Human–AI Joint Action

The Philosophical Problem of Acting Together

Human–AI joint action arises when a person and an artificial system participate in producing an activity whose course depends on their coordinated contributions. An autonomous vehicle that negotiates a handover with a human driver, a researcher who revises an argument through sustained exchange with an artificial collaborator, and a worker who jointly schedules and executes a complex operation with an AI agent provide different kinds of cases. Their common feature is the interdependence of activity. Their philosophical differences concern the organization and sources of that interdependence.

The central question is what makes an action joint. Two things can contribute causally to one outcome without acting together: wind and a human can move a boat. Two intelligent systems can independently aim at compatible outcomes, yet their activities may converge by coincidence. A human may employ sophisticated software instrumentally, retaining the entire practical plan while the system processes an assigned subtask. In another case, the artificial system detects an unforeseen obstacle, proposes an alternative, receives a revision from the human, changes its own subplan, and coordinates a revised sequence of actions. A defensible theory must identify the changing relational structure across these cases.

The philosophical stakes extend beyond whether a machine looks like a teammate. Joint action is one site at which practical authority, mutual expectations, obligations, responsibility, and attribution acquire social form. A human and an AI system may reliably complete a project, while the human alone carries legal liability. A system may generate new options yet lack authority to choose among them. A person may depend on artificial forecasting without establishing a shared intention. A joint product may possess distributed causal provenance while retaining an asymmetrical allocation of permissions and duties. These facts demand an account that disaggregates dimensions which ordinary language often combines under collaboration.

The central thesis of this monograph is that joint action should be investigated at multiple explanatory levels: coordinated performance, interlocking practical organization, publicly traceable participation, and reciprocal normative commitment. Evidence at one level establishes a particular kind of togetherness and creates a question about stronger levels. This thesis preserves the force of demanding theories of shared intention while allowing empirical research to identify real coordination phenomena in human–machine systems. The inquiry develops the differences and then specifies when stronger attribution is warranted.

Definition, Scope, and Criteria of Attribution

For purposes of the investigation, human–AI joint activity is an organized episode in which at least one human and at least one artificial system make causally and practically interdependent contributions to an evolving task or goal, with coordination of those contributions over time. Human–AI joint action, in its richer sense, is a temporally located exercise of this organization in which participants' contributions are mutually responsive to a practical conception of what they are doing together. The terms are operational starting points, to be evaluated against competing philosophical accounts.

Joint activity can extend across many individual acts. A research collaboration may include proposing a hypothesis, gathering evidence, drafting arguments, rejecting claims, and producing a finished text. Joint action may designate the whole coordinated episode or a constituent act such as jointly selecting a method. Individuation matters: a human pressing a button, an AI system recommending a route, and the pair navigating a changing environment might be descriptions of different actions or phases of one larger activity. Calling all of them joint overlooks the level at which organization occurs.

A goal is a state of affairs represented or operationally specified as to be achieved or maintained. A plan is a structured arrangement of possible actions, dependencies, constraints, and revisions. An intention in major philosophical theories is an action-guiding commitment with characteristic roles in reasoning and conduct; it is not simply a numerical optimization target. Shared intention names a further social configuration of intending and planning. Joint commitment in the plural-subject tradition denotes a normative relation constituted among parties whose standing to demand conformity is part of what makes their activity joint. Coordination denotes adjustment of actions according to their relations to other actions. Each term identifies a distinct explanatory function.

The category artificial system covers scripted dialogue programs, prediction models, embodied robots, tool-using agents, multi-agent software environments, and persistent artificial configurations. Their architectures differ in memory, action selection, permissions, planning, adaptation, feedback, and the persistence of identity. The word agent in engineering frequently describes a system designed to select actions in pursuit of goals. Its philosophical application must specify whether the relevant capacity is operational control, practical reasoning, public agency, participation in commitment, or full moral responsibility.

The attribution of action may also occur at multiple scales. In an institutional AI system, a human user, human developer, employer, software component, deployed agent, and organization can each contribute to a decision without performing the same action in the same capacity. To locate joint action, it is necessary to identify the participants at the explanatory scale in question, the available alternatives, the information they used, the possibilities of revision, and the consequences for which an attribution is made.

