Nsisong Udosen
Large language models, such as ChatGPT and Claude, are trained on aggregations of text scraped from the internet. This process compresses the genealogy of ideas into outputs with a difficult-to-trace lineage. When these models become the primary interface through which people compose thought, the cumulative effect is a dissolution of what writer Morgan Harper Nichols calls the “personal bibliography”, known as one’s ability to trace the source of their formative beliefs or perspectives (Nichols, 2022). This paper defines this AI-enabled erosion of traceability as ‘epistemic flattening’. Beyond the impacts on individual voice, this paper analyses the effects of this flattening on shared knowledge infrastructures.
Systemic designers are uniquely positioned to demand and design for different choices, and in the words of Toni Cade Bambara, “make the revolution irresistible” (Bambara, as cited in Lewis, 2012, p. 3). The two-track structure proposed here, building edge alternatives while calling for stronger levels of ethical responsibility within dominant systems, is offered as an opening, in addition to the recognition that everyday practices of refusal and care are already constitutive of the change. The question of whose words these are departs from being a rhetorical one, and transforms into a design brief that calls on practitioners and the wider public to participate in building an alternative epistemic infrastructure.
Keywords: epistemic flattening, conversational AI, systemic design, epistemic lineage, large language models, data sovereignty