Material and Linguistic Practices in the Shaping of AI-Based Collective Agency

Abstract

This article explains how material practices and naming shape collective agency in organizations that depend on artificial intelligence. It addresses an apparent tension. An AI system can be a constitutive condition of what an organization can do at a given time. The same system can remain replaceable across the organization’s history. This distinction separates present agency from persistence through change. This article develops affordance facts as an analytically distinct category within explanations of social facts. These facts concern the possibilities for action that a material or computational arrangement makes available to situated users. They differ from anchoring facts because they do not establish the rules that ground a social fact. Naming has a related but narrower role; it supports recognition and coordinates claims of continuity, yet it is neither necessary nor sufficient for organizational identity. A revised thought experiment on structural relocation shows that continuity cannot be determined by an AI system alone. Brief illustrations involving IBM Watson Health and Google DeepMind support this claim. The resulting view avoids technological determinism because it treats agency as relational. AI systems can shape organizational capacities without becoming autonomous identity bearers. They are often constitutive conditions of present activity and contingent operational infrastructures across time. This framework clarifies how responsibility can remain with an organization even when an AI system has become indispensable to its current operations.

1. Introduction

Questions concerning collective agency and collective intentionality occupy a central place in social ontology. Social ontologists are, after all, concerned with groups, corporations, states, and many such similar entities. We recognize that these entities act in various ways; however, we must ask whether their actions should solely be reduced to what is done or intended by their human constituents. Perhaps by some anthropocentric bias, the contributions of material objects to social phenomena have been hugely neglected in social ontology. When it comes to metaphysical explanations of collective agency and the nature of social facts and kinds, the literature relies heavily on seemingly unique human contributions, such as attitudes and intentionality, to the constitution of such phenomena. We can see this not only in ontologically individualistic explanations of collective agency, as in Bratman [1] and Shapiro [2], but also in constructivist accounts of social facts, such as Searle [3]. This reliance on human intentionality shapes how collectives are distinguished, which often happens through acts of naming that reinforce anthropocentric assumptions about what counts as an agent. The naming of collectives, therefore, does not merely describe a social phenomenon, but instantiates it. Recently, there have been some attempts to introduce the notion of material objects into the field, in the form of discussions revolving around ontological hybridity and heterogeneity for social phenomena. Elder-Vass [4], for example, provides an account of how sociotechnical entities, as a subclass of social entities, can acquire their causal powers through the spatiotemporal organization of their material parts. Similarly, Brouwer, Ferrario, and Porello [5] provide an account of collective intentionality and agency that acknowledges hybridity, suggesting that material objects can embody case-agential attitudes, thus providing a venue for them to join the base upon which collective agents supervene. We propose to label such approaches as approaches of hybridity. However, as we will see in the following section, this view cannot adequately explain how non-human factors play a role in the constitution of social facts.

Recent social ontology gives material objects a more direct role. Elder-Vass argues that sociotechnical entities can acquire causal powers through the organization of human and material parts [4]. Brouwer, Ferrario, and Porello allow nonhuman components to contribute to the attitudes of hybrid collective agents [5]. These accounts weaken the assumption that collective agency supervenes only on human attitudes. They also direct attention to the way in which an organization distributes memory and control across people and artifacts. A difficulty remains. The contribution of a computational system should not be explained by quietly assigning it the same attitudes as a human participant. Such an explanation only recognizes the artifact after translating its role into intentional language. It can therefore miss the distinctive contribution made by storage capacity or automated constraint. The relevant question is how those properties enter an account of collective action without turning the system into a person.

Actor–network theory offers an important response to this problem. Latour treats humans and nonhumans as participants in networks of association [6]. The principle of symmetry directs inquiry toward what each participant does within a network. It prevents technology from disappearing into a passive background, and also discourages an account in which technology determines social form by itself. The present article accepts these methodological lessons. It does not follow that every participant contributes in the same way. A person can undertake a commitment. A database can preserve a record that later participants must consult. A workflow can prevent an action that an unauthorized user attempts to perform. These contributions belong to one sociotechnical arrangement, but their explanatory roles remain different. Recent work on artificial intelligence sharpens the issue. List compares artificial agency with group agency and separates agency from responsibility [7]. Courtenage examines intelligent machines within collective structures of responsibility [8]. Holmström and Hällgren describe the mutual shaping of AI systems and organizational contexts [9]. Hedfeld develops a sociotechnical account in which AI mediates institutional action without thereby becoming a moral person [10]. This literature supports a relational analysis of AI in organizations. It also warns against moving too quickly from technical participation to moral status. The fact that a system changes how an organization acts does not settle the question of whether the system is an agent in its own right. It also does not settle the question of whether the organization would cease to exist after replacing that system.

