Why Computers Do Not Make Art
We’re living in an age of amazing advances in Artificial Intelligence (AI) technology, and amazing new visual art that uses AI. We see this in the new neural stylization apps and technologies, in the trippy, evocative DeepDreams, as well as as the intriguing new work of artists like Mario Klingemann, Helena Sarin, and Jason Salavon.
Together with these new technologies comes the hype: technologists who claim that their algorithms are (perhaps) artists and journalists who suggest that computers are creating art (or might be soon). These discussions usually betray a lack of understanding about art, or about AI, or both.
This essay is meant for anyone who wonders about whether any current “AI” could be called “an artist.” I will explain why today’s technologies do not create art; they are tools for artists. This is not a fringe viewpoint; it reflects mainstream understanding of both art and computer science. Yet, it seems poorly understood by many of those who comment on software art, perhaps because it requires some understanding of both fields.
This leaves open the question of whether computers could ever create art in the future, using some as-yet-undiscovered technology; I discuss this question in the following essay. In this essay, I focus solely on the kinds of technology that exist today.
Authorshop in Art
Before talking about AI, it’s worth understanding how authorship in art is currently understood, particularly in the context of procedural, crowdsourced and other types of systems.
In the current Western tradition, the artist is the mastermind behind the artwork. No matter how much of the work came from other people, or from automation, or from natural processes, we ultimately assign credit to the individual (or collaborators) behind the work. The artist is the captain of the ship; the head honcho; the ringmaster. They set up the system or rules for the artwork, and they, as desired, exercise final decision-making control over every step of the process, including which outputs of the system (if any) to finally select as the artwork.
Starting from around the invention of photography, and in the Modern art era that followed, the definition of art broadly expanded. The most famous example is Duchamp’s Fountain, which was a urinal that he (or possibly a female friend of his) flipped over, signed, and submitted as a sculpture. This, and other work like it, ultimately set a precedent that any object can be called an artwork by an artist, regardless of where it came from or who made it. In environmental art, like that of Robert Smithson and Christo, the artwork can be a site, rather than an object. In performance art, like that of Marina Abramović, there does not even need to be a physical artifact. The contributions of other people often go unrecognized; for example, in Damien Hirst’s extravagant show Treasures from the Wreck of the Unbelievable, which surely involved the work of an army of highly-skilled collaborators, only Hirst is credited as the artist.
In popular arts, the same is true: a single artist is normally credited, no matter what materials they used or who else helped. A DJ that samples and remixes sounds is the artist behind a new track. Since auteur theory, the director is created as author of a film, despite the recognized contributions of many other artists, including screenwriters, actors, cinematographers, set dressers, visual effects artists, and so on, as well as the software used to create visual effects.
In the fine art world, there is a long tradition of procedural and software art. John Cage used random procedures to compose music. Many of Sol LeWitt’s wall paintings are written as lists of instructions that anyone can follow. Harold Cohen’s AARON software generates paintings based on a set of randomized procedural rules that he wrote. Karl Sims’ and Scott Draves’ evolutionary artworks involve automatic creation of images, virtual creatures, and procedural animations from crowdsourced feedback. Many artists have created lovely abstract interactive artworks that respond to the viewer’s movements, including the work of Daniel Rozin, Camille Utterback, Scott Snibbe, and Golan Levin. (My own version of this is a canvas that paints a picture of you as you move in front of it.) New Media arts programs typically have entire programs around software and interactive art.
Even though software, crowdworkers, and/or artisans may have executed on the artist’s instructions, the artist is the person who instigated and coordinated the work. None of these software systems is called “an artist.”
Artificial Intelligence is Not Intelligent
For AI software to be called an artist, one would expect that it is somehow intellegent, conscious, or similarly operating at a human level.
There is enormous hype around AI technology, and enormous misconceptions. People think that AI technology is intelligent, but it is not. This is understandable, given the misleading name, and the misleading media presentation of it.
