What generative AI boosters misunderstand about art
As I see it, the people who promote generative AI under a promise that it will “democratize” artistic / creative work or make personal communication easier are misunderstanding a very important thing.
I can see the part of the argument that says that a lot of people have good artistic ideas that they’re simply incapable of properly expressing. I, myself, have a few that probably will not ever turn into complete works because I’m not good enough at writing that kind of stuff. The problem is that even if generative AI is developed to such an extent that it will no longer make errors or have significant negative externalities, a lot of people will still prefer to avoid any artistic content that seems generated.
Analogy: targeted advertising
Businesses that make products want us to buy them, but we, the consumers, (1) might not be aware that the product exists, (2) have a finite amount of money. Advertising provides a solution to these problems by making people aware of the product’s existence and also having the business provide some information relevant to the product – for free to the potential consumer, but at cost to the producer/distributor.
Theoretically, a consumer can get the same information about the product through other means – read through the business directory to find which companies in a specific area exist and inquire them manually, read / listen to / watch third-party articles describing or reviewing said products, visiting (physically or online) the store where such products would be sold (though, of course, it is worth noting that placement of products in stores is also often a form of advertising)… But these cost time, energy and sometimes extra money that the consumer doesn’t have.
And, while advertising is usually covered by laws that ensure businesses can’t just lie to consumers, there is still a lot of leeway for them to make accurate-but-misleading statements or otherwise present their product in a way that seems more appealing to the consumer (consider how many cosmetic or luxury products are advertised under an implication that they will make the buyer seem more romantically attractive). As a result, while most consumers treat the “text” (in a literary theory meaning) of advertising with suspicion, they will tend to look at other signals to form an opinion.
One of such signals is the expense (both material and in terms of effort) at which the advertising was made and published.
An advertisement for a luxury product made cheaply and placed in a low-brow magazine will make the readers suspect that the product itself is of lower quality. Likewise, a video ad for a local low-end store filmed with the same techniques you’d expect from a highly-profitable nation-wide chain and aired during the most prestigious sports game of the year would make consumers think that, with so much money spent on it, the store would have to mark up the costs of their products to compensate for it.
When targeted advertising – the ability to provide different ads on the same spot to different consumers based on collected data – became a thing, it significantly reduced costs for advertisers who wanted to reach specific audiences, as instead of having to provide the same spot to just one advertiser or cycling them randomly, these can be selected for what the individual viewer is most likely to respond to. What this has resulted in practice is the ability for the creators of shoddy overpriced products to promote themselves as if they’re the kind of mid-range or luxury products that would normally advertise in that spot – and, as reported by The New York Times, is just one of the least worrying consequences of targeted advertising (in addition to privacy concerns, the lack of review that a more generally-aimed at would invite, etc., especially when applied to political advertising).
Back to generative AI
Right now, generative AI boosters promise to “democratize” content creation and make it more available to people who don’t have the time/money/effort/skill to do so otherwise. But the same problem exists: consumers of creative works don’t have the time and money for everything that comes out. The worldwide film industry has produced 9628 movies in the year 2024 alone1, and assuming a minimum length of 90 minutes per movie (the typical definition of “feature-length”), that much cinema would take 601 days of uninterrupted watching. And there’s even more music, books, TV shows and video games being made every year. People will need to have some way to filter through stuff and determine which specific works are worth their time, money and attention.
The act of reading a book, watching a film or video, playing a video game, etc. is, in a way, a kind of investment (or a bet, depending on your view) – the end user spends their time and/or money in expectation of receiving some kind of enjoyment, education, information, etc.
You can tell by the fact that a disappointing creative product often gets described as “a waste of my money” even if it is, on an “objective” technical level, a functional product (a movie with a coherent plot, decent acting and no VFX errors; a video game that is playable from start to finish without bugs, etc.), and how people might overlook some of the technical issues if the product’s more subjective qualities appeal to them sufficiently (for instance, I personally like Sonic Adventure 2, but I also understand that it’s a very janky game, and both it and its later ports have their share of bugs and annoyances that can ruin someone’s fun).
Much like how a commercial for a product is normally supposed to show the product, tell what it’s supposed to do and signal who it appeals to, same happens for movie trailers, TV show advertisements, demos and videos for video games, and so on. Often, these include the most “expensive” scenes from the product, or ones that the creators are proud of the most, in an effort to indicate that this kind of quality is to be expected from the entire thing.