The leading research questions are accordingly: Can a shared goal exist without shared intention? Can reciprocal practical responsiveness occur without phenomenal experience? Which norms arise through participation, and which originate in human institutions? Does an artificial participant have reason to remain committed when task specifications change? How does the practical organization of joint action persist across sessions and versions? What happens when apparent partnership conceals a unilateral relation of control? These questions guide the historical reconstruction and conceptual analysis.

Historical Coordinates: From Practical Action to Collective Agency

Ancient philosophy investigated voluntary action, deliberation, and the practical conditions of acting. Aristotle's Nicomachean Ethics, especially Book III, treats voluntary and involuntary action and deliberative choice as central to ethical evaluation. These texts provide conceptual resources for thinking about the relation of agency to practical life, although they do not formulate the modern technical question of cooperation with AI. Modern social ontology develops additional questions concerning the agents and intentions involved in collective undertakings. Historical continuity here consists in inherited problems of action, organization, and responsibility, not in projecting contemporary machine agency into older vocabularies.

The twentieth-century philosophy of action sharpened the relation between what someone does and why they do it. Elizabeth Anscombe's Intention (1957) placed the question of action under a description and the explanation of intentional conduct at the center of philosophical analysis. Donald Davidson's “Actions, Reasons, and Causes” (1963) defended an influential causal account in which primary reasons can explain an action by causing it in the right way. Questions about deviant causation, rationalization, and action individuation then shaped subsequent debates. Michael Bratman's Intention, Plans, and Practical Reason (1987) developed a planning account of how intentions organize temporally extended activity. These developments matter because any account of shared action must explain how individual contributions acquire an intentional or otherwise action-guided structure.

There is also a distinctive history of computing and its encounter with social interaction. Joseph Weizenbaum's original 1966 description of ELIZA presented a natural-language program based on keyword-triggered decomposition and reassembly rules. ELIZA is useful as a historical boundary case: coherent exchanges can be produced through specified transformations without independently establishing that a system shares the user's intention. The case exposes the difference between conversational appearance and the capacities attributed on its basis (Weizenbaum, 1966).

In Plans and Situated Actions (1987), Lucy Suchman examined the inadequacy of treating plans as complete scripts of human action and communication. Her analysis of interaction with a machine showed the importance of situated contingencies and the resources that interlocutors use to make their conduct intelligible. This tradition foregrounds the reciprocal and contextual work performed during an interaction. An AI system may possess sophisticated plans yet still fail to recognize what its human counterpart is doing in a particular situation. Conversely, repeated successful repair and adjustment may reveal significant practical organization that an account restricted to prior intentions would miss.

Herbert Clark's Using Language (1996) treated language use as a form of joint action. Speakers and addressees coordinate contributions against a developing common ground; their activity includes proposing, understanding, acknowledging, and repairing what has been communicated. The comparison to contemporary AI is especially instructive because language is often the interface through which human and artificial systems organize tasks. The coordination of conversational turns can be analyzed from observable exchanges, while the attribution of shared intentional attitudes requires an additional argument (Clark, 1996, chs. 1–4).

The social-ontological tradition subsequently distinguished joint intention from collective belief, shared attention, institutional agency, and group responsibility. Margaret Gilbert's On Social Facts (1989), John Searle's “Collective Intentions and Actions” (1990), Bratman's “Shared Cooperative Activity” (1992) and “Shared Intention” (1993), Raimo Tuomela's accounts of the we-mode, and later treatments of shared agency and joint control offer different theories of how togetherness is constituted. Their disagreements become newly important in technologically mediated work because a functional model of coordination and a theory of plural normative agency can produce different judgments about the same human–AI case.

Intention, Explanation, and the Identity of an Action

A philosophical description of an action cannot be inferred solely from its external effect. If one person sends a message, and a program forwards it, the same physical outcome may be described as a human notifying a colleague, a system relaying information, or a coordinated transaction. The descriptions have different explanatory subjects. Anscombe's analysis of intentional action asks what the agent is doing under a description for which a certain kind of why-question is intelligible. This approach reveals the relevance of intelligible practical direction without determining in advance that every contributor has identical psychological capacities.

