Exploring aesthetic theory of new media AI generative networks. Follow-on post from “Herding Electric Sheep“, which was inspired by the work of Refik Anadol Studio.
Anadol creates AI networks which, when fed data sets, are able to analyze and interpret the data into a new work of art. These AI networks are a type of GAN, or Generative Adversarial Network, which work against themselves to “learn” or “grow” and adapt. The artist/programmer defines the parameters of execution and can have significant to little influence on how the AI then processes the information. In the case of Anadol and information presented in artist statements, it seems that he and his team are highly hands-on in helping to curate and guide the AI responses into something which they find to be artistically “successful”.
In this article I want to explore some of the aesthetic choices being made by Anadol, as well as some of the questions that might arise around the art theory of his projects.
Some of these questions are specific to Anadol and his work, such as why the output of the systems looks the way that it does. Other questions are perhaps more general in terms of how AI systems are being put into use to generate spaces, contexts and content.
I am no expert in the technicalities of computer programming, nor am I broadly versed in AI. These questions don’t, I believe, require that sort of expertise in order to open up useful ideas.
Several of Anadol’s AI works (such as “Artificial Realities“, “Unsupervised” and “Renaissance Dreams“) are presented on digital displays which seem to depict shallow visual spaces. Within these spaces, opening into the image, are dynamic visual representations of colors and forms that are constantly in flux. Many seem to be presented as the surface of a liquid pool as it undulates and shifts in color and texture.
These representations are being generated in real time and do not repeat, as the AI continually churns out and adjusts the visual output. What is represented on the website is necessarily only a selection of moments preserved as still images, or short sections of video, which show how the visuals shift and change.
Underpinning these visuals are a curated data-set of visuals which have been fed into the machine. It is the selection of visuals as well as the setting of parameters which constitute what the artist is hoping to achieve.
For “Renaissance Dreams” the program was built around a data set of paintings created in Italy between 1400 and 1700 within what is generally agreed upon as the Italian Renaissance. For the MoMa show “Unsupervised” the artists was able to use a data set of all works cataloged within the museum. From these data sets the AI is “inspired” by and “learns” the images which it has been fed, and uses elements from those images to create something new.
It is a sort of meta-view of the works taken as a whole.
But what is the significance of the fluid dynamics of these representations? Certainly none of the works within the data set (from the Renaissance anyway) were video or contained actual movement of any kind.
As a method of constantly generating images one might see the need to move from one representation to another, the flowing makes conceptual sense but isn’t the most obvious option. Simply showing a slideshow, or using now-common animation techniques to simply morph one image into another, would have been just as easy.
Aside from flowing one image into another we are presented with the curious choice that these dynamic visuals would also appear to have dimension. Not only would they ripple and flow, but they would actually seem to splash out at us, undulate and roll in the depths.
What does watching psychedelic whitecaps tell us about the oeuvre of renaissance oil paintings?
I find both of these choices to be somewhat at odds with the projects as described. If we are interested in looking at what AI sees as connections between works which are static and two-dimensional, how are we to perceive these aspects when they are rolling and bubbling across the screen?
I see this as a problem of relations, in which the source material is so far removed from the output as to be unmoored and unable to hold dialogue with itself. It is interesting to know that these images derive from a meseum’s collection of famous works, but so far removed from our ability to make those connections that it might as well have derived from images of jelly beans.
As I will continue to push on, the capabilities represented in AI may be less limited in their technical abilities than they are in our ability to connect them back to useful relationships within our lived experiences as humans.
Not all of Anadol’s works take place as pools of liquid. Another version of the “Unsupervised” show at MoMa did take a two-dimensional approach to the outcome, with flattened images morphing into one another.
Removing the element of depth it becomes instantly easier to begin focusing on elements within the images that seem to be marks which might have come from human hands. At one moment the image may present itself as calligraphic textured brush strokes, while at another it might appear more as modernist color blocks. Some moments become more architectural before dissolving into impressionistic glowing patches without structure.
The content always shifts and changes. Seeing elements that remind us of other painters, or movements, time periods or locations, is where I found the most resonance with the piece. It was much easier for me to engage in a dialogue about how the work was interpreting the museum’s collection as a whole.
Like watching TV, or doom scrolling, the game becomes one of waiting for these next moment of connection to show up. If seeing an combination of visual elements forming on the screen which resonate with other works that we know is somehow satisfying, then I find myself waiting around for the next moment like that to occur. Thankfully the videos which have been posted are necessarily a very small portion of the “finished” work.
Why this interest in continuous generation of content?
This seems to be something which exists beyond the circle of AI in art, and appears as an element in other AI projects. Perhaps because it is because generation of the new from the existing is a major technical strength of the technology.
I suspect the desire to continuously generate new content is tied to the human desire for “more”. Why generate one image, when it is hardly any extra energy to generate 2, or 5 or 1000?
In the case of Anadol it seems as if continuous generation is also a hallmark of how the function operates. The AI, in a constant competition within itself, is adapting all of the time. Perhaps it is the only method by which one can truly appreciate the tools in real time. To see an image generated today will not tell us what it might have learned to generate tomorrow.
And yet, that is how every human artist must operate, only ever able to create what it is that they have been practicing to be able to create.
Never-ending art creates at least two significant problems.