The central question is whether an AI system can be constitutively important for an organization without being necessary for its identity. The answer depends on a distinction between two time scales. A synchronic claim concerns what enables agency at a given time, while a diachronic claim concerns what must remain for the organization to persist. An organization may be unable to exercise one of its current capacities without a particular model. It does not follow that the organization must disappear when that model is replaced. Confusion arises when dependence within a present procedure is treated as a condition of identity across time. The distinction also explains why calling a system constitutive need not amount to technological determinism.

This article makes four contributions. It gives a working criterion for an AI-based organization and explains why the ordinary use of an AI tool does not satisfy that criterion. It distinguishes affordance facts from grounding and anchoring facts while acknowledging that these explanations can depend on one another. It limits the ontological claim made for naming by separating public recognition from numerical identity. It also recasts the identity thought experiment as evidence that system continuity cannot settle organizational continuity. The argument begins with material affordances and the extended mind analogy. It then examines naming before turning to the separation between an AI system and an organization. The final sections connect collective agency with persistence and state the legal implications of the account.

2. Material Affordances and Collective Agency

2.1. Hybridity and Relational Contributions

The problem with the aforementioned hybridity approaches is that, due to the parallels they draw with extended cognition, the role they attribute to material or artificial objects does not do justice to the unique and intrinsic ways in which such objects contribute to social phenomena. Both approaches humanize the non-human by abducting an originally human understanding of intentionality and mental attitudes and applying it to material or computational systems. This undermines the thesis of social heterogeneity rather than supporting it and fails to do justice to the particular ways in which non-human elements contribute to social facts and entities. More importantly, it neglects how naming functions as a bridge between intentional and material domains, often fixing identity where material continuity is absent, or dissolving continuity where the material base persists but linguistic identity shifts. While we are sympathetic to hybridity approaches, to properly establish the thesis of heterogeneity, there is a need to go beyond and show the unique ways in which non-intentional spatiotemporal properties of material and artificial systems contribute to the constitution of social ontological phenomena such as collective agency and social entities. One way to achieve this is by drawing a parallel with the Gibsonian notion of affordances. If what can be done with an object, or an artificial system, is determined by the interplay between its properties and its users, we can see how such affordances become relevant to social ontological questions. Following Epstein’s [11] framework of grounding- and anchoring-type facts, we contend that we need a third class of facts to accommodate heterogeneity and to give a full explanation: a class we call affordance facts. These capture the specific ways in which the material or computational organization of objects affords possibilities of action that shape collective agency. For example, a digital system used for governance does not only instantiate existing rules, but also enables new forms of coordination through its computational affordances. However, the continuity of the institution that employs such systems is often preserved through naming practices that maintain the same linguistic identity despite the replacement of its material and technological components. We therefore suggest that social groups can be ontologically conditioned by material and linguistic practices.

Actor–network theory supplies a broader account of distributed action [6]. Its principle of symmetry directs attention to associations rather than a fixed division between active subjects and passive objects. This approach is especially useful when an outcome cannot be traced to one decision maker. The present argument adopts that methodological point. It does not infer that every participant contributes in the same manner or bears the same responsibility. Human members can undertake commitments and answer for their use of a procedure. Technical components can enable an action or prevent one. Institutional rules determine which outputs count as decisions of the organization. Retaining these distinctions prevents methodological symmetry from becoming an undifferentiated theory of agency. This relational view also avoids technological determinism. An affordance is not a force that produces one social outcome by itself; it is a possibility that arises from the relation between an artifact and a situated user. The same prediction model can inform a professional judgment in one organization, yet in another organization, its score can trigger an automatic refusal. The difference does not lie in the model alone, but rather in the procedure that assigns authority to the output and in the options available to the people who receive it. Technical properties matter because they open some courses of action and close others. Organizational uptake matters because it determines which of those possibilities become part of collective practice.

2.2. Affordance Facts

This proposal has two benefits. It draws proper attention to the fact that the constitution of social entities is, first and foremost, material, extended through technological infrastructures. These infrastructures shape what collective agency can be. Drawing attention to physical and artificial structures helps explain how properties that appear indispensable to social entities really come to be. On the other hand, naming functions as the linguistic correlate of material and computational affordances, in the sense that it binds otherwise transient systems into enduring collectives, allowing for both theoretical and practical persistence. Naming operates as an implicit condition of coordination and recognition, allowing both humans and artificial systems to act under the description of a collective whose boundaries are otherwise materially and computationally diffuse. Even if this question might not seem ontologically important to some, it has clear ethical, legal, and political implications. For example, if we conceptualize corporations as including AI systems as constitutive members, actions performed by those systems could, under certain naming practices, be attributable to the collective itself. Conversely, if, as Lawford-Smith (2019) [12] argues, the inclusion of such heterogeneous components undermines the possibility of coherent collective agency, they must be treated as extensions rather than members. However, the boundary between inclusion and exclusion is often sustained by naming itself, as the name of a collective determines whether an AI system is seen as internal or external, essential or replaceable.

