This article was previously published on fx(hash).
Generative AI for image creation has divided the world into two camps. On the one hand, there are people who are enthusiastic about this novel set of tools, who incorporate AI into their artistic practice, and/or who collect works from the people who do so. On the other hand, there are people who consider AI-generated imagery low effort, not art, and/or an affront to visual artists who have spent years honing their craft. I am squarely in the camp of the AI enthusiasts, and here I would like to argue in favor of this position.
One common critique is that AI is inherently predatory, stealing works from other artists. Let's get this critique out of the way first, as it is not actually inherent to AI and this article is not about the topic of potential copyright violations in AI art. I think we can all agree that if the AI was exclusively trained on images that were explicitly licensed for such use then there is no issue. The training images could be all in the public domain, or a large image repository such as Getty Images could decide to train an AI on their image catalog, or an artist could train their own AI exclusively on images they have taken themselves. (Ivona Tau does exactly this.) So for the remainder of this article, let's assume the artist has both the legal and the moral right to use the AI to generate images for commercial use.
The more interesting question, from my perspective, is whether AI-generated imagery can be art. To address this question, we need a definition for what is art. I'll give you mine:
A person is making art whenever they put more care, craft, skill, or thought into something than is required from a purely utilitarian perspective.
By my definition, a sushi chef who has honed their skill of cutting rolls to exact dimensions is creating art. A scientist who carefully refines and polishes their visualizations or their writing is creating art. A gardener who carefully plants, trims, and prunes may be creating art. You may not agree with my definition and find it overly broad, but if so I'd invite you to consider why you want to deny so many people their artistry. Do you lose anything if a chef, a scientist, or a gardener is an artist?
If you allow me to proceed with my definition of art, it directly leads to the question of how much care, craft, skill, or thought AI practitioners put into their work. And in fact, the most commonly raised critique against AI imagery is that it is low effort. Now it is certainly possible with AI to create images without putting in much effort, if you don't particularly care what you get out. Provide a few quick prompts to DALL-E or Midjourney and they'll hand you some images. But just because a medium can be used this way doesn't mean it's the only way it can be used. I can make low-effort collages, but nobody would doubt that there are true collage artists that create amazing works. I've never heard anybody accuse Wangechi Mutu of producing low-effort art, and yet many of her works are just cuts from magazines glued onto a canvas.
One way for an artist to get around the low effort critique is to build and train their own models. Many of the AI OGs do that, because as little as a few years ago this was the only way to create images with AI. Let's consider Ivona Tau once again. She takes a collection of her own photographs, uses them to train a neural network that she has coded herself, and then uses the trained neural network to generate unique artworks that reflect but also transform the original photographs. Whatever you think of her work, it is not low effort, that's for sure.
However, in my opinion, while training your own models is laudable, it is not a requirement for making meaningful AI art. Let's consider what people do who use primarily or exclusively some of the widely-available, off-the-shelf models such as Midjourney, DALL-E, or StableDiffusion. First, they may spend extraordinary amounts of time giving the model different prompts, over and over, to explore the model's latent space, that is, to determine exactly which prompts have what effects. In addition, they may explore how to fine tune prompt weights to get specific outputs or compositions, how to modify model parameters, which sampling algorithm to use, and so on. It is not unusual to generate hundreds or thousands of outputs to fine-tune everything just right for one final image. Second, in addition to basic prompting there are so many other tools these artists can use to gain more control over their creations. There is image-to-image, upscaling, inpainting and outpainting, ControlNets for spatial conditioning, the list goes on and on and more tools are added every day. The AI space provides plenty of room to make this a high effort activity, and in fact I would like to invite you to try this out for yourself to see how much effort and skill is actually required to get images you are satisfied with.
Another critique I have seen is that the artist has no control over the image output and composition and is just randomly sampling possible images from the total set of outputs the AI can produce. Well that's certainly not true at all for image-to-image approaches, where the artist provides an input image that guides the AI. This is a commonly used technique. My collection AI Artifacts was made this way (see below). If you compare the AI output to my input image, you can see that the AI followed my image prompt quite closely and mostly provided texture and depth. If I had created texture and depth by gluing magazine clippings onto a canvas, Wangechi Mutu style, nobody would bat an eye and accept my work as art. But if I instead use a generative AI model that doesn't count? The AI was somehow able to take my code-generated artwork and turn it into non-art? That's quite a feat.
Finally, I've seen some critics draw a sharp distinction between the people that write their own neural network code and/or train their own models and the people who use off-the-shelf models, arguing that the former may deserve the label of artist but the latter do not. Of course everybody is free to draw their own boundaries but I cannot at all agree with this assessment. From my observation, people using off-the-shelf models produce works at widely differing levels of quality. It's not all the same "low effort" stuff. Instead, some people consistently amaze me with what they can coax the AI to do, while others produce work that is much less impressive. Seeing these differences I feel comfortable concluding that there is true artistry and skill involved at the upper end of the performance scale. Not everybody is equally talented or capable when it comes to working with AI.
A great place to observe this difference in skill levels is the long-form AI platform EmProps. Everybody who publishes on this platform has access to the exact same handful of models and there is neither manual post-processing nor (for the long-form drops) manual selection of outputs. The EmProps OpenMarket Beta has released several hundred long-form collections. I'd like to encourage you to explore the different artists and see to what extent they get very different results using the same tools and methods.
In my mind, working with generic AI image-generation models is almost the purest form of artistic expression. Everybody has access to the exact same tools. The ability to generate imagery is not limited by somebody's physical ability to move a pen or brush in a precise manner or by their mental ability to write sophisticated graphics rendering code. Instead, all that matters is artistic intent, imagination, and judgment. In the musical arts, we are perfectly happy to celebrate composers who have written masterpieces, even if they never play any instrument themselves. I believe we should similarly celebrate the most impressive AI artists. They are visual composers who tell stories, create emotional reactions, or otherwise delight us with their artistic perspective and vision, even if they never actually draw, edit, or touch up the images they produce.
If you enjoyed this article, you may also like my earlier article discussing how little AI is necessary to call an artwork AI art.
