Category trouble: AI’s queer problem

This essay subjects AI algorithms and code more broadly to a critical code studies analysis to argue that coding architecture in its current form is unable to accommodate queerly fluid visions of the world because of the elemental importance of categories for code to function. I first demonstrate how coding is necessarily a categorical enterprise, whose condition of possibility depends on eliminating uncertainty and in-betweenness. Then, I turn to an analysis of image recognition models (specifically, the convolutional neural network LeNet) and the word embedding architecture of large language models to show the strategies with which these models encode fundamentally ambiguous and fluid concepts through fixed and clearly demarcated categories. Finally, I apply the insights from the classification and generative models examined to a case study of recommendation algorithms in social media to demonstrate the real-life effects of their categorical logic. Ultimately, those effects amount to the construction of digital worlds that exclude fluidity and enforce strict boundaries between people and identifications, siloing different groups and silencing alternative perspectives. To achieve a more just vision for the future, then—one that respects and furthers the coexistence of diverse, uncertain, and shifting views of identity and belonging—it will be paramount to fundamentally rethink the structures of contemporary AI models.