An affordance fact concerns an available possibility for action within a situated sociotechnical arrangement. The term follows the relational insight in Gibson’s theory of affordances [13]. An affordance belongs neither to an artifact considered in isolation nor to a user considered apart from an environment. It arises from their relation. An affordance fact does not create a constitutive rule. It concerns what participants can do once the relevant arrangement is in place. A secure archive affords durable retrieval for authorized users. A voting platform affords rapid aggregation when its interface and access rules are accepted. A prediction system affords ranked comparison among cases, although it may not afford a reasoned explanation of that ranking. Affordance facts are therefore analytically distinct from anchoring facts. The distinction does not require a new fundamental kind of fact, nor does it imply that an affordance can be described without reference to an institution. Some affordances depend on an authorized role, while others depend mainly on physical design. The point is explanatory. Anchoring identifies why a rule has authority. Affordance identifies how an arrangement enables or limits action under that rule. One fact can participate in both explanations when an institutional decision changes access to a technical capacity. Even then, the questions remain different. One asks why the arrangement is authorized, and the other asks what action becomes possible through it. Consider a public agency that uses an AI system to rank applications. The procurement decision can anchor the system’s authorized role. The applicable law can ground the legal status of a decision that follows the required procedure. The system’s capacity to sort thousands of cases is an affordance available to the agency. A threshold built into the interface may also make some applications easier to exclude from further review. The ranking does not become authoritative merely because the system can produce it. Authority depends on the institutional arrangement in which the output is received. These explanations concern different relations, so treating them separately prevents the system from appearing inert. It also prevents technical capacity from being mistaken for institutional authority.

2.3. Artifacts and Collective Capacities

All of the above invites an examination of how the inclusion of material objects into social ontology transforms our understanding of collective agency. To achieve this, let us start by drawing a parallel between the discussion on extended cognition and social ontology. The idea of extended cognition originates from Clark and Chalmers [14], although the origins of that very idea can be traced back to semantic externalism and functionalism in philosophy of mind. The classical example they provide is of an Alzheimer’s patient named Otto, who relies on his notebook to make it to an exhibition in a museum. While his biological memory is impaired, Clark and Chalmers observe that the notebook fills in the same functional role as Otto’s biological memory. If we think that Otto’s biological memory is part of his cognition, we have all the reason to think the same for the notebook. The lesson here is that the limits of cognition cannot be drawn around what is inside the skull. However, there is another important lesson to draw, and it is about the nature of agency. In Otto’s case, we can see that without the help of the notebook, he cannot accomplish the action of going to the museum, but this is not a simple case of inability due to some external circumstances. We can very much imagine that without the help of the notebook, he cannot even begin to form the intention to go to the museum. If we think that memory is crucial for agency and the forming of appropriate intentions to exercise that agency, then it should also be clear that the notebook is essential for Otto’s agency as it fills in the functional role of his memory.

Now, let us imagine Otto in some collective agential explanations. One such prominent and individualistic theory comes from Bratman [1]. According to Bratman, there is nothing above and beyond individual intentions in collective agency. Following from this, to form a so-called collective intention, it is enough to form an intention in the form of “We intend to go to the museum”. This intention should be common between Otto and the people with whom he intends to go to the museum. Furthermore, each participant should also know that each other has such an intention. Without going further into Bratman’s discussion of how the participants’ subplans should all mesh together, we can ask about the role Otto’s notebook plays in this scenario. It is clear that without the notebook, Otto can neither form the appropriate intentions (“we intend to go to the museum”) nor the common knowledge of others’ intentions. Without the notebook, Otto cannot go to the museum for reasons pertaining not the external circumstances that are out of his control, but due to the very conditions of his agency. As such, the notebook is, in this sense, essential for Otto’s agency and the overall collective agency of which he is potentially a part.

What about more properly collectivist approaches to collective agency? Lawford-Smith, drawing from List and Pettit [15], asserts that a strong criterion for collective agency involves autonomy and rationality on the part of the organization at hand. Autonomy refers to the fact that an organization’s intentions and beliefs should be independent from the intentions and beliefs of its members. The rationality criterion, on the other hand, involves consistency between that organization’s beliefs and intentions and its previous beliefs and intentions, so that the organization does not haphazardly change its stances and plans on the fly. As we cannot really talk of an individual Otto here, we have to ask about the ‘notebook’ of that very organization, be it a state, corporation, or a union. The most obvious non-human factors are the very constitution of that state, some records of the past actions of the corporation, its databanks that provide the necessary evidence for forming rational beliefs and intentions, etc.