First, the scope of the work can never be grasped as a whole. With not official end, and little discernable relationship between the visual representation of one moment and the next, it is impossible for the viewer to anchor on any sort of perspective or relationship with the work. Is it enough to watch it for one minute, or thirty, before one begins to appreciate it? Each analysis is pre-empted by the variations which have not yet come to pass.
Second, without an anchor point, it is impossible to hang analysis or interpretation on any one visual aspect. Much as I am struggling with here, the work itself becomes not only difficult to properly describe, but also difficult to internalize. No method of compositional analysis can be put into place, nor historical perspective, comparison to film or TV. Perhaps it can be compared to the never ending feed of social media, but it does so in the gestalt of its existence rather than in specific moments of representation.
In that context, anything that is unending and constantly in flux, could be read in the same manner.
If our interest is in exploring black box systems that follow their own rules about development, and do not cease to develop in the face of their own complex environments, we might be just as satisfied to spend time a garden.
Our entire world is a generative adversarial network of chaotic forces acting within their own systems. While it may be colloquial to think of our world as static and stable, the reality of any given moment is the struggle for life or decay of what has been constructed. What does the creation of a digital version tell us about the world we already inhabit?
In Anadol’s work “Artificial Realities” he and his studio create AI networks trained on images of coral reefs in order that they might in turn create digital versions of non-existent reef-like structures. The images do manage to attain a sense of beautiful structure that one might expect to find under the sea. Doesn’t it, however, beg a question about the alien beauty of the actual coral reefs which exist in the world around us? Rather than collect data to feed into an artificial version, would we not similarly marvel at the wonders of what nature has designed over millennium?
Computers can already be programmed to generate the structures of a reef based on the actual real world models of growth patterns. It is not a fundamental understanding of how they are created, or the inability to create convincing facsimiles which AI brings to the table.
The trick which AI seems to perform most uniquely, is the generation of novel versions of things based on parameters. It is the sheer scale and speed at which inputs can be developed into finished versions. What does the creation of reef-like structures (again, shifting and transient) tell us about the examples on which they are based?
That line of questioning is where I find myself often getting stuck. AI can create outputs which are fascinating in their novelty, their newness or foreignness, and the magnitude of their presentation. That, however, presupposes that what is novel is also more worth while. It leans on the idea that something we have never before seen is valuable for that very reason.
What, for humanity, is novelty? It is not a question of basic addition. Adding one new experience does not always add new information. Humans learn new information within a context. We rely heavily upon stories and previous experience to integrate something new into the paradigm within which we operate.
If it was enough to read lists of dates, names and places, then history could be reduced to a list. For most of us we need to understand a broader perspective on the information. We need to think about the story, the origin, the driving factors and influential context.
Leonardo da Vinci was insatiable in his quest to learn about optics, anatomy, atmospheric effects and any number of other factors which go into how objects may be perceived under different conditions. He wrote about “composition” not as the orientation of different elements into a picture plane, but rather as the imaginative activity of the mind as it makes decisions about how a figure will appear on the canvas. For him, the act of composition required deep knowledge of the subject. It was an act of being able to hold the subject within one’s mind’s eye, not as the two dimensional perception that we see before us, but as a complete whole in the round. The artists greatest act, as I choose to read him, is to be able to take what we perceive and understand about a subject, and make mental adjustments to it so that it can be represented back to the viewer in a new configuration.
This conception relates two important aspects which I find missing as we try to align ourselves towards such works as we have been discussing.
First, the the knowledge with which the artist must operate should be as deep as possible. In order to build any knowledge, and in order to understand anything beyond rote memorization, humans need to build mental connections to that thing. We may develop knowledge through a number of different avenues, not least of which is a non-rational aesthetic understanding akin to muscle-memory or the lived experience we have working with the information in our daily lives.
Second, for Leonardo the act of composition wasn’t the act of the novel from nothing. It was the adjustment of what was known in order to present it in a new way, or the way most appropriate to the situation. The best leaps in understanding are not sudden exposure to what is alien, but incremental exposure to what we thought we new so that we can deepen our connection and understanding.
AI networks are often put to use in the analysis of massive amounts of data. In some of these cases the works were trained on hundreds of thousands, or even millions, of images. What they are generating out of connections made between those images, are visual solutions to questions that humans could not hope to pose, or understand the answers to.
In many ways, AI generated art seems so meta that only other AI networks could possibly have any sort of understanding of the results (though that sort of understanding is also beyond their capabilities).
Though fascinating to look at, I fear that the generated art works of this sort of programming is landing, if indeed it has an aesthetic value to be mined, at a level so far beyond human comprehension as to make it functionally inhuman. Pretty to look at, but ultimately less informational than an aquarium full of exotic fish. Staring at alien creatures from the deep at least comes with the limits of our scientific knowledge of, and aesthetic connection to, similar sea creatures and experiences in the ocean.
It is a tricky area to walk, for I do not want to deny aesthetic value to anything which exists alongside of us.
I do find it very difficult, however, to see where one might be able to form a meaningful connection with these works. Especially given their penchant for continual un-resting change at pace.
Perhaps we must simply keep watching in order to build up our experiences with this kind of art. I only fear that that instinct, to keep watching one more moment lest we pass up the one vital spark we have yet missed seeing, is more than anything a commentary on the attention-commodity marketplace, in which FOMO and click-bait have stripped the virtue from not-looking.