As literature and language scholars have begun to learn, what is commonly called artificial intelligence (AI) today is usually a form of data-driven machine learning. Such “intelligence” does not derive from a human-like capacity to critically reflect on the world but, rather, through the mining of data through arsenals of computational power. Universities have embraced AI as the next big thing. Dominated by marketing hype and misunderstanding, discussions of AI cry out for the critical approaches and historical perspectives of literary critics and others. This panel follows on last year’s MLA-organized roundtable, “AI and the Human,” and proceeds in concert with MLA’s discussions on new “working conditions.”
Sophia Hsu opens the panel by framing the temporal ramifications of biased systems; Pamela K. Gilbert tightly focuses on a specific genealogy; and Lauren M. E. Goodlad closes with AI’s ontological implications. The panel offers an informative “way in” to a topic of urgency while connecting to ongoing research questions among nineteenth-centuryists and others. As AI increasingly dominates strategic initiatives in teaching and research, humanists must join the growing conversation over this so-called fourth industrial revolution.
In “Statistical Reasoning and the Limits of Predictive Modeling in AI,” Hsu connects the statistical problem of induction to debates about AI’s predictive capacity. AI’s usefulness lies in its supposed ability to accurately predict individual behavior. Yet these predictions are highly reductive since AI systems depend on data that embed racist, sexist, and classist biases. As algorithms reproduce these biases, they create feedback loops in which the future emulates the data on which the systems are trained. The question surrounding AI is hence ethical and efficacious: can and should AI support future action given its limits and biases? This question echoes the reflections of nineteenth-century statisticians. The epistemological and ontological gap between the data they collected and the abstract principles they offered threatened the legitimacy of their field. Like AI scientists, early practitioners struggled to determine whether statistics could discover universal laws and future happenings or whether it merely quantified phenomena that had already occurred. Novels contributed to this debate by both utilizing and critiquing statistical thinking. Victorian novels thus grapple with the problematic working conditions that arise when data collection falls behind the phenomena it claims to predict and explain.
In "Common Sense, AI, and the Whiteness of Affect," Gilbert discusses discredited racial hierarchies underlying the discourse and practices of data collection related to emotional expression. Historians of affect have long discussed nineteenth-century evolutionary contexts of standardizing emotional expression (e.g., Ruth Ley). More recently, scholars interested in the social entailments and implications of AI (such as anthropologist Kate Crawford) have described the questionable working conditions under which data is collected and mobilized for affect recognition systems. However, less attention has been paid to prior conversations around affect and theology which was partly displaced and partly disguised by discussions of Darwin and the eventual attempt to standardize observation of affect (e.g., Galton, Lombroso, and Mosso). In fact, unexamined eighteenth-century "Common Sense" theories, elaborated by such thinkers as Lord Kames and Thomas Reid, about divine design and the emotional basis of social cohesion still undergird this thinking. Gilbert considers recent debates around AI and affect recognition in the light of this history to identify how theocentric and racial assumptions underlie the working conditions (both ethical and practical) of these enterprises.
The idea of modeling the world by collecting data to predict and normalize outcomes dates back to the biopolitical thinking of the nineteenth century. British researchers in physical anthropology and statistics elevated white (Anglo-Saxon) masculinity to a vanguard place in an ongoing history of progress which culminated in a pseudoscientific, eugenist, and imperial governmentality. In doing so, they innovated a “datafying” process built on the atomized homo economicus of political economy and utilitarianism. In our own day, this atomistic social ontology underwrites the working conditions in which “AI” is produced and implemented. Jathan Sadowski defines data as “a recorded digital abstraction of something or somebody” and datafication as “the process” of “reducing and turning things into data.” In “Ontologies of Datafication,” Goodlad explains how the biopolitical turn to race and statistics instituted “datafication” in concert with the deterministic frameworks of literary naturalism. Part of a project on the ontological work of fiction, the paper argues that a critical grasp of ontology, inside and outside literature, will make scholars better readers of Victorian writing, as well as informed respondents to new machine-driven regimes of datafication.
Our in-person panel leaves time for discussion and invites spillover into hallways.