The question is not whether what is recorded on such documents and files could be distributed among the human constituents of the organization—we already know that—but rather the question is about the very possibility of them being distributed and offloaded onto material objects. Just like in the individual case, this too seems correct. After all, the very function of a state’s constitution is to grant the basis for its autonomy and rationality. When the leader of the state, or any other legal or political human constituent, steps out of line, we can refer to the constitution and what is written there and judge the constituents accordingly. This way, the autonomy of the state from its individual constituents is guaranteed.

Now, one criticism of the overall approach we showcase here would be to say that notebooks, constitutions, databanks, etc., are all too human, so to speak. They are not showing how the non-human can have an essential contribution to the collective agency, but rather they ‘humanize’ the non-human. The notebook serves as a collection of intentions and beliefs, and so does the constitution for a state. Here, we are not stepping out of the human, and not showing what can be uniquely contributed by the non-human, but only the ways in which they can be humanized. So, in a sense, all of the intentionality that can be attributed to seemingly material records or similar artifacts are derived from the original intentionality of human participants of the collective at hand. As long as the focus is still on the beliefs, intentions, agreements, acceptances, and practices, the examples we have provided cannot really be employed to resist the anthropocentric tendencies of social ontology.

3. AI-Based Organizations

3.1. Working Definition

Professional organizations that integrate artificial intelligence systems seem to be more complex than being just groups of human members, technical staff, and management, and things of that kind appear to exhaust the notion of an AI-based collective (although such a collective can be seen as a non-summative abstract whole consisting exclusively of such things). Therefore, it appears that it is not problematic to assume that the physical and computational extensions of abstract objects such as AI-augmented organizations are flexible, in the sense that each of them can change, while the organization still retains its identity. For example, organizations can change their employees, management, and even their core AI systems, as well as their data infrastructures, platforms, and even names, while still holding on to the same identity. Of course, we are not forgetting the legal aspect of their identity, which brings us to the issue of institutional licenses and AI operational systems, which, unlike all of the concrete things above, and similar to the organizations themselves, can also be considered abstract entities. But are they the same thing as the organizations? With this question in mind, it seems that the issue of the relationship between AI-based collectives and their AI operational systems deserves attention. Our main focus in the remainder of the paper will therefore be to show that AI-based collectives cannot be equated with AI operational systems, even if those AI-based collectives are defined and sustained by those particular AI operational systems. We will start with the naming problem, which is crucial in defining collectives in such a manner.

This definition is functional and institutional. It does not require the name of the organization to contain the term AI. It also does not require the system to possess intentions. A conventional firm can become AI-based when a system enters its regular decision structure. For example, a firm may authorize a model to rank applications before human review. The model then shapes the set of cases that decision makers encounter. By contrast, an organization that uses a chatbot for occasional drafting remains an ordinary organization that uses AI. The difference concerns the place of the system within an authorized procedure. It does not concern how prominently the organization advertises the technology.

The criterion is deliberately demanding. Dependence on electricity does not make every organization an electricity-based collective. The same applies to generic software. These technologies enable activity without occupying a specific role in the organization’s decision structure. An AI system becomes relevant to the present argument when its classifications or recommendations are built into that structure. Even then, the organization should not be identified with the system. Institutional licenses and legal personality can persist when technical components change. The definition therefore marks a strong form of operational integration while leaving the question of identity open.

3.2. Four Roles of AI Systems

A constitutive condition is part of what enables the present capacity of the organization. An AI system has this role when an organization cannot perform the same authorized activity in the same way without it. The claim is indexed to a time and a practice. A prediction model can be constitutive of the present capacity to rank a large volume of cases. The claim does not imply that the organization could never develop another way of performing that task. An operational infrastructure is the organized technical basis through which an activity is performed. The term draws attention to stable use rather than occasional assistance. A system has this role when members rely on it as part of an established workflow. This description concerns function within the organization. It does not decide whether the system is a member of the collective or whether the system is essential to identity.

A contingent extension is a component that can be replaced while the organization persists. The term does not suggest that the component is unimportant. Replacement may be costly and may require extensive retraining. It can also transform the organization’s capacities or public reputation. A component can therefore be practically indispensable during one period while remaining contingent across a longer history. Difficulty of replacement is evidence of dependence, but is not sufficient evidence that the component bears identity. An identity bearer is a feature whose continuity is necessary for the same organization to remain in existence. The relevant feature may involve legal incorporation or an authorized rule of succession. It can also involve continuity of obligations recognized by other institutions. A particular AI model does not usually satisfy this description. The model can be transferred to another entity while the first organization survives. The organization can also adopt a new model without becoming a numerically different organization. The central thesis can now be stated precisely. An AI system may be a constitutive condition of present collective agency. It may also function as operational infrastructure. The system can still remain a contingent extension across time. These descriptions answer different questions and can therefore apply together. The first two describe the system’s role in a current practice. The third describes its replaceability across organizational change. None of these claims treats the system as the bearer of organizational identity. That further claim would require an argument about what must remain through succession.

