Beyond black boxes and the AI sublime: critically assessing the code behind commonly used machine learning models

Abstract Discourses on artificial intelligence often fall into two interrelated obfuscating narratives: the black box and the sublime. The concept of AI as a black box shrouds algorithms in mystery as the sublime blinds subsequent critical scrutiny of how these models are integrated into important systems. Scholars embracing a sociotechnical orientation counter the “unknowability” of algorithms by analyzing the complex constellation of elements that underpin these systems, along with how these arrangements function in practice. One such understudied element by scholars of the humanities and social sciences is the coded architecture of commonly used AI models. This article is inspired by critical approaches that survey AI models through a combination of technical definitions, critical theory, and how different algorithms operate via actual code. This includes evaluation of the various technical and sociotechnical strata at which algorithms exist, followed by distinctions between algorithms, symbolic AI, and connectionist AI. I specifically highlight two machine learning models of linear regression and decision trees to examine what these models can and cannot do via questions they can be used to answer, their limitations, and examination of how outputs can become biased in common use settings. Consequently, this article aims to demystify the opacity surrounding one important aspect of AI by assessing the code of commonly used models, and, in doing so, avoid the discursive trap of opacity that reifies the mystique of unknowable and all-powerful AI.