A painting created using GANs (generative adversarial networks) sold for $432 000 at Christie’s today.
There is a thread on twitter raising complaints about the originality of the piece. The collective that produced the painting (Obvious) used modified versions of open source code, a lot of which were based on Robbie Barrat’s data scrapper and implementation of dcGAN. This work is itself based on a chain of research originating from Ian Goodfellow back in 2014.
The controversy is a nuanced issue and a very interesting one for me. On one hand, the work sold at auction is by no means the state of the art in GAN results, as even the authors admit. And much of the media coverage around the piece does not pay homage to how much the work is based on Robbie’s code. But successful art requires context; it is more than the piece itself. For instance, Damien Hirst employs a factory of workers to produce his dot paintings. But he conceptualized the context for their success. It is oftentimes not even the artist that casts their work in the right light to highlight a broader significance; It’s a dialogue between art critiques, institutions, collectors and auction houses that imbues a piece with that, like offering social commentary, or challenging ideas of old and new. The fact that the work made it to the prestigious auction house and the price it commanded at auction is a reflection of the public’s interest in AI and also it’s misunderstanding about what AI generated art is.
The method of creation, namely by training GANs and using carefully selected datasets, is also very important to the context for the artwork and therefore important to its sale price. Which is why the complaint about the work being derivative carries weight. This is a catalyst for a broader conversation about AI augmented art.
A part of what makes novel art is to explore how to create with new constraints and new tools. Now the tool is a neural network model. We focus not on the physical act of painting but on curating data for the model in order to tease out interesting visual outputs. This process is similar to exposing ourselves — a model — to new data from the world to draw inspiration from. To create with that model is to learn more about its limitations and expressive capabilities, and to manipulate that with intent. All of these will define what it means to create meaningfully with AI tools.
Obvious’s original story to sell this piece was that “AI” created the body of work. A question now is whether the real story about creating art with AI — which is very much with human involvement, curation and input — will remain interesting for the art world. The price commanded at auction by this work has at least demonstrated that there is significant interest in the tech that created this content. That interest may be misinformed about how truly intelligent the tech is, but if the appetite is there, then there’s an opportunity to communicate this tech more broadly and perhaps make it more accessible. Is there a way to tell the story so that it’s engaging and educating for a broader audience?
If so, then “Edmond de Belamy” will not be the only piece created with the help of GANs to be sold so lucratively at auction.