4. Naming and Organizational Identity

4.1. Naming as Recognition

Naming itself carries another kind of agency—one that has often been overlooked because of its apparent simplicity. The act of naming seems to be only linguistic, but it has a structural effect on how collectives persist and transform. When a collective gives itself a name, or receives one from others, this linguistic mark becomes the point where its material and intentional aspects converge. The name allows continuity where the actual structure changes, and it allows replacement of parts without dissolution of the whole. However, this continuity is fragile. It depends on whether the name binds the identity of the group to its contingent members or to something that can outlast them. The next section turns to this issue in more detail.

Naming is not itself a form of agency. It is a linguistic practice that helps organize claims about an agent. A name can serve as an anchor when an authorized act establishes its institutional use. It can also provide evidence about continuity when later participants deliberately preserve it. These effects explain why names often matter in disputes about succession. They do not make a name necessary or sufficient for identity in every case; an organization may survive a lawful change of name. A newly created entity may receive an earlier name without acquiring all of the earlier entity’s obligations. Kripkean accounts explain how names can continue to refer despite changes in description [16]. Organizational cases add a complication because institutional rules govern authorized succession. The continued use of a name may reflect the persistence of one entity, or it may instead reflect a transfer of branding or a legal settlement. Public usage can then diverge from the history of obligations. The referent of an organizational name therefore depends on more than linguistic form. It depends on the practice that connects the name with an institution and on the authority of those who maintain that practice.

4.2. Naming and Apparent Necessity

When it comes to naming social groups, the modal profile of persistence is deeply affected by whether the name ties the group’s identity to one of its extensions. A company called Bill Gates and his workers would make Gates appear inalienable to the group’s very being, while a name like Microsoft allows the company to endure his departure without strain. Yet there are many cases where collectives have in fact been named after members: Johnny and the Moondogs being an early case of the Beatles. The puzzle is immediate. Had Lennon left, would the group’s name have demanded revision, or could the group have persisted under a description that no longer matched its actual structure? Now, imagine a collective that integrates an artificial agent so deeply that the formal social group adopts its name. For example, lets imagine a fictional system called SearleAI, and the corporation is named after it as SearleAI & Partners. At first, the AI is one contingent extension among others, filling a role analogous to an associate or senior partner. It appears that by embedding the AI into the very name of the collective, the group has altered its modal profile. The AI now appears as an essential component of the group’s persistence. If the system (SearleAI) were retired, replaced, or rendered obsolete, the question arises whether the group could continue under the same name without generating a kind of structural dissonance. The practice of naming here tacitly shifts a contingent extension into something that behaves like a non-contingent core.

This tension mirrors what happens with human members but with a distinctive twist. Human figures like Lennon are paradigmatic persons, whose leaving creates a mismatch between group name and group structure but without threatening the metaphysical persistence of the collective itself. With AI, however, the dynamic is sharper because the artificial agent is already ontologically ambiguous. Is it a member, a tool, or an infrastructural substrate? Naming a group after it forces the structure to treat the AI as if it were a non-substitutable node, even when in practice it may be replaceable by another system with the same affordances. What naming does here is not merely linguistic—it has ontological weight. It reclassifies a contingent extension into a seemingly essential one, making what should be a standard replacement of one system by another appear instead as a disruption of the group’s identity.

The metaphysics of group naming, once refracted through artificial intelligence, highlights how our linguistic and institutional practices can reconfigure the structural roles of non-human agents. By embedding AI into the very name of a collective, we create the illusion that what is, in principle, replaceable becomes indispensable. This results in a form of modal distortion, in the sense that a system that ought to remain a contingent extension of the group is treated as its structural essence. This distortion arises because linguistic and institutional practices can easily reify functional occupants, obscuring the difference between what is necessary for the realization of a group’s structure and what is only contingently filling a role.

Seen in this way, naming emerges as a dual operator of social ontology that simultaneously stabilizes and distorts. It stabilizes collectives by granting them linguistic persistence across material and artificial change, but it also distorts ontological hierarchies by elevating certain contingent elements such as artificial systems into the position of apparent necessity. The theoretical novum of this lies in showing how collective agency is mediated not only by material affordances, but also by the semiotic affordances of language itself. The goal is to understand how institutions and AI-driven entities can manipulate naming to manage perceptions of autonomy and continuity. Naming, therefore, could be seen as both a metaphysical and institutional ‘tool of persistence’. More specifically, the naming problem brings into focus the question of what constitutes the identity of a social group, a question that, as we will see, becomes even more intricate in the context of analyzing the problem of structural relocation and renaming of AI-based organizations that are defined and sustained by a particular AI operational system.