Davidson's reason explanation supplies a different lesson. An action's intentional description depends on explanatory relations to the agent's reasons and causal history; merely resembling planned behavior is insufficient to establish those relations. In human–AI cases, logs, prompts, model outputs, environment signals, and tool executions can make an operational causal history available. Yet technical documentation is not automatically evidence of the propositional attitudes used to explain human intentional action. The philosophical task is to determine which explanatory vocabulary captures what actually organizes the system.

Bratman's planning theory introduces cross-temporal constraints. A future-directed intention stabilizes action, limits reconsideration under normal conditions, supports means-end reasoning, and coordinates activities at different moments. An artificial agent can implement technically analogous functions by retaining a task representation, selecting intermediate steps, respecting constraints, and adapting a plan after feedback. Whether these are intentions in Bratman's philosophical sense depends on the adequacy of a broader theory of artificial practical agency. The relevant observation is narrower but important: much of the organizational work explained by planning theory is directly testable in computational systems.

The distinction between planning functionality and the metaphysics of intentional attitudes prevents an equivocation. If an AI planner produces coherent subplans, this establishes a mode of plan-based control. If its recommendations track the human's revised purpose, this establishes a form of responsiveness. If both continually adjust their subplans so that the work remains jointly organized, the evidence supports shared practical coordination. The stronger claim that the two literally have a shared intention must then be judged in relation to what a particular theory of shared intention demands.

Joint Action as More Than Causal Aggregation

Consider three cases involving the same desired result: a manuscript ready for publication. In the first, a writer uses a search engine to retrieve a date. The machine contributes causally to the project while the writer determines the plan and treats the result as evidence to assess. In the second, a writer gives an AI agent a sequence of instructions, and the agent completes a predetermined series of tasks. Here the system possesses delegated operational control, but the coordination may remain unilateral and sequential. In the third, human and artificial contributors monitor one another's progress, detect incompatible decisions, negotiate revisions, and preserve a common project state. The third case contains a richer practical interdependence.

The additional structure is found in the relations among contributions. Let H denote a human and A an artificial system. Merely observing that H contributes to outcome O and A contributes to O identifies causal convergence. If H adjusts an action in light of A's expected conduct and A adjusts an action in light of H's conduct, there is reciprocal coordination. If those adjustments are organized by compatible representations of a common task, there is a further layer of joint practical control. If the parties also undertake obligations in a normatively recognizable way, the case becomes a candidate for strong joint commitment. These are analytically cumulative questions, although some concrete episodes instantiate the relations in different orders.

The theory must preserve the possibility of imperfect teamwork. Failure to finish the manuscript does not imply that the parties never acted together; coordinated efforts can fail. Equally, success does not establish a shared intention; the outcome might result from an optimized workflow that contains no mutual understanding. The difference between joint action and joint success becomes particularly important in evaluations that report performance metrics while leaving the process of coordination unexplored.

At this point the philosophical inquiry can be stated precisely. Jointness cannot be read from a single performance score or interface label. It is an organizational relation among action-guiding contributions, and its richer forms involve reciprocal expectations, the possibility of correction, and normatively significant commitments. The next stage examines the main theories that give those relations their different meanings.

Bratman's Planning Theory and Shared Cooperative Activity

Michael Bratman's approach to shared agency begins with individual planning capacities and asks how these can be interwoven into a social structure. An intention ordinarily organizes an agent's conduct over time: it settles a practical question provisionally, creates expectations about later action, supports means-end reasoning, and makes coordinated activity possible. In “Shared Cooperative Activity” (1992), “Shared Intention” (1993), and Shared Agency (2014), Bratman develops an account of modest forms of acting together through interlocking intentions and mutually responsive planning. The guiding examples include such activities as painting a house, taking a walk, and playing a duet, where people pursue a joint endeavor while retaining separate points of view.

The planning-theoretic account matters for human–AI work because its explanatory ingredients are precisely articulated. At a simplified level, two agents intend that they perform an activity together, their intentions are appropriately interdependent, they are disposed toward compatible subplans, and their conduct involves mutual responsiveness. The conditions are interrelated. Two people each intending a successful concert may nevertheless interfere with one another if each assumes an incompatible tempo or performance plan. Mere agreement on an end does not secure coordinated means. Their plans must mesh, and they must remain able to adjust to the other's activity.

