Intent in Generative Art

A generative artist faces unique challenges trying to convey intent through their art. In other visual mediums the artist directly manipulates the work. Viewers can then look at the subject matter, composition, and technique used in the work to try to glean the artist’s intent.

How intent usually flows to the viewer from the artist.

In contrast, a generative artist manipulates not the visual output, but an algorithm which is responsible for creating that visual output. This adds an extra layer between the artist and the viewer that the intent must filter through.

How intent flows to the viewer from a generative artist.

It may seem that because the artist is responsible for every detail of the algorithm, there is no reason to believe that any of the artist’s intent would be obscured by this additional layer. However, even in the most extreme case where there seems to be little difference between the algorithm and the visual output, this pipeline places an additional burden on the viewer to try to interpret the work. Consider a deterministic algorithm responsible for drawing a series of shapes in particular locations on the canvas. This algorithm produces the same output every time. The same visual work could easily be achieved by someone using image-editing software like photoshop. What distinguishes these two works?

Left: a digitally drawn work. Right: a generative work.

The answer is: nothing... until you tell the viewer that one of the works is generated by a computer. Now they know that an algorithm produced this work, but they know nothing of the inner workings of that algorithm, which begs many questions. Do certain natural ways of expressing ideas through code influence the composition? How do the limitations and difficulties of working with code affect the artist’s ability to express themself? Could this program produce other outputs? Satisfactory answers might be found in the algorithm, but if the viewer can only see the visual output, they must either proceed under assumptions that could lead to a false impression of the work.

This uncertainty is exacerbated by the fact that generative programs incorporate pseudorandomness to produce an array of possible outputs. The artist designs an algorithm not with a single output but the entire space of possible outputs in mind, which means their intention is most directly reflected in that entire breadth. Viewing a single output does not give access to the artist’s thoughts on this higher level of organization. Even viewing multiple outputs provides a skewed view of the probabilities governing the algorithm because there is an infinitesimal chance that a finite sample is perfectly representative of underlying probability distribution.

Archetype #61 by Kjetil Golid. Viewing only this output might suggest the entire project is about repetition.

The larger the sample size becomes, the more representative the sample would appear to be, but is also becomes increasingly likely that outliers present a new challenge. How should rare outputs that differ wildly from the norm be considered in the context of the work as a whole? Are they possibilities fully considered by the artist and therefore just as representative of their intent as any other? Do they naturally emerge from the algorithm even though the artist never conceived of them? Or are they edge cases which went unaccounted for, resulting in bugs or otherwise unintended behavior? A viewer might have difficulty discerning the role of outliers even if the algorithm is understood in its entirety, as it requires them to guess at the artist’s thoughts in designing the algorithm in a particular way.

Fidenza #388 by Tyler Hobbs. How intentional is this output?

All of these questions are natural ones to ask in an attempt to understand a generative work. Most of them are questions about the algorithm itself and how it functions. They are not questions about the artist’s intent, but merely a prerequisite before that intent can be interrogated. The algorithm, and not the artist, becomes the natural starting place for viewing a generative work, and focus is drawn away from the human element in the art.

It is key to note that the viewer’s prior knowledge is hugely influential in how they perceive a generative work. They were only able to distinguish the two identical images when they learned one is computer-generated; had they never learned this fact, they may have understood the meaning of the work to be entirely different. This burden of knowledge extends to the viewer’s understanding of algorithms. A viewer unfamiliar with generative art might ask why the artist chose to use fluid forms in their work, while a familiar viewer might visually identify the flow field algorithm used in the work and ask what distinguishes this work from others with the same algorithm? Trying to give both of these viewers the same experience is a daunting task.

There are many challenges introduced by attempting to translate artistic intent through generative means. The purpose of highlighting these challenges is not to suggest that making generative art is somehow more difficult than other forms of art. I believe analyzing how different generative works navigate or subvert these challenges can lead to a better understanding and appreciation of those works. Take Maya Man’s FAKE IT TIL YOU MAKE IT as an example. While the individual messages in each output can be interesting or funny, the main point of the work is that these kinds of messages can be constructed entirely through algorithmic means, pointing out how robotic affirmational content on social media can be.

I also hope that making generative art with these challenges in mind will help others to consider novel approaches in their own work as it has helped me. In my upcoming work Through Curved Air I want to make it clear through the visual output alone that there is a human responsible for its creation. Since it is natural for the viewer’s focus to begin with the algorithm, I took two complementary approaches when designing mine. The first approach is to make the algorithm as simple as possible. The outputs are constructed entirely from copies of the same shape. The result is that each output can have a unique texture, but the origin of this texture is very plain to see, so the viewer does not have to spend much time asking how it works. The second approach is to design an algorithm which cannot be understood through visual observation. How the algorithm arranges these repeating shapes cannot be intuited even from many outputs. If the viewer tries and fails to understand how it works, they are forced to move on to asking questions about the visuals themselves; why did I choose to use these forms in my work? It will always be obvious that a computer was involved in the construction of my work, but I have attempted to make it equally clear that I am sitting on the other side of that computer.

Pop/Generative Art in Linescapes