4.3. Real World Illustrations

The transition from IBM Watson Health to Merative illustrates a separation among assets and organizational identity. IBM announced the sale of healthcare data and analytics assets to Francisco Partners in 2022 [17]. The acquired business later operated as the standalone company Merative. The products and datasets moved into a new institutional setting. The Watson name did not determine the identity of that setting. The creation of Google DeepMind provides a different illustration. Google combined the DeepMind organization with the Brain team from Google Research [18]. The new unit joined people and computational resources under a revised name. No single model supplied its continuity. The organizational settlement determined which activities and histories were carried into the new unit.

These examples do not resolve a metaphysical theory. Corporate transactions can involve legal details that a short illustration cannot capture. They still show why real organizations should not be identified with a model or brand. Assets can move across organizational boundaries. Names can also persist after technical change.

5. From Collective Agency to Persistence

5.1. Synchronic Agency and Diachronic Identity

Collective intentionality and organizational identity concern different questions. The first asks how a group forms an attitude or performs an action at a time. The second asks which later entity inherits the group’s commitments. A complete account must connect them without treating them as identical.

Even when we find such an individualistic approach to collective agency to be highly constraining, we can turn to Shapiro’s [2] account. According to him, it is simply enough to have a commitment to the plan at hand without really having the intention to reach the end goal. This way, Shapiro’s account can accommodate cases of alienated participation into collective agency, such as an office worker who is committed to do his file work whose contributions are integral for the intentions of the higher-ups but nevertheless does not share those intentions. In such a scenario, we can imagine that Otto’s commitment to see through his part depends entirely on what is written in the notebook. So, it seems that artifacts such as notebooks can play an integral role not only in the very success of the collective intention (as explained individually), but also in the very possibility of one’s participation in the collective agency.

An AI system can affect both questions. It can help generate a decision at one time. It can also preserve data used in later decisions. Neither role proves that the same system must remain for the organization to persist. The distinction between capacity and identity remains decisive.

5.2. Structural Relocation

To better understand the relationship between a hypothetical AI-based organization, defined and sustained by a particular AI operational system, and that same AI operational system itself, we will present a revealing case of structural relocation and renaming. Imagine the case where administrators of an AI-based organization SearleAI & Partners, due to some particular reason, decide to migrate the organizational structure (i.e., the de facto organization) to another operational domain. Although almost the entire computational and managerial infrastructure of the collective was transformed, due to the persistence of the organization’s public identity, a legal settlement was signed to keep the name and historical continuity of the original operational system SearleAI (although it was ‘deactivated’ until a new organization was formed). Eventually, because the original AI operational system was no longer being used, and it was a very expensive and powerful system, a new collective was formed in order to maintain it. This collective was treated as a new organization. When analyzing this case, we might assume that the default I-relation likely tracks the AI operational system, in the sense that organization A persists to become organization B if and only if A and B are based on the same operational system (SearleAI). But in this case, the persistence of the organization comes apart from the system (the AI operational system ‘changes’, but the organization remains). At this point, we may simply conclude that ‘AI-based organizations are not their systems’. Indeed, this statement might seem trivial at first glance. However, the very triviality of this observation indicates that the case in question does not constitute a genuine obstacle to the claim that AI-based organizations and the AI operational systems affiliated with them can still be regarded as the same entity, since no real paradox would have emerged even if the I-relation had ‘tracked’ with the system. Therefore, to highlight the problematic nature of the claim about the inseparability of AI-based organizations and their systems, it seems necessary to turn to another case. For that reason, we will now point out the paradoxical nature of such a claim. This is a case involving two successor AI collectives, involving the problem of tracking the I-relation associated with these collectives and their respective AI operational systems. Imagine that the first organization was founded as an AI consortium at one point in time and was later restructured and renamed under a new legal entity. A decade later, the newly renamed collective allowed a new organization, formed in partnership with another institution, to take over its original name and even appropriate the historical record of the first consortium. So, unlike the previous case, it appears that we have two different collectives that could have continuity with the original consortium. If we muse upon this case, we can distinguish three distinct positions regarding the identity continuity of today’s organizations in relation to the original one.

Neither of the current organizations are the original consortium since the consortium has ceased to exist.

The new legal entity is the original consortium.

The reestablished organization under the original name is the original consortium.

