Agents Without Agency: Anthropological and Sociological Lessons for Contemporary AI Research and Policy

Abstract

Recent hype about artificial neural network-based agents, fomented by tech corporations and AI startups and increasingly parroted by public sector actors and even critics, is premised on the idea that large deep learning models can act more and more autonomously — with minimal or no human supervision — and/or with increasing agency. In this paper, we contest the assumption — embedded in decades of research on the topic of agency in various AI subfields — that agency should be understood purely as a property of an individual. To challenge these claims, we will combine a historical analysis of past AI research with evidence from social-scientific scholarship, specifically from empirically-grounded sociological theory and linguistic anthropology, to make a case for a conception of agency that is ontologically relational and inextricable from social accountability. Agency is thus revealed as not a locatable and/or quantifiable property of an individual subject, but as a determination emerging from social interactions, whose status can dynamically vary with the observer and/or the accompanying sociotechnical (and/or legal) context. This perspective can a) illustrate the potential incentive to convince the public of the agentic status of AI models, which can defer accountability and blame away from corporations and/or developers and onto instances of opaque neural network architectures; and b) provide empirical grounding for the regulatory ``social licensing'' of purported agentic models.