None of the AI software we have today is intelligent in any real sense of the word. The massively successful AI tools we have are essentially very sophisticated data-fitting procedures. This applies to the biggest successes of AI today: image recognition, natural language translation, web search, recommendation, and ad targeting. For each of these tools, humans select the problems to work on, engineer models, fit these models to data, and iterate and repeat the engineering process until they are happy with the results.
In short, when people talk about “an AI,” it is best to just think of it as “software that someone wrote and carefully adjusted until it worked reasonably well for a single task, possibly including some data-fitting steps.” AI software is not much different from any other software that we use; it just tends to be much more successful for certain problems.
The Practice of AI-Based Artwork
All of this leads to the inevitable conclusion that AI-based artwork is still artwork made by a human. AI software is just software, and there is a long precedent of art made with software.
The earlier computer-generated artwork, such as AARON and Sims’ evolved virtual creatures, were based on “classic AI” methods, i.e., hand-coded rules and numerical optimization. The same goes for procedural image stylization algorithms.
The recent neural stylization algorithms use data-fitting as one step, but they are each ultimately the result of a human writing software, and then experimenting with and improving the software, data, and fit repeatedly until they get results they like.
We can get a window into some of today’s GAN artists’ experimentation via their Twitter feeds. Artists like Mario Klingemann and Helena Sarin tweet about their experiments using the latest image transformation software to create art. They are tinkering, experimenting with code, parameters, and data-sets, playing with the tools until they get great results. To assign authorship of their work to the data-fitting software that they are using would be perverse, and deny the effort and intelligence of the artists’ themselves.
In a few cases, computer scientists have claimed that their software is (possibly) the artist. In each case, they are writing the code, running an optimization, tuning the algorithm and the optimization, and selecting the favored results… just as in all previous computer artworks, like those listed above. I suspect that these authors make these claims due to a lack of awareness of the history of computer-generated art (which this essay attempts to help remedy). Regardless, in each case the claim runs counter to how we define art and artists in the modern world.
If these software packages are artists, then so is the font rendering and page layout package that renders your Word or PowerPoint documents; so is the game engine that beautifully renders an infinite set of new 3D landscapes; so is Harold Cohen’s hand-coded painting generator; so are most of the SIGGRAPH papers we’ve written on Non-Photorealistic Rendering we’ve written over the past few decades; so is any piece of software that creates an image or a sound.
People sometimes talk about the possibility of “collaborating” with an AI. This is misleading, because “collaboration” implies co-ownership and joint high-level decision-making. One does not collaborate with software any more than one collaborates with watercolor or Photoshop.
New AI will be New Tools for Artists
There is a positive lesson from the history of art and technology: new technology gives new tools to artists, who in turn will use it to invent new visual styles and new artistic media. The infusion of new technology into art is one of the main ways that art remains vital. I predict that the new AI technology will play a major role in invigorating art and empowering artists in this way.
Ever since (at least) van Eyck’s experiments with oil paint in the 16th-century—which demonstrated numerous advantages over the tempera and fresco then in use—new artistic tools have enabled new art. Modern technologies like photography, moving pictures, computer graphics, electric guitars, and music synthesizers, have each had profound impacts on our popular and fine arts. Moreover, predicting the effects of these tools is impossible. For example, when Les Paul made his innovations to the solid-body electric guitar in the 1940’s, could he have predicted how the musicians of the 1970’s would use it? Predicting how AI tool will evolve as art forms is equally impossible.
We are lucky to be alive at a time when artists can explore ever-more-powerful tools. Every time I see an artist create something wonderful with new technology, I get a little thrill: it feels like a new art form evolving. Today we are seeing the birth of a new artistic medium, that artists will use to create ever newer and more wonderful creations.
In this essay, I’ve focused solely on our current understanding of AI technology, and argued that it does not create art on its own. Could this ever change—could future developments in AI lead to a true “AI artist”? The definition of art changes across times and cultures; could it expand to allow AI artists? I address these question in Part 3 of this series.