Selected Works:
- Joh30461220
- 3361175
- E Mountain Rd
- 10.82720243:1436.7095.993725
- MO3010060
- Documenting Paragon for Systems of Release
Essays:
About
by Briar Smith
...The copy or representation of the self becomes an “other” to then act as a tool in the practice of comparison to the “real” or embodied self, furthering a hegemonic structure of systematic binaries. For marginalized bodies, this othering of the copy is detrimental to the ways that we are able to understand and analyze the copy as a virtual object. The othering of the “false” representational form deems it unable to be further analyzed beyond just its “falseness.” When the false is treated as a “true” form, only then are we able to explore the way in which these misrepresentations distort lived realities through the dematerialization of the physical form into binary information to be used, mistranslated, and regurgitated through the machine’s vision.
In order to dissect this hierarchy, first we must understand a technological fixation on the representational proximity to lived reality, or to be more concise, the verisimilitude of virtual representations. Year after year, developers of virtual worlds aim to more accurately represent the “real” world than those had previously. Harun Farocki’s films Parallel I & II explore the linear progression of computer graphics and their drive to most closely simulate the “real.” Through the span of just a few decades, trees evolved from blocks, rectangles and single lines of pixels to complete armatures able to respond to the virtual environment in which they exist. Water represented through grids of lines and dots developed into complex systems of hyper-realistic, algorithmic waves. Clouds moved from fixed, white, clustered pixels to expansive, translucent masses that shift and evolve as they pass through the virtual sky.
Farocki, Harun. Parallel I & II. 2012. Germany: Farocki Filmproduktion, 2012.
Representation as a technological tool mirrors the development of other digital tools; placing value in objects that are meant to make our lives easier. Therefore, recognition in this sense is made easier by stripping the objects of all nuance through a representation of the virtual form that is closer to its relative “true” form. The aim toward verisimilitude acts as an equalizing structure that reinforces a general distaste for a copy that appears as just that, a copy. Instead of allowing the “false” to exist as false, the developmental drive toward better graphics and higher fidelity software aims to transform the copy by defining its worth based on its proximity to the “real,” creating an object of hyperreality in the process that over time will become closer and closer to imitating reality. For many, this becomes a deep-rooted fear, intrinsically taking shape as a sort-of fear of automation.
The strive toward verisimilitude informs the hierarchical structure that defines the copy as less than its true counterpart. When navigating a system that promotes the creation of tools that aim to make the virtual copy dissimilar from its origin, the copy is not only undervalued, but becomes othered by notions of the material or “real.” In Anne Freidburg’s the Virtual Window, she delineates the real from the virtual through the relationship between materiality and immateriality, a “key marker [of] a secondary order in the relationship between the real and its copy.” While what is seen as “real” is visible to the eye, its copy occupies temporal space, constituting it as a virtual object. The virtual object’s liminality, and therefore lack of materiality, causes it to become othered, separate from what we know or can recognise as “true.”
The object’s often intangibility forces it to be separate from the rest of the tangible world. In order to redefine this hierarchical structure, we must inverse this order, instead valuing the copy over the “real.” Only then, when inverting this structure, can we fully dissect the copy and its ability to not only have representative substance and value, but also address the misrepresentative weight and harm that it can (re)produce.
To fully examine the copy and its misrepresentations and mistranslations, we must again return to the idea of verisimilitude. Not only does our fixation on the proximity to reality affect the resulting representational form, but also shapes the way in which the systems are produced, specifically within the training of language models, diffusion models, computer vision, and image recognition models.
Having already established the developmental focus on systems that produce outputs more closely associated with the real world, the goal of these neural networks is to generate outputs unrecognizable as AI. Those developing these systems rely on insurmountable amounts of data to train and ultimately shape these networks, believing larger amounts of data produce models that are closer to replicating the work of a human–more data equals more diversity, more representation equals better treatment.
However, when scraping large amounts of online data, blindspots tend to be overlooked. In a study conducted by several professors at the University of Washington, they found that most of the data (in the US and UK English) was taken from user-generated websites such as Reddit and Twitter. Specifically on Reddit, the users are more likely to be men, a higher percentage being of a younger demographic. Because of these overrepresented demographics, misogynistic, white supremacist, and generally hegemonic viewpoints tended to be more abundant and consequently retained.
In the case of underrepresented marginalized demographics within the training data, many of the terms used, such as reclaimed derogatory terms and slang, were flagged as hate speech or vulgarities, causing large amounts of data to be omitted. Similarly, when scraping data related to social movements and activism, groups that have widespread media coverage tend to be overrepresented, while groups or movements less covered by the media are subsequently washed out.
When the biases within the training data become encoded within the algorithmic genesis of these models, their output, text or image, becomes equally as biased, furthering the violence against marginalized groups through the creation of virtual representational media.
Simply put, all it takes is a request to generate an “image of a woman” in ComfyUI to produce pornography. Or when asking it to generate an image of a doctor, nine times out of ten, the image will be of a white man. By focusing on the verisimilitude of generative models, much of the data goes overlooked, not only reflecting the violent reality we occupy, but virtually recreating it in the process. Why focus on creating a tool that's worth is based in its proximity to reality when reality is unsafe for a large amount of the population?...
Prompted imagery made using ComfyUI.