System continuity favors the third description because S1 is now used by the reestablished group. Legal continuity favors the second description because the successor retains obligations. Name continuity also favors the third description. The conflict is the point of the case. No single factor determines identity.

5.3. I-Relation and Nonbranching

It appears that (1) tells us that this case constitutes a type of fission or branching. Lewis [19] uses the I-relation to describe continuity among stages. For the purpose of showing that (1) does not hold, we will introduce our own constraint, which borrows its basic principles from the already-established psychological nonbranching clause introduced by Derek Parfit [20]. This constraint will also be considered as a necessary condition that an AI-collective must meet in order for its identity to persist. However, to adequately elaborate our viewpoint, we first need to understand what branching (in terms of identity) generally means. Branching, according to the so-called psychological approach, is the case in which “one person ceases to exist and two new people come into existence” [21]. So, the nonbranching psychological clause specifies that a person must not branch in order to retain her identity, or alternatively, “that the fact about personal identity consists in the holding of a psychological relation that does not take a branching form” [22]. But when we talk about the analysis of AI organizations, as well as any other non-personal entities, psychological clause has no relevance because there is no psychological continuity (nor survival in the literal sense) in the first place. With this in mind, the conclusion that emerges is that I-relation, as well as the transitivity implied by it, is important when analyzing the continuity of at least those entities that are not considered persons. Following this, we could articulate our nonbranching condition for the persistence of metaphysical objects, which serves as a constraint. So, the criteria for an AI organization to retain its identity would, expectedly, be that it does not branch into two new organizations that come into existence after the branching. With all of the above in mind, if we consider the continuity of these two organizations, we will see that there is no point in time in which they truly coincided in any way; therefore, we can conclude that no branching had occurred. This, in turn, implies that (1) does not hold.

The nonbranching condition has a similarly limited role. Discussions of personal identity use nonbranching to prevent two later persons from being identical with one earlier person [20,21,22]. Organizations do not possess psychological continuity in the relevant sense. The present argument therefore does not treat Parfit’s account as a ready theory of organizational identity. Nonbranching can still operate as a negative consistency constraint because numerical identity remains transitive. It rules out treating two simultaneous successors as numerically identical with one predecessor. The condition tells us which combination of claims is impossible. It does not tell us which institution has the stronger claim to succession. Sports franchise cases illustrate the value and the limit of this constraint. The Cleveland Browns case separates team continuity from the movement of a franchise [23], and the Charlotte Hornets case involves later reassignment of a name and historical record [24]. These cases show that legal rights and public history can diverge. They do not prove that organizations share the identity conditions of sports teams.

As for (2), the situation is somewhat different. Even for those who find it obvious that (2) cannot be the case, there is still one large factor to consider, and that is the fact that the successor organization actually did claim the original consortium’s history prior to the renaming. The successor does not claim that history now, but the fact that it once did is quite enough for the problem to be raised. It seems that there is a clear I-relation between the pre-renaming consortium and the immediate successor entity. But the fact that the entire public, as well as both the new and old organizations, act as if they accept that today’s consortium is I-related to the original collective raises a problem for (2). To understand the depth of what we are discussing, let us consider the following. Conventionally and legally, the head of the reestablished organization has been considered its director since t1. Following the implications of the acceptance of position (2), while taking into account the overwhelmingly accepted fact that this person has led the reestablished organization since t1, it could be claimed that they were the director of the successor entity back in t1. But it appears that something is not quite right here. In other words, even though the director was the head of the same AI operational system the whole time, that system, in reality, has nothing to do with the successor entity and has previously been affiliated with a completely different collective. On the other hand, the system affiliated with the successor was also affiliated with the original consortium; that is, the system and the consortium were affiliated with each other up until their legal separation, when that bond was denounced and their system stopped being affiliated with the organization. It was not until then that the reestablished organization ‘came along’ and became affiliated with the original name. With this in mind, we seem to have solid ground to reject (2), unless someone thinks that the members of the successor organization actually belong to the reestablished one or, to formulate it even more peculiarly, that the reestablished organization is not itself (the successor is the original, and the reestablished is some other). We obviously cannot accept this farfetched notion. So, to summarize the analysis of the first two positions, it appears that by accepting (1), we have the unnecessary denial of identity continuity between the original consortium and today’s reestablished organization, for which there is an unproblematic consensus. By accepting (2), we would also go against the consensus, as well as against the will of the successor entity itself.