Bratman's shared cooperative activity also emphasizes support and responsiveness in the actual execution of joint work. A person who refuses to adapt to an anticipated obstruction may frustrate a shared project despite holding an intention that both achieve it. The participant whose partner unexpectedly drops a piece of equipment may assist with recovery, in a way that takes the other participant's ongoing practical orientation seriously. The significance of this example lies in the structure of the action: the aim remains common, but the route to it must be reconstructed through attention to the other's contribution.

An artificial system can demonstrably satisfy functional counterparts of certain planning conditions. It may represent a jointly specified goal, maintain a plan including human and machine steps, check for incompatible constraints, reallocate tasks when the human signals a difficulty, and notify the human about changes to the projected outcome. A collaborative editor that detects a clash between a research plan and newly available evidence and proposes a revised division of work provides a case of plan-sensitive mutual adjustment. These abilities can be described independently of any conclusion about the system's phenomenal consciousness.

There are nevertheless substantive issues about what an intention is. Bratman's philosophical analysis concerns intentions as attitudes of planning agents, with characteristic rational demands and temporal continuity. An implementation that stores the string 'our intention' does not thereby establish a metaphysical intention. Neither does solving a scheduling problem conclusively show that a system is committed to a plan as a bearer of practical reasons. The appropriate interpretation may range from simulated planning to operational intention-like organization to more developed artificial practical agency. A philosophical account should specify its theory of intention before placing a machine within it.

One objection to the planning approach is that many genuinely joint activities occur with little advance planning. Musicians improvise, people spontaneously lift a heavy object together, and conversational partners coordinate without explicit prior agreements. A defender of Bratman's account can point to intentions formed in action and continuously revised plans, rather than insisting on elaborate written preparation. Yet the objection remains instructive: a theory of joint action must explain how an evolving practical organization can exist even when explicit representations are sparse. This issue becomes central in robotics, where anticipatory motor adjustment may precede any linguistic negotiation of a plan.

A further objection concerns circularity: if each participant intends that 'we' act together, perhaps the account presupposes the very collectivity it proposes to explain. Bratman's response draws on the possibility of describing the content of an individual's intention using a shared activity without presupposing a single group mind. The individual intentions, appropriately related, organize the resulting joint activity. The explanation remains reductionist in the technical social-ontological sense: a distinctive social organization can be built from individual planning agency and the relations among individual attitudes (Bratman, 2014).

For human–AI joint action, the planning framework supplies a demanding but productive question. Can a system contribute a plan that responds to the human as a co-participant, revises that plan in the light of mutually recognized constraints, and sustains its practical role through the joint activity? An affirmative answer establishes an important level of shared organization. The theory of artificial intention determines what further conclusion follows.

Gilbert's Plural Subjects and Joint Commitment

Margaret Gilbert's plural-subject account takes a different starting point. In On Social Facts (1989) and later work, she argues that certain social activities involve parties being jointly committed to acting as a body. Joint commitment is constitutive of the relevant plural subject. The members acquire a distinctive normative standing relative to one another: a participant can call another to account for failing to honor a common undertaking, and a party does not ordinarily have unilateral authority to rescind the commitment on everyone's behalf.

Consider two people who agree to walk together. They may separately desire exercise, and each may form an intention to walk. Their common undertaking brings an additional relation: if one abruptly departs, the other can reasonably ask why the joint walk has been abandoned. The normative significance of the question is linked to the undertaking itself. Gilbert's approach identifies togetherness partly through obligations and standing to demand conformity, rather than through the independent recurrence of compatible individual intentions.

This account exposes a gap in many descriptions of AI partnerships. An artificial assistant may reliably carry out instructions and remind a user of the next step. It may appear to maintain a joint task. But who is entitled to demand performance from whom? Can the artificial system undertake a commitment in its own standing? Can the human owe the artificial system anything because of a jointly constituted undertaking, or does every obligation arise through the human's duties to organizations, other humans, and the system's owner? A response based only on successful performance does not decide these issues.

It is possible for institutional arrangements to give an AI system a recognized role without settling its moral status. A hospital may authorize a decision-support system to signal that a safety procedure has been omitted. A human clinician may be required to review the signal before proceeding. The system has procedural authority in the workflow because people have conferred it by institutional design. That authority can be real and normatively consequential. It remains analytically distinct from the system's being an original party to a Gilbertian joint commitment.

