Last October was something of a watershed moment for contemporary generative art with the first sale by auction of a portrait generated with the help of artificial intelligence (AI) for $432,500, titled “Portrait of Edmond Belamy”. The event garnered a huge level of public interest, and received widespread media coverage across several major outlets, including the NY Times, the Washington Post and the Miami Herald, and from leading online Art news platforms Artsy and Artnet, among others. Aside from the final sale price greatly exceeding the original estimated sale price of $10,000, one of the most interesting aspects of this episode was how it was characterized by the media at large.
“Portrait of Edmond Belamy” that was generated with the help of AI. Source: Christie’s
The headline from Artsy claimed that the portrait was “generated by artificial intelligence”, a view perfectly consistent with the NY Times of a “portrait produced by artificial intelligence”. Prior to the actual sale, Artnet had posed the question “Is the art market ready to embrace work made by artificial intelligence?” Christie’s, the auction house who offered the artwork, asked whether artificial intelligence is set to become art’s next medium.
Framing the story in this way - that artificial intelligence was responsible for a fictitious portrait that fetched a handsome sum on the market - is arresting and creates the impression that the portrait was developed without human involvement. It also runs counter to the idea that art is, in essence, a medium for human expression; a highly creative and engaging way for human beings to express themselves, their experiences and how they perceive the world around them. Most importantly, however, such characterizations are fundamentally misleading. The notion that intelligent machines are the sole authors of the art they produce, with little or no human interaction, isn’t an accurate portrayal of how art is generated using AI. In this article, we seek to explore and clarify the process of creating generative art using AI techniques.
Generative Art: Stepping Through the Creative Process
Generative art is a classification for art created using an autonomous system such as a machine learning (ML) algorithm. AI is a large field, but when discussed in relation to art, what is usually meant is ML. In the context of generative art, ML is used as a creative tool for artists to explore and push the boundaries of existing art forms and modes of expression. Pushing known boundaries is of course something artists have always sought to do in practice. At Christie’s Art + Tech Summit this past June, Mike Tyka likened ML as a creative technique to the use of a system that the artist does not fully control, such as the technique of spilling paint onto a canvas: the artists directs the process, but the final result is determined in part by the laws of physics. One might therefore argue that generative art using ML is simply one more manifestation of the typical trend of artists looking to probe, to experiment, to innovate in new and exciting ways.
How Exactly Might Artists Use ML as a Creative Tool to Generate Art?
According to one definition offered by Jason Bailey at the Art + Tech Summit, ML is models built on lots of training data used to make predictions when new data is added. These models ingest large amounts of training data, they learn from the data by finding correlations and patterns in the data to build a correlation structure, and they then leverage that correlation structure to make predictions from new data. In the context of generative art, the ML models employed in the creative process are trained on many examples of various kinds of artworks, and they make predictions in the form of new artworks generated as output.
As suggested by the press headlines covering the sale of the Portrait of Edmond Belamy, there is a tendency to see these ML models as functioning in a fully autonomous fashion, and therefore to see their final output purely as the product of “artificial intelligence”, but in reality, human agency is in fact critical to the ML process for generative art in all aspects.
The artist working with the ML model selects and assembles the training data set for the model to learn from. In the specific case of the Portrait of Edmund Belamy, for example, 15,000 portraits painted between the 14th and 20th century were selected to train the underlying algorithm or model, according to Obvious, the team of researchers who worked with the model to produce the portrait.
The artist also selects the ML model used to interpret and find correlations in the data and to learn about the visual elements that constitute good artwork. In the case of the Portrait of Edmund Belamy, the Obvious team elected to use a Generative Adversarial Network (GAN) model (based heavily on code by AI artist Robbie Barrat), a technology around since 2015 that pits two ML algorithms against one another: one (the discriminator) is trying to detect fake artwork and one (the generator) is trying to generate images that look real, not fake. (An important aspect to note about the GAN model is that the discriminator provides a continuous feedback loop to the generator, nudging it towards more authentic looking images until the discriminator can no longer detect the generated images as being fake).
And finally, the artist selects the output generated by the ML model to be shared or published; Portrait of Edmond Belamy was in fact just one of a group of portraits of the fictional Belamy family curated by Obvious team members for display and potential sale.
Many of us feel ambivalent about, possibly even threatened by, the impact of AI advances in various fields of human endeavor due to suspicions that humanity may be marginalized by them. These sentiments appear to hold true for some in the field of art. As Ian Bogost explained in the Atlantic recently, some deem the promise of AI for art to be a threat: “Given the general fears about robots taking human jobs, it’s understandable that some viewers would see an artificial intelligence taking over for visual artists, of all people, as a sacrificial canary.” Some traditional-minded critics may even fear that contemporary generative art heralds a brave new world in which the fundamental relationship between artist and artwork, or the connection made through the artwork between artist and viewer, is undermined, and the contextual and symbolic meaning of an artwork, derived from the human thoughts and choices of the author and the cultural influences at play, is lost.
There are good reasons however to be optimistic about the impact of AI on fine art. The central role of human agency in the ML process for generative art demonstrates that AI, rather than taking over the artist’s role, presents a great opportunity for human artists to capitalize on. And while the use of AI techniques by human artists is a relatively new phenomenon, the use of technology by artists is a common practice historically that has led to highly positive outcomes. Fine art photography is just one example of a new form of art to have emerged from the practice of artists working with new technologies. As more of today’s artists engage with AI techniques like ML to create art, more opportunities for technology companies to support them in their endeavors will inevitably surface. Collaboration between artists and new technologies is nothing new, and if the past is anything to go by, we should hope that the current embrace of AI as a creative tool for artists continues in earnest.
What are your thoughts about the potential impact of AI on fine art? Please share them in the comments below.