An AI Art Ethos
As someone who uses “artificial intelligence” techniques in my creative work, I feel the need to write out a document that clearly defines my ethos when it comes to creating art using “AI”.
(note the majority of this article as written a month ago, things have been changing rapidly but I think the basic thing I lay out here holds.)
Quick Background
I’ve been drawing and painting since I can remember. I saved up for months to afford my first digital art program, Macromedia Flash 8 when I was in middle school (No software piracy allowed in our house!). I went to school for studio art with a focus on painting. I’ve been making imagery using neural networks since 2015. I learned how to run and install Linux on my old desktop specifically so that I could use style transfer models like Justin Johnson’s torch implementation of neural-style. I’ve written previously on this subject. I work as a UX Designer by day. I’ve exhibited my neural network images publicly in 2018 and again in 2022 in Austin TX.
My Ethos
Ethos, in the classical Greek sense, refers to the values, beliefs, and customs that define a person or group and shape their actions. In this context, my ethos refers to the values and beliefs that shape my approach to creating using “AI”. It is not meant to be a universal ethical mandate for others to follow. What I’m seeking here is to express my perspective, my view of ethics and the facts of the matter that we face.
I avoid the term “AI”
First I personally avoid the term “AI” when referring to this stuff. The word “AI” is notoriously vague and imprecise, it raises questions about what is meant by “artificial” and “intelligence”, and muddies the water with mystique.
What makes something intelligent? I believe that debates about the uniqueness of human consciousness and the sentience of AI, while interesting and potentially meaningful or inspiring for practitioners, often distract from the more pressing ethical issues that should be considered when discussing neural network techniques.
What makes something “artificial”? Humans are tool users. Every art form is engaged with our tool making nature. Even pure performance like dance or recitation of poetry relies on our ability to leverage language as a tool. The division between “natural” and “artificial” is not useful, and often veers people into a classical logical fallacy, the Appeal to Nature. We are a part of nature, and our tool use is continuous with our experience in the natural world. I do not see a sharp division between nature and artifice.
Instead of using the term “AI,” I prefer to use the term “neural network” to describe the family of techniques I engage with. This term is more precise and focuses on the specific technique that I use, rather than opening up abstract discussions. By using this term, I hope to ground discussions about my work in something concrete and avoid unnecessary distractions.
I have on occasion used the term AI art in posts to make it fit the current trending discussion on social media, but I have preferred this term going back to at least 2015 for the reasons outlined above.
I avoid the term “Art”
I believe discussions about the definition of art are often unproductive. I have a broad and vulgar definition of art. I don’t use the term “Art” when referring to my work. I make images, whether you call that art is up to you. I don’t care.
I align with Larry Shiner’s perspective in “The Invention of Art” when it comes to the concept of techne. Techne is the Greek word that’s often translated as “art” or “craft”. It refers to the practical skills and techniques required to create a particular type of work. In my mind, it’s a way of knowing (episteme) gained through practiced engagement rather than reading or instruction. You can read about the history of knives or the physics of propulsion, but the only way to really know how to throw a knife at a target is to do it over and over again. This is what I mean when I do use the term “art”.
I don’t sell images, I share freely, and I strive for transparency
I don’t sell my images. There really isn’t much point in even trying. Supply is infinite. If you want a print feel free to right click and save it to your computer. I doubt I could really get that many people to pay even if I tried. I believe we are seeing a paradigm shift in our society and I would prefer to strive for de-commodification rather than fruitlessly try to re-commodify it. I hated NFTs before it was cool.
In line with this, I believe that just as I have learned from others who have shared their knowledge, I have an obligation to share with others freely. I don’t always post my prompts with images but I am happy to give them to anyone who asks. I also open source any scripts I write.
In addition to sharing my experience, I also try to be transparent about my use of neural networks in my image making. I do not hide or pass off work that was generated using a neural network as something that I created entirely on my own. I think it is important to be transparent about the use of technology in “art” making in order to avoid obfuscation and to recognize the role that technology has played in art production throughout history.