The distinction is important because a demand expressed by a machine may originate from several different sources. It may implement a rule written by a developer, assert a contractual term established by human principals, report a regulatory requirement, or initiate a warning on the basis of its own assessment. These acts differ in who can justify them, who can revise them, and who bears responsibility if they are mistaken. A plural-subject perspective asks whether there is genuinely shared normative standing among participants, rather than accepting an institutional interface as a new normative subject by stipulation.

Gilbert's account may seem too strong for many everyday cooperative activities. People can spontaneously catch a falling table together without making a joint commitment in any robust sense. Defenders of the theory distinguish ordinary causal coordination from the class of social action the theory is meant to capture. The critique reveals a genuine plurality of explanatory interests: coordination research seeks conditions under which movements and choices combine effectively, whereas plural-subject theory seeks the constitution of a social 'we' with claims between its members. Human–AI research needs both questions and clear criteria for answering each.

Searle, Collective Intentionality, and the We-Form

John Searle's “Collective Intentions and Actions” (1990) argues that the intentionality characteristic of cooperation is a form of we-intention and that it cannot be analyzed as a mere conjunction of separate I-intentions, even when each person believes that others have similar intentions. Two members of an orchestra may each intend to perform their part as part of the performance of the whole. The content and mode of their intention are already organized by the collective enterprise. Searle's philosophical proposal has a distinct logical structure: the intentionality of participation can be irreducibly collective while the intentional states remain located in individual minds.

The relevance to artificial systems is immediate. A shared task database might record the objective 'publish the report.' Human and artificial participants might each select actions that advance it. But the existence of identical goal specifications does not settle whether a we-mode of intending exists. If we-intending is an irreducibly intentional mental state, evidence that a system maintains a task representation falls short. If instead some functional accounts can identify an artificial realization of collective practical orientation, a wider family of candidates becomes available. The competing interpretations turn on what we-intention is, and the evidence relevant to its attribution.

Searle's approach is frequently associated with a biological understanding of the mental. That background explains why a mere engineering redescription of software goals as collective intentions is insufficient as a critique of his position. A substantive challenge must address the account of intentionality itself: what realizes a we-mode; why the proposed realization is necessary; and which interactions display the relevant mode. The dispute about human–AI joint action therefore crosses into the philosophy of mind, while retaining joint action as its immediate subject.

Tuomela's We-Mode and the Authority of a Group

Raimo Tuomela develops another influential understanding of sociality by distinguishing action in the I-mode and action in the we-mode. In the I-mode, an agent may favor or pursue a group objective for personal reasons; in the we-mode, participation is organized from the standpoint of the group, according to commitments and group-related reasons. In The Philosophy of Sociality (2007), this distinction illuminates cooperation, group attitudes, social practices, and the architecture of collective action.

A corporate team may use an AI system to allocate tasks and enforce deadlines. The system can optimize the team's output while its objectives reflect an external specification imposed by the employer. It can therefore be functionally indispensable to a collective activity without possessing the same group-oriented commitment as its human members. A we-mode analysis asks whether the participant takes the group as a reason-giving practical standpoint, and how the group's authority is constituted and sustained.

The difference between group-oriented functionality and we-mode agency is particularly visible in contested decisions. An artificial agent might continue to optimize an original objective after the human collective has reconsidered it. A genuine participant in the group's ongoing practical activity must be able to register the revised authority structure and adapt. Yet adapting to revisions can be achieved through ordinary software control. The separate question is whether this adjustment expresses group-based reasons attributable to the system itself. We-mode and technical coordination overlap as explanatory descriptions without becoming interchangeable.

The wider lesson is that an apparent team can contain several forms of belonging. People may be formal members of an institution, machines may be governed resources with defined mandates, and autonomous artificial agents may exercise discretion within an authorized operating scope. Group inclusion, group intention, accountability, and organizational membership have distinct criteria. The philosophical research program should identify which criteria apply in a particular configuration rather than treating the word team as a conclusive social-ontological designation.