When it comes to the only remaining option, position (3), the problem, as might be expected, seems to be that today’s successor organization was once considered to be the original consortium, which, it appears, was undoubtedly I-related to the original. This is where the thesis of the separation of systems and organizations really comes into play. There is no doubt that what is considered today’s successor system was, strictly speaking, once associated with the original consortium, in the sense that the system was affiliated with it at other times (the consortium was once affiliated with the same system that the successor is affiliated with today), but the fact of the matter is that it is not affiliated with it any longer. How is such an ‘affiliation break’ possible? If we remember the earlier case, we realize that the persistence of that organization was not related to the system it was affiliated with prior to the migration. In other words, I-relation does not necessarily follow the system (although that does appear to be a common case). Therefore, there has been an instance of a ‘shift to a different I-relation’ in that earlier case, but, more importantly, this also occurred when the consortium severed ties with one system and established an affiliation with another. So, if every constituent of an AI collective is flexible, as we assumed at the beginning, and this also applies to separate abstract entities such as systems, it should be possible for shifting to occur in this case as well. One additional thing in our favor is that there is no reason for us to think that the shifting is not in accordance with the nonbranching constraint. In fact, the shifting phenomenon is only possible if the two organizations (or any other objects) never coincided, otherwise that would, following the nonbranching constraint, be considered a case of I-relation violation. The shifting requires that there is no instance of branching, as there indeed is not in (3). For this reason, we believe that (3) is also in line with the nonbranching constraint. So, taking all of this into account, we can conclude that the identity shifting of the organization from one system to another explains the motivation behind accepting (3) quite satisfactorily. Since (3) is the only plausible option, and this option is, as we have seen, only viable if AI-based organizations and their AI operational systems are treated as separate entities, we have offered an argument in favor of the thesis that calls for the separation of AI-based organizations and their AI operational systems. This brings us to the conclusion that if we do not assume the separation of AI-based organizations and their systems, collectives such as those discussed above could not be considered the same entities as they were before their respective transformations, so we can draw the inference that such a separation is philosophically, and possibly legally, necessary. Systems could, therefore, be treated ‘only’ as operation-like entities that provide infrastructure to AI-based organizations (and usually treat them as instruments of operational optimization), and as such, they do not represent an essential part of the organization; rather, they can be seen only as its contingent extension.

This result supports the separation thesis. An organization can move from S1 to S2 while preserving an institutionally recognized I relation [24]. A system can also move to a different organization. The relation between an organization and an AI system is therefore an affiliation that may change. It is not identity.

6. Legal and Ethical Implications

The separation thesis has consequences for responsibility. An organization should not avoid accountability by describing an AI output as the act of an external tool when the organization has assigned that output an authorized role. A lender that makes a model decisive in its approval procedure remains responsible for the procedure it adopted. This attribution does not require the model to be a moral person. It follows from the organization’s authority over the process and from its reliance on the resulting decisions. The system’s material contribution can therefore be recognized without allowing the organization to place responsibility outside itself.

The opposite mistake is also possible. Treating the AI system as the organization’s identity bearer can displace responsibility onto the system. List shows that artificial agency and fitness for responsibility require separate analyses [7]. Courtenage likewise argues that responsibility within human and machine collectives must be allocated through their organization [8]. Naming can intensify both mistakes. A name such as SearleAI & Partners may encourage clients to treat the system as a member. It may also encourage managers to describe its outputs as independent decisions. Clear governance should identify the human and institutional authority behind the system’s role. Technical dependence can still become severe. An organization may be unable to replace a model without losing data or expertise. This practical dependence can constrain institutional choice and does not make the model metaphysically necessary. Law and governance should address the dependency directly. The analysis therefore supports a restrained legal conclusion. AI systems can be part of the facts that explain organizational action yet need not receive the legal identity of the organization. Responsibility can remain with the collective that authorizes the system and benefits from its use.

7. Conclusions

This research has shown that the existence and persistence of collectives depend on both material and linguistic practices. Material infrastructures shape the operational possibilities of these collectives, while linguistic practices, such as naming, determine how they are recognized and transformed. The case of structural relocation and renaming demonstrates that the relationship between an AI-based collective, defined and sustained by a particular AI operational system and its operational system itself, is not fixed but flexible and dependent on how these practices are performed. When administrators decided to move the organizational structure to another operational domain, the organization retained its name and public identity, even though its operational system had changed. This shows that collective identity is not simply grounded in physical or institutional continuity, but also in the linguistic and symbolic structures. Artificial intelligence makes this dynamic visible, since AI systems can serve as constitutive elements of collectives. By embedding AI into their material and linguistic structures, organizations reshape their own conditions of existence. In this way, naming and infrastructural arrangements do not merely describe what an AI-based collective is, but are essential part of its ontological structure. Understanding AI-based organizations, therefore, requires attending to how material configurations and linguistic conventions together generate and transform collective agency and identity. An AI system may constitute a present capacity of an organization and can still remain a contingent extension across time. Operational infrastructure describes how the system functions. An identity bearer describes what must remain for the organization to persist. The argument does not assign both roles to the system.