I believe this obfuscation of technique in the history of art has led to a misconception of art as a test of skill, an athletic competition that can be won. I think at least part of the backlash against neural network art is because of the perception that it reduces hard won and sacred skills to a couple of lazy clicks of the button. However, I believe that this view of art and technique is fundamentally misguided and that the mechanical reproduction of art has always been a part of art production.
It is worth noting that this article was written in part using ChatGPT.
I avoid using loaded prompts
I try to avoid using the names of specific artists in my work. This is not because I think it is necessarily immoral, but rather because I find it uninteresting and a bit cringe. I have used the work of artists like Lisa Frank on occasion, but that goes back to my initial experiments with style transfer and honestly her style is so distinctive that I think most viewers are able to grok what I’m doing. In general I think heavy reliance on an artist’s name hinders the full potential of a lot of prompt writing.
I also avoid using loaded terms like “good”, “trending on Artstation” or “beautiful”. I just think it’s a little silly and frankly a misunderstanding of how “good” art/techne operates. What is a “good” image? Good is a metaphysical value, it’s contingent on a definition of telos, which a neural network cannot provide.
I see neural networks as a generative tool
I try to be careful about how I describe my relationship to these images as well, avoiding phrases like “I made this image” in favor of “I generated this image.” I believe that this phrasing is more accurate and avoids arguments about what “making” art is. Overall, my approach to creating art using neural networks is inspired by the work of early algorithmic artists like Jean-Pierre Yvaral, who saw algorithmic transformation of imagery as being akin to the scientific process, a process of variable experimentation and the generation of new insights.
I don’t think the main value of neural networks is generating specific images I have in my head. Rather I prefer to use neural networks as a generative tool for exploring the collective unconsciousness, reflecting on the biases of society and the model it reflects, and for generating new insights by mashing up different ideas. I think trying to be overly prescriptive in one’s approach with neural networks is to avoid the thing that actually makes them interesting. Using a neural network as a tool is a lot more like a collaboration with another mind, it’s a conversation that goes back and forth, not just a series of unidirectional commands.
The legal issues
Legal issues are distinct though not separate from moral issues around neural network art.
Intellectual property (IP) law as we know it did not really exist 500 years ago. Shakespeare didn’t invent Romeo and Juliet, he stole that from another poet who stole it from another poet. The 2010 “Everything is a Remix” by Kirby Ferguson was formative in the development of my views on this matter.
It is important to recognize that IP law is not a test for moral action. IP law was created to incentivize innovation and does not necessarily reflect any kind of transhistorical moral right. Medical patents are used to gatekeep life sustaining technology for millions every year. Famously the inventors of insulin sold their patent to the University of Toronto for a single dollar because “Insulin belongs to the world”. I do not believe that anyone has a moral claim to IP outside of the strictures imposed by law and the pragmatic goals we seek to achieve through our design of those strictures.
An important technical note on neural networks training models
Neural network models are not databases. They do not “sample” or “collage” images together from scraped datasets. The LAION 2B dataset that this version of Stable Diffusion was trained on has 2 billion images. The final trained model weights is 7.7 gigabytes. That’s about 4 bytes per image. To give you an idea of what that means, this Mario sprite gif is 1,684 bytes. The image below is 421 times bigger than 4 bytes (a 40,000% increase).
This isn’t a very rigorous argument, there’s lots of additional parameters you could play with (the original NES Mario sprite is almost certainly smaller than the facsimile I made), but I include it to help illustrate how wrong it is to conflate the LAION dataset with the final trained model or to assume this thing is just copy and pasting stuff.
Rather than thinking of neural network models as a database, think of them as a generative algorithm similar to the Library of Babel. The model does not contain the images it scrapes. Instead it contains an algorithm that can be used to describe almost any image possible. (The Library of Babel is not a direct 1:1 of a neural network but it’s a vastly simplified way of illustrating algorithmic possibility space.)
Fair Use
As it stands the only real test on the legality of this stuff will be the courts and who is presiding in that court and what their opinions and disposition is towards the specific case. IP law is civil not criminal. The enforcement of IP law is often arbitrary, with different standards being applied to different cases. Andy Warhol’s work was considered “transformative use” under IP law, while street artist Shepard Fairey’s work was not. In the future, companies like Disney or Warner Brothers may hold the power to enforce these standards.
