Towards an Anthropology of Agentic AI Sandboxes: Emerging Agency in Experimental Governance.

Catalysed by widespread technological disruption, large-scale plans for national level AI regulatory sandboxes across the European Union (EU) are on the ascendant. Articles 57, 58 and 59 of the newly created EU AI Act are mobilising what represents “one of the most important technological developments across all member states.” Although their main objective is to supply the global community with safe and innovative AI systems, built upon an ethos of responsible development, economic growth, technological advancement and societal benefit programmes, these experimental governance frameworks are encountering a mixture of mistrust and ambivalence from the AI Governance community. For many stakeholders, however, interaction with sandbox initiatives may bring necessary innovation opportunities and regulatory clarity for their organisations and communities, with the prospect of a fruitful and technologically progressive future. Yet, in the process, tech companies, regulatory bodies, research institutions and civil society groups have entered into a process of speculation about the future of AI experimental governance which has fed into concerns regarding technological control and democratic participation.

Conversely, the process of ontological negotiation that different stakeholders bring to sites of experimental AI governance may represent an opportunity to address fundamental sociotechnical challenges including algorithmic accountability, democratic participation in technological decision-making, and the equitable distribution of both AI benefits and risks. While simultaneously revealing how current sandbox frameworks may inadvertently reproduce exclusions that limit their transformative potential for fostering more just and inclusive technological futures. Rather than pitching corporate interests against civil society, this presentation will examine the technical ontologies centered on computational/AI emergence, regulatory ontologies highlighting legal containment, commercial ontologies emphasising market value, and social ontologies acknowledging relational agency that may generate productive tensions that may both facilitate and limit opportunities for democratic technological governance.

Drawing on Anne-Marie Mol’s work on ontological politics, we can understand how these various ontological framings do not merely compete and collaborate within sandbox environments, but actively enact different realities of what experimental AI governance is. The sociotechnical assemblages that may emerge through these regulatory practices may reveal both the limitations of purely technical approaches to AI safety and the generative potential of hybrid governance arrangements. Following Mol’s insight that ontologies are performed rather than given, these sandboxes may become sites where multiple, potentially contradictory versions of Agentic AI agency, risk, and control are simultaneously enacted and coordinated, acknowledging multiple forms of human and AI agency. Critical decisions around exceptions and exit governance from sandboxes exemplify these relational agencies, for example when an algorithm misbehaves, will it be the legal ontology that serves as a site of intervention, algorithmic agency or both? By anticipating how these ontological politics could emerge in practice, it attempts to move beyond reductive framings that treat AI systems as either passive regulatory objects or autonomous technological agents but as situationally linked in the sociotechnical network with the potential for a more than human Agentic AI Governance.

By Christine Louise, MRes