Xenoreproduction: Exploration and Recovery of Collapsible Modes as Core AI Safety Objective

Abstract

Generative AI models reproduce the biases in the data and further amplify them through mode collapse. AI scholarship often overlooks conceptually rich perspectives, such as those from Queer Theory and Black Studies, when theorizing about those phenomena. As a result, our field lacks a theory with teeth, one sincerely committed to pluralism. In this paper, we introduce Xenoreproduction as a core AI Safety objective, aimed at avoiding homogenization failure modes. Succinctly, Xenoreproduction is the task of strategically recovering and exploring collapsible modes in Gen AI models. To illustrate it, we sketch how this task is formulated for LLMs. Our formalism ties queerness and subalternity to the collapsible modes. Our abstractions readily spark conversations about CoT dynamics and improvisation capability in agents. To render AI xenoreproductive is to ensure it adapts under uncertainty, invents new solutions, and behaves creatively. We invite future AI scholarship to form more unruly connections between disciplines.