As far as I understand it, there is no reason to believe that training a neural network on publicly available images is not protected under Fair Use law. There is no copyright protection for the “style” of an artist in the abstract. Any move towards an opt-in/opt-out model on the part of model developers is more of a public relations matter than a legal one. It does not validate the claims of detractors on the legality of training these models. Additionally, I do not believe that opt-in/opt-out options will significantly hinder the development of neural networks long term. The incoming waves of public domain art are inevitable, and neural networks are getting better at training on less information.
While training a neural network on art scraped from the internet is not IP infringement (or at least is not so without any established precedent to back it up), using a neural network to generate images of copyrighted characters and then sell them is most definitely IP infringement in the US (Japan has slightly more complicated history with this). This has nothing to do with whether or not images of Mickey Mouse were scraped for the training data set. The illegal action is not the training of the neural network or even the creation of the image but rather profiting from a copyrighted property. Anime conventions banning AI art are not doing so because neural network art infringes on IP law. They are doing it because of the vocal backlash from artists (who in reality are the ones unambiguously infringing on copyright).
Finally, just like media piracy, the only way to actually enforce IP law on the free and open Net would require draconian surveillance and invasion of privacy far worse than the problem it seeks to solve. Even then I doubt it would do anything to stem the flood. Genie is out of the bottle ya’ll.
The Real Issue
The real issue here is “How do we live in a post-scarcity economy?”
Let’s be clear, most artisan labor was automated over a century ago. Being a professional artist in the modern day is a combination of incredible talent, hard work, nepotism, and luck. The luddites were right that the automated looms would destroy their way of life. What many think of as “art” is but a remnant of an older system.
I prefer to say “displace” rather than “replace” because it avoids the endless meaningless arguments about whether or not “AI can really replace humans in totality”. It doesn’t matter if neural networks can completely replace humans in totality. A mechanized combine harvester doesn’t need to possess the indomitable human spirit in order to put 50 human harvesters out of a job. It doesn’t matter if neural networks have the “spark of creativity”. The production of art in commercial contexts is not driven by some high human value, it is driven by the necessity of production. Humans will remain in the loop with automation just as they have in the past with previous forms of automation. This will still have an effect of putting people out of work. Labor productivity goes up, labor demand goes down. It’s simply supply and demand.
This has been repeatedly demonstrated by the MIT economist Daron Acemoglu whose research shows a net loss of jobs to automation since at least the 80s. Digitization has already produced massive productivity gains in the production and distribution of imagery.
We are not standing at the edge of the cliff. We already jumped.
The good news is a deep well of history to draw on with this subject. Artists have reacted to industrialization in a variety of ways.
Impressionism was made possible by the invention of lead tube paints by John Goffe Rand, which revolutionized the process of mixing paints by industrializing the mixing and packaging of paints in a portable form with a wider range of colors than artists previously had access to. This is creative destruction at work. No more need for assistants to mix and prepare paints, instead the Master is free to work unchained from the studio en plein aire.
Alphonse Mucha was a classically trained painter and his application of his skill to industrial printmaking for theater print advertising helped define the iconic Art Nouveau movement. Much of the art movements of this period were artists synthesizing their training with the industrial formal mediums newly available to them.
Even William Morris, who wrote explicitly against the evils of mass production, found himself leveraging the new technologies of industrialization through his machine-woven carpets. These carpet designs were very likely encoded by punchcard, a direct precursor to modern computer memory systems, making Morris in some ways one of the first digital artists and programmers.
Digital reproduction has only magnified this trend, with the creation and distribution of digital artifacts becoming easier and more widespread. Attempts to reverse this trend, such as NFTs and “buy local” art market campaigns, have largely failed. We must move beyond the old paradigm. This is not utopian. It is the opposite of utopian. Utopian is thinking that you can put this genie back in the bottle.
We have made a mistake in conflating price with value. Some things like diamonds command a high price but are of little value if you are starving to death in the desert, while other things are extremely valuable and command no price. We must face the real issue if we are to save those things which are priceless.
An IBM slide from 1979. Source
The responsibility for preserving what is priceless is in our hands alone.
