Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
There is a certain type of artist who yearns to go beyond the contents of their own head. I have long wished for a truly malleable painterly space, one that has the velocity and dynamism of great abstract painting, only fashioned out of figuration, made with images.
There is also a history of artists embracing technological innovation: silkscreen, video, neon. Everything was new once, even oil paint. Photography was thought to signal the death of painting until artists as diverse as Thomas Eakins and Edgar Degas started using it in their work. Now we have artificial intelligence. Could AI help realise my vision of a fluid pictorial space, exempt from the laws of gravity?
The AI art that I had seen was not promising. The more futuristic the imagery, the more banal the effect; the images seemed to have no inner life. Was it the fault of the technology or the user, a failure of the car or the driver? What did the machine not “understand” about visual art? For one thing, digital imagery is not to be confused with painting. Pixels have no surface texture, no physical presence at all. Painting, no matter how cerebral, is an object in the world. I am a painter, not an engineer; I’m on the side of things made by one author, by hand. And yet, I’ve spent the past two years teaching a computer to think like an artist.
Artificial intelligence is essentially an averaging machine for the universe of images taken from the internet. The basics of representational painting were not part of the machine’s neural network. Two things, both essential to visual art, had gone missing in AI’s indiscriminate data scrape: specificity and intentionality. Most critically, it had no sense of how to treat the edge, the place where two shapes meet. How a painter resolves the edge goes a long way in determining his or her style. Is the edge brushy and expressive, or tight and constrained? In a digital image, an edge is often approximate, as the machine tries to guess the next move, but in any event is just differently coloured pixels; it lacks intention and harbours no inflection.
The adage from the early days of computing, “Garbage in, garbage out”, was still relevant. To get anywhere at all, the machine would have to be sent to art school. Through the venture adviser Tom Cohen, I was introduced to a talented young technologist named Grant Davis. Together, we set about designing the curriculum.
We first trained the machine on a handful of paintings by Giorgio de Chirico, Edward Hopper and Arthur Dove, three 20th-century masters who exemplified, respectively, the drama of perspective as an organising device, the ability of value pattern to convey a sense of volume, and the lyrical use of the colour black. To those luminaries, we added a tightly edited selection of my paintings from the 1980s and ’90s which emphasise a collage-like, juxtapositional approach to composition. The machine’s visual IQ went up several points, but the edges were still weak.
We dialled back the art history lesson and retrained the machine on a series of my paintings from 2001: extravagant, fragmented compositions loosely based on a 19th-century opera backdrop. This romantic, pastoral image — a man and a woman sitting on an embankment, with a lake in the middle ground and mountains in the distance, was painted in eye-popping colour harmonies. As a final step, we added paintings of mine in which forms are rendered in decisive, expressive brushstrokes.
As we fine-tuned the model, Davis was independently developing a software program for visual artists called “Wand” which eliminates the need for verbal prompts. This works like a joystick, a lever that moves along a continuum from “similar” to “dissimilar”. We fed the machine a complexly hued “Pastoral” painting, placed the dial at the midpoint between the two extremes and held our breath.
The machine now proved to be an excellent student. My compositions were melted down, put into a blender and passed through a sieve. The results were clearly derived from my work, with many familiar elements, but the reconstructions were also unlike anything seen before. Since the machine had no idea what it was doing, it could break the rules of continuity, scale, anatomy and pictorial logic with impunity. (The sense of the machine not knowing its transgressions was key; artificial unintelligence is more to the point.)
The images were still just pixels on a screen. I selected the most promising scenes, and had them printed on canvas. These prints then formed the backgrounds on top of which I painted as I normally would, layering a collage of figurative elements that seem to emerge from the background swirl. Male and female torsos; gesturing hands; precariously balanced teacups are painted using a big brush with highly saturated colours. The compositions are dynamic; everything sails through the deliquescing, topsy-turvy space. The paintings speak with the gravitas of art. But they are also linked to an uncanny “future-present”; they are palimpsests from a dialogue with my earlier work — like “a duet of one”, as a curator friend put it.
AI is a powerful tool, but it is still just a tool. It is not, for now, the author of anything. It works with what it has been taught, in the way it has been directed. The “thinking” part of the machine, the algorithm, does seem to have a drive to break pictures down and recombine their component parts, to scramble their spatial orientation but preserve and even amplify their emotional subtext. What could be more expressive of this moment?
David Salle’s ‘Some Versions of Pastoral’ is at Thaddaeus Ropac, London, to June 8, ropac.net; and his work is at Frieze New York, May 7-11, frieze.com
Find out about our latest stories first — follow FT Weekend on Instagram and X, and sign up to receive the FT Weekend newsletter every Saturday morning