifurcate, with premium markets offering the superior, human product, while other services offer the cheaper, lower quality AI product. In this regard, it would be much like the portrait market today — high-end portrait artists continue to be able to charge for painted portraits, but photographic portraits replace much of the market, while also making it cheaper and more widely available. High-quality generative AI art can make bespoke or small market productions profitable. Today, commissioning high-quality written fiction is out of reach of most people, who instead need to choose from whatever is available on the market. Personally commissioning television shows is the realm of the ridiculously rich — Jeff Bezos may have commissioned additional seasons of The Expanse and commissioned Rings of Power as much for his own consumption as because he thought they would be money makers, but that’s out of reach for people who aren’t billionaires. If a generative AI can cheaply produce a high quality television show out of the whole cloth — something that is by no means certain, but well within the realm of possibility — that could be something that everyone can do in twenty or thirty years. From a consumer’s perspective, therefore, generative AI displacing human creative work is likely to be largely positive, although it may not be entirely positive. And for amateur creators, generative AI may open up new possibilities for creative expression. I can write a short story or a TV script, but I can’t have that made into an actual, professional quality TV show. But if generative AI can replace actors, directors, set designers, and all the rest with ersatz artificial versions, it may be possible for me to write a script and have a generative AI produce a television show for me. New technologies, including new creative technologies, are disruptive, but the benefits on average outweigh the costs.
No amount of careful negotiations or labor union solidarity will be able to overcome sufficiently large cost savings from switching to AI creation. I support the WGA and SAG-AFTRA in their strikes. I hope they claw a higher share of the revenue for TV and movies away from the AMPTP, and that they get fair residuals for streaming services distributing their works. Creative workers deserve substantial portions of the revenues from their creations, and even if it increases costs to consumers — not a given — we should be willing to pay a little more for those creators to be able to make a living and to be compensated for the risks in their careers. It also makes sense that the WGA and SAG-AFTRA are deeply concerned about the use of AI to replace some of their jobs. However, if generative AI can do their jobs at an acceptable quality level for a significantly lower cost (or at a significantly higher quality level at only slightly higher cost), no contract can protect their jobs for long. If these strikes end in victory, but the next three years allow AI to actually replace humans in some of that work, the AMPTP will insist on changes to those terms in the next contract. And even if it doesn’t, if OpenAI can make professional quality television for pennies on the dollar through generative AI, even if Disney and Netflix agree to human-only creation, TV-GPT will spring into existence to compete with them. By all means, the WGA and SAG-AFTRA should fight for the best terms they can get — that’s what unions are for — but we should be realistic that if the capabilities and cost of generative AI improves to the point where it can do more for less — or close to the same amount for much less — economic pressures will still squeeze out many of the human creators. That shift would likely be good for humans in general, albeit bad for the members of the WGA and SAG-AFTRA.
The one case in which the general welfare improving effects of more efficient technologies might fail is if AI displaces not just a substantial portion of creative work, but rather a substantial portion of all work, in a way where new and better jobs don’t replace the old ones in the way they have generally in the past. I think Brad DeLong has a parable about how the rise of the internal combustion engine displaced the farriers and grooms who were employed working with horses, at the same time as it generated on net more and better jobs as car mechanics and drivers and engineers and the like.2 The question, if AI displaces much of the work done by human brains, is whether it will open up new and even better jobs. As I believe DeLong phrased it, the great question is whether humans are still the farriers and grooms, with their jobs displaced but new jobs replacing them, or whether there is a technological shift coming where we are the horses, rendered largely economically obsolete by new technology. If AI displaces not merely creative work, but a large portion of all work, then we will need to wrestle with issues of equitable resource allocation, finding meaning in our lives when professional work cannot be the source of that meaning, and similar problems related to living in a largely post-work world.
1See Lingjiao Chen, Matei Zaharia, and James Zou, “How Is ChatGPT’s Behavior Changing over Time?”, available at https://arxiv.org/pdf/2307.09009.pdf. Various commenters have criticized parts of this paper, and it cannot determine the causes of changes in ChatGPT’s behavior.
2Thoughts similar to this are available at J. Bradford DeLong, “Technological Progress Anxiety: Thinking About ‘Peak Horse’ and the Possibility of ‘Peak Human’”, Grasping Reality on TypePad, available at https://www.bradford-delong.com/2015/09/highlighted-the-history-of-technological-anxiety-and-the-future-of-economic-growth-is-this-time-different.html. I can’t find the actual “are we the grooms or are we the horse?” post right now.