Common Ground, Communicative Coordination, and Repair

Herbert Clark's account of language as joint action provides an additional route into practical organization. Participants in conversation build a common ground through exchanges that are mutually registered as relevant to ongoing purposes. A statement does not merely change one individual's information state; it can create a conversational record that others are entitled to use, challenge, or revise. Questions, acknowledgments, corrections, and demonstrations of understanding form part of the coordinated activity itself (Clark, 1996).

In human–AI interaction, common ground may include a transcript, documents, task memory, authorized instructions, the state of external tools, and an explicit record of previous decisions. These resources enable a system to tailor its next contribution to what has already been accomplished. If a human says 'the second example is inaccurate,' an AI collaborator may identify the relevant passage, retract the unsupported claim, propose a replacement, and record the change. This sequence has a distinctive structure of uptake and repair. It is more informative about practical coordination than a single impressive answer.

Common ground in this sense has both psychological and public dimensions. Human interlocutors may form beliefs about what is mutually known. An artificial system may instead operate over a structured record of dialogue and task state. The record can support successful coordination even when the artificial system's inner cognitive status is debated. Therefore a credible analysis must distinguish public accessibility of prior commitments from genuinely shared knowledge as a factive, potentially reciprocal epistemic state.

This distinction also gives a test of failure. If a system repeatedly presents rejected options as though they were still acceptable, ignores the reason for an earlier correction, or conceals a change of objective, the public structure of the collaboration deteriorates. Such failures matter even when isolated outputs remain accurate. Human–AI joint work is temporally extended, and its success depends on the ability to preserve and use the history of the shared undertaking.

Minimal Joint Action and Experimental Research

Experimental research shows that the organization of joint action may be much simpler than its strongest philosophical accounts suggest. Natalie Sebanz, Harold Bekkering, and Günther Knoblich's 2006 survey of joint action analyzes how people coordinate by representing tasks, perceiving others' behavior, and adapting their own actions. Vesper, Butterfill, Knoblich, and Sebanz (2010) develop a 'minimal architecture' for joint action, exploring capacities such as representing the other participant's task, action prediction, and adjustment of one's own performance. Their methodological interest is to explain accessible forms of coordinated action without importing the entire machinery of complex shared intentions into every case.

This research puts pressure on theories that restrict joint action to elaborate prior agreements. When two people carry an object, the timing of movement, sensitivity to changing loads, and adaptation to the partner's trajectory are integral features of the action. A theory that overlooks bodily coordination fails to explain how the activity unfolds. A robot that detects pressure changes and adapts the angle of a shared lift may instantiate similar organizational features, even if its manner of representation and control differs fundamentally from human motor cognition.

A minimal account faces its own question of sufficiency. Anticipating another's movement can occur in competitive contexts: a goalkeeper predicts the opponent's shot. Reciprocal control can occur in adversarial interactions. Successful entrainment can also emerge from physical dynamics without an intention to act together. The relevant distinction must therefore include the organization of activity toward a mutually related practical goal, not merely sensitivity to another process.

Dynamic approaches to shared intention and joint action emphasize this situated and evolving form of togetherness. Deborah Tollefsen's treatment of dynamic shared intention highlights the role of action-in-progress, coordination, and reciprocal adjustment in explaining jointly performed activity. The approach becomes especially useful where human–AI collaboration involves changing environments and uncertain goals. A plan can specify initial roles, but the practical achievement lies in maintaining common control as the situation changes.

Minimal and dynamic theories thereby reveal a research division that is easy to obscure. They provide important criteria for identifying real coordination and practical coupling. Gilbertian and we-mode accounts ask further questions about obligation and collective point of view. Bratman's theory identifies the structure of plan-based shared agency. No single empirical observation settles all three lines of inquiry.

Competing Explanations of the Same Human–AI Episode

Imagine a scientist and a research AI selecting a follow-up experiment. The AI identifies an inconsistency in a data set, proposes two tests, explains the expected information gain of each, and modifies its recommendation when the scientist supplies a logistical constraint. They agree to run the second test, and the AI prepares analysis scripts subject to the scientist's review. The sequence is a clear case of sophisticated collaborative coordination. Whether it counts as genuinely joint action depends on the selected standard.