And in this case, generative AI acts like a tool that can create a decent-enough promotional material for anything, or “fill in” the gaps where the creator was unable or unwilling to do the work. Even if the generated content is good enough – though, at this point, it often is not – the matter of fact is, the “cost” / effort of producing it has dropped down to near-zero. Much like shoddy product makers that have learned to take advantage of targeted advertising, shoddy content creators are taking advantage of generative AI to produce content that looks alright at first glance, but could contain pretty much anything inside.
In a world of victorious generative AI, someone browsing through a book or movie store, watching a movie trailer or glancing at a piece of software in their app store / repository would be nearly incapable of telling how much effort was put into anything – and therefore, whether it’s worth interacting with. This is not to say that generated works can’t ever be good, but that the only way to determine that would be to consume it in its entirety, which most people won’t do.
The fact that people are now finding common traits of generated content and treat anything made with AI involvement as “slop”, regardless of whatever its actual quality might be, is entirely understandable in this context. The fact that people would rather look at someone’s mouse-drawn scribbles or basic-level pencil sketches instead of a more “decent” looking generated artwork shows that, in these cases, it’s the effort that went into making something that counts more than the “quality” of the art itself.
Anyone remember NFTs?
The latter part seems to me like the same mistake that a lot of NFT boosters made several years ago. The overall idea of having some kind of digital token that tracks “ownership” over an image, in my opinion, keeps making sense even if the image itself can be easily duplicated. Someone looking at one of many online scans of the Mona Lisa would probably get a better view of it than if they visited the Louvre and stood in a tight crowd of people eager to take photos of it behind a protective glass, and yet the original is what matters the most and what still motivates people to journey to Paris and stand in (and be part of) those crowds.
And yet, what did most NFT collections offer? Thousands of similar-looking algorithmically generated images, whose only substantial difference from each other is the RNG seed that was used to create each specific image. Sure, some artist might have spent some time drawing up variations of faces, hats, clothes, etc. that make each individual “ape” or “punk”, but once these are done, it’s possible to create millions or even billions of these at any moment.
Consider that if Leonardo da Vinci, a world-renowned artist, suddenly turned out to have had algorithmically made thousands of variants of the Mona Lisa in different clothes, with different hair, faces, backgrounds, etc., these would not gain the same value as the one Mona Lisa we know today – if anything, it might even devalue it! The people who study the painting in detail and research its history would not rush to then study these new thousands to the same extent.
One might respond to this by mentioning trading cards, which often gain significant value despite being frequently resold. But even then, the way they gain said value usually lies outside of the creator’s direct control. A card featuring a sports player often gains value if that player likewise becomes famous for something. A playing card gains value if its in-game abilities become unreasonably powerful – sometimes only after years of the game’s competitive environment developing one way or another. Either way, it continues to hold value outside of the direct context of trading. And unlike NFTs, a physical card is often produced in restricted amounts, in a way that can’t be easily restarted once production is over – whereas the way most NFT collections are generated means that the same algorithm that made a thousand “unique” images can also be used to generate another thousand at pretty much any moment.
Art is not a product
Some people talk about generative AI as if it will result in the next Industrial Revolution. Indeed, the Industrial Revolution has also resulted in a lot of crafts that were traditionally handmade or developed by individuals or guilds being turned into mass-produced goods. It also involved a significant centralization of wealth and power, to the point that a large portion of the 20th century was spent fighting back against it in one way or another. But the ultimate end result of the Industrial Revolution was that a lot of products became way more affordable for the average person, and overall quality of life still improved. The boosters promise that a similar thing might happen with generative AI, but at least when it comes to art, we are not starving for content to consume. Instead, we are trying hard to figure out which “content” is or isn’t worthwhile, and generative AI is only making things worse in that regard, not better, by flooding the landscape with what’s functionally garbage. This is not to say that genAI, or other improvements in machine learning, are incapable of helping people at all, but the biggest consequence of it right now, as seen by the average person, is an overabundance of zero-effort slop and statements from big tech companies about how dangerous it is and therefore they’re the ones who should be put in charge of regulating it, because imagine what would happen if someone else were to build one (which, on some level, sounds a bit like a threat a mafioso would do).