Art in the AI of the Beholder: How Queer Art and Artists Challenge Algorithmic Aesthetics

Critical computing scholarship on generative AI often focuses on harms, such as representational harms against queer people and allocative harms against artists. However, queer artists' agency to make use of or resist these technologies is under-considered. In the first two studies of this dissertation, I examine queer artists' perceptions of and material engagements with generative AI. First, I conducted a medium-term workshop study where I provided 13 queer artists with access to two major corporate generative text (GPT-4) and image (DALL-E 3) models. Participants struggled to use these models due to various normative values embedded in their designs, such as hyper-positivity and anti-sexuality, but developed strategies to work around these limitations. One of the limitations of this workshop methodology was that I required participants to use the provided generative AI models, potentially excluding artists who refuse to use generative AI. To understand these more critical perspectives, I interviewed 15 queer artists to see how they are responding to the proliferation of generative AI. I detail how interviewees refuse and resist the development and deployment of generative AI models that misalign with their aesthetic and moral values, such as calling out suspected generative AI use. My prior work demonstrates the power of queer artists to re-appropriate, resist, and refuse generative AI. However, their agency is still shaped by generative AI developers' design choices. In particular, participants in both my workshop and interview studies praised the surreal artistic potential of text-to-image-models, while critiquing the increasing realism of newer models. This begs the question: What do developers aesthetically value in the production of text-to-image models? The final study of my dissertation examines this question by conducting an audit and trace ethnography of the LAION-Aesthetics Predictor (LAP), an AI model widely used to curate training data for and evaluate the performance of visual generative AI models. By evaluating LAP on three datasets of images, I find that LAP considers photorealistic images of landscapes or women to be most “aesthetic.” This is perhaps because a majority of the images and aesthetic score ratings used to train LAP were scraped from a photography competition website in 2012. I conclude by discussing the politics of taste amid the rise of generative AI, highlighting the potential for queer aesthetics to imagine generative AI otherwise.