AI and Institutional Action

Recent debates about AI and speech acts have focused almost exclusively on whether Large Language Models and other AI systems can assert. Yet this focus obscures a more consequential feature of these systems: many are already embedded within institutions in ways that allow them to create and alter social facts. In this paper, I argue that some AI systems can perform declarations. I defend this claim through the Human–AI Declaration Symmetry Thesis, according to which an AI system performs a declaration when it can create or alter the same social facts as a human speaker through a process that tracks the relevant worldly features of the social domain. I then address three objections. First, I reject the claim that declarations require intentions or other robust mental states that current AI systems likely lack. Second, I argue that sanctionability requirements for declarations attach to institutions and their roles rather than to individual speakers. Finally, I reject attempts to weaken the view through concepts such as proxy-declaration, quasi-declaration, or proto-declaration. Recognizing AI declarations helps illuminate the authoritative role these systems increasingly play in altering and structuring the social world.

AI systems are increasingly embedded in institutional roles once occupied by humans. Yet the question of whether such systems can count as members of the groups into which they are integrated remains underexplored in social ontology. In this paper, I argue that current and near-future AI systems can be, though are not always, members of institutional groups. Existing theories of group membership face a dilemma: permissive accounts over-include machines and artifacts, while restrictive accounts that require consciousness or human-like agency under-include AI systems that occupy authoritative institutional roles. To resolve this dilemma, I develop a conferralist theory of institutional group membership. On this account, membership is not grounded in an intrinsic property of the role-occupant. Rather, an entity counts as a member when it occupies a role within a group’s structure and is conferred a social status connected to that role. I distinguish recognitional membership, conferred through formal institutional recognition, from agential membership, conferred when an entity’s outputs are treated as determining the group’s decisions, commitments, or plans. This view explains how some AI systems can be considered group members without implying that every tool or artifact is also a member. It also clarifies what is distinctive about AI replacement compared to historical instances of machine automation: AI systems are not merely replacing human labor but increasingly occupying agential roles through which institutional groups plan, coordinate, and act.

The integration of artificial intelligence into institutions creates a neglected problem for those who attribute moral properties to corporations and other group agents. If AI systems cannot themselves bear certain moral properties, their incorporation into groups may alter whether those groups can bear those same moral properties. I introduce the concept of a morally hybrid group: a group whose members differ in their capacity to bear a particular moral property. Focusing on moral blameworthiness, I examine the conditions under which such a group can be blameworthy. I argue that a morally hybrid group can be blameworthy only when members capable of bearing blame are suitably distributed across relevant roles and participate in its actions. I conclude by showing how the possibility of morally hybrid groups provides a distinctive metaphysical justification for maintaining humans “in the loop” and generates new – and potentially insurmountable – epistemic challenges for attributing moral properties to increasingly AI-integrated institutions.