The architecture of anomaly: agential errancy in human-AI relations

This paper examines errors arising amid human and AI interaction, theorizing them as “agential errors,” namely, as generative performances of the human-nonhuman co-developmental relationship and not as mere failures to be expunged. Drawing on posthuman rhetorics and new materialism, the authors argue that errors are crucial operations that actively reshape the norms, identities, and infrastructures through which the human–AI relationship unfolds. In particular, the paper identifies three types of agential errancy: the grammatical-stylistic, or errors strategically deployed by humans inside of AI texts to perform human authenticity; epistemic errors, or AI-hallucinated sources and unverifiable knowledge claims that expose the rhetorical nature of epistemic solidity; and social errors, or decorous practices of politeness in AI interactions, which Large Language Models tend to over-perform but do not require. In each case, the error operates not as a simplistic breakdown demanding repair but as a recursive, co-developmental event that reshapes human-AI systems, causing them to “reload,” resulting in new standards, habits, and possibilities. The paper then concludes that “the error” will shift over time alongside each “agential” realization; however, each instance exposes just how the human is not equivalent to the AI nor the AI to the human.