Why Algorithms Will Never Make Great Art
This A.I. portrait just sold for $432,000 at Christie’s. Why is it so valuable?
The composition is awkward and the brushstrokes are a touch clumsy. Even so, the “Portrait of Edmond de Belamy,” which sold at Christie’s Auctions & Private Sales in New York City for $432,000 (£337,000), looks suspiciously like art.
The creative energy behind the image is not a person but an artificially intelligent system built by a French digital collective called Obvious. The A.I. system was programmed to compare its portraits to thousands of authentic, painted portraits, and to keep adjusting its image until it is unable to detect a difference between the two.
In the province of modern art, there are countless examples of art that have been created by mechanical mediation, or even completely surrendered to the vicissitudes of a predetermined process.
Undoubtedly, the A.I. painting looks like art. Not good art, I would argue, but since it fits into the convention of three-quarter view portraiture, is contained within a gilded frame, and is hung on a wall in an environment where we expect art to be displayed and sold, it looks like the real deal.
But is it?
Countless works of art — particularly modern art — have been created by mechanical mediation. Some pieces completely surrender themselves to the vicissitudes of a predetermined process. Perhaps this is how we should view the A.I. portrait. John Hilliard, for example, made Camera Recording Its Own Condition (1971) — a work consisting of 70 snapshots taken by a camera aimed at a mirror. Each snapshot showed the moment of exposure, and the shots differed by film speed, exposure time, and aperture size. More recently, Anish Kapoor created Shooting into the Corner (2009). It consisted of a canon that periodically fired 11-kilogram balls of wax into a corner of a room, allowing them to hit the wall and splatter.
These are just a few examples of “process art,” or art made according to methodology rather than explicit intent. On these terms, the A.I. portrait could be considered part of the tradition of process-driven art.
Some commentators have argued that the real artists behind this work are the computer programmers of the French collective. They used the A.I. system as a tool, a vehicle of artistic realization, like a complex paintbrush.
And yet, I suspect the French A.I. programmers had no intention of contributing to the acclaimed history of modern art, or “process art.” I suspect they wanted to create something else altogether.
What seems to set this so-called “artwork” apart is how programmers enable the A.I. to learn for itself, thereby giving it a degree of autonomy. They may have intended for the A.I. to generate a portrait, but the exact nature of the end result was unknown. In this way, the artwork was created by the computer, not the human programmers.
Or was it?
What if the portrait had been less conventional? What if it were as chaotic and distressed as a Willem de Kooning painting or as awkward as a Picasso? Would the programmers have called it art? I suspect not. I suspect the end result we see today is the culmination of a long process of programmatic tinkering and fine-tuning. I suspect significant human intent to make sure the A.I.’s work conformed to what we typically think of as art.
Taken this way, the A.I. portrait is much less of a challenge to our ideas of what constitutes art. Instead, it seems to be a tediously rendered confirmation of it.
Think of it another way. Imagine a computer program written to compose a piece of music. This is by no means a far-fetched idea. Today, it’s possible to listen to music that has been created entirely by A.I. Tempos and dynamics can be tweaked with a simple shift of parameters. One such team of developers says they’re “training deep neural networks to understand music composition at a granular level.”
Now, imagine this musical A.I. algorithm came up with a piece similar to John Cage’s Variations II (1961) — a composition that some deride as nothing but random sounds or meaningless noise. I suspect we would not credit the computer with much creativity, even if the real composition is a high point of modernist classical music.
The core conundrum is how we, in the 21st century, determine what is and is not art. As A.I. generates more of what entertains and delights us, how do we make an informed distinction?
In the place of a clear-cut definition, it may be more accurate to think of art like any wide-ranging human activity: as a family of similar but assorted practices that manifest in different ways at different times.
Could we ever imagine a computer having intention, passion or sophistication, or indeed being able to think in any way?
To draw the strings of this vague picture together, the philosopher Guy Sircello asks, “In what manner do we normally assess a work of art?” He looks at the parameters we tend to use to evaluate an artist’s output, such as “competency,” “coherence,” “seriousness,” “maturity,” “sanity,” and so on, and suggests that it is through these parameters that we are able to approach a work of art as a work of art. His philosophy even helps us makes sense of John Cage’s Variations II in the context of the artist’s intentions and the piece’s historical moment.
In Mind and Art, Sircello writes, “Naming these parameters simply points out the sorts of considerations which might be relevant in particular descriptions of artistic acts, leaving it an open question which of these parameters are relevant, and to what degree, in particular cases.”
Could we ever imagine a computer having intention, passion or sophistication? Could we imagine a machine that’s able to think? According to philosopher Ludwig Wittgenstein, to ask the question, “Is it possible for a machine to think?” is to misunderstand the criteria by which we ascribe such terms. It is as nonsensical as asking, “Has the number 3 a color?”
Wittgenstein suggests that if we are to ascribe the capacity of thought to a machine, then it only makes sense to do so alongside the full scope of associated nuances — that is, the qualities of mind that often accompany thoughts.
After all, my human thoughts are cut through with hesitancy, fallibility, open-mindedness, rashness or prudence, and shrewdness or stupidity. For an action to be described as resulting from a thought, it must sit within a broader scheme of capacities and dispositions, not least of which is the thorny concept of free will.
I suspect that both Wittgenstein and Sircello would claim that artistic intent, according to its full definition, is a capacity that can only usefully be ascribed to human beings and not A.I. systems.
This returns us to the A.I. portrait of Edmond de Belamy. I began by saying that the painting looks like art. As an art historian, my next set of questions would be: What tradition is the artist working in? What was the historical context? Is it a sincere painting or does some degree of irony underlie it? What is the attitude of the artist with respect to the sitter: respectful, sympathetic, or disdainful? How does the artist use the materials? And so on.
These are hardly questions one could seriously ask an algorithm, any more than it would make sense to ask a book why it chose the pictures on its pages.