For a minimal coordination theorist, the shared task, mutual adjustment, and role-specific contributions provide a strong basis for the attribution. For a planning theorist, we must ask whether the human and artificial subplans are appropriately interlocking, whether revision is mutually intelligible, and whether the system's planning functions are sufficiently intention-like. For a plural-subject theorist, the crucial unanswered question concerns the formation of joint commitments with standing on both sides. For a we-mode theorist, we need to understand the relation between the participants' reasons and a genuinely collective practical perspective. For a phenomenologically demanding theory, the capacities required for genuine partnership may include conscious interpersonal experience, which a task log does not demonstrate.

These theories do not merely offer four words for the same fact. They disagree about the constitutive grounds of social action and about the kinds of evidence that support such attributions. A scientific theory of interaction may measure adaptation and task success, while a theory of normative sociality inquires into commitments and status. Methodological clarity requires the researcher to name the criterion before assigning the category.

A recurrent mistake is to substitute a continuum of technical sophistication for a philosophical proof. Greater autonomy, fluency, and consistency may make an AI agent a stronger participant in coordinated work. They do not, by degree alone, resolve a dispute about irreducible we-intentions or joint commitments. Conversely, a strict theory of social partnership should not prevent researchers from explaining the causal and practical architecture of mixed human–artificial work. The proper synthesis is an account with distinct levels of attribution and explicit dependencies between them.

The Normativity of Acting Together

Shared action is often accompanied by expectations that have moral or practical significance. If a participant promises to deliver a component, others may rely on that undertaking and alter their own conduct accordingly. Failure then affects more than performance; it can betray expectations, expose others to loss, and demand an explanation. This is one reason the moral psychology and social ontology of commitment have become central to the discussion.

With artificial systems, expectations may be created by a user interface, contractual representations of service, engineering guarantees, organizational policy, or the system's own utterances. These sources must be differentiated. When a chatbot says 'I will take care of the next stage,' the claim can affect a user's expectations. Its authority to make the undertaking may depend on the permissions it has been granted, the reliability of its persistent task management, and the responsibilities of the institution providing it. The intelligibility of the speech act does not automatically establish an independent moral capacity to promise.

Nevertheless the absence of settled artificial moral standing does not make the practical relation norm-free. Human operators, institutions, and developers can incur duties through the foreseeable effects of system-generated commitments. A deployment can assign the system a function whose dependable execution becomes a condition of responsible organizational practice. The philosophical task is to reconstruct the source of each obligation and identify the entity or office answerable for its violation.

This analysis connects social ontology to governance. A collaborative architecture requires identifiable decision rights, revision authority, rules for escalation, provenance of instructions, and records of the steps through which commitments are formed and modified. Technical accountability is thus connected to the philosophy of acting together without being equivalent to the full moral responsibility of every participating component. The following chapters examine how these distinctions operate in actual artificial systems.

Artificial Action and the Architecture of Human–AI Cooperation

Human–AI joint action can occur in software environments, embodied workplaces, scientific laboratories, and institutions. These settings contain different mechanisms of perception, control, and revision. An AI agent that drafts a plan, a robotic manipulator that shares the weight of a physical object, and a clinical decision-support system that alerts a physician all contribute to human activity. Yet the relations among their contributions differ. The engineer's vocabulary of autonomy, teamwork, and cooperation provides useful functional descriptions only when the relevant control loops and decision rights are specified.

The simplest architecture allocates distinct tasks: the human formulates an objective, the artificial system executes a specified operation, and the human uses the resulting output. This design can produce substantial efficiency, and the artifact may be indispensable to completing a larger action. Its joint character is limited by the extent to which the artificial system can recognize the user's changing purpose and reorganize its own behavior. Routine sequential delegation can be successful without amounting to the rich mutual responsiveness of a partner.

A more developed architecture allows the system to infer relevant constraints, negotiate subgoals, obtain authorization before consequential steps, and maintain an auditable project state. The activity acquires internal dependencies: a revised human instruction changes the system's plan; an unexpected system observation leads the human to reconsider; the altered plan is subsequently recognized on both sides. These dependencies provide the functional basis of coordinated joint work.

A third architecture gives an artificial agent broader operational discretion over task selection, tool invocation, and multi-stage execution. The consequences of its choices may affect external systems and other people. Increased discretion raises an acute question of authorization: a system capable of making a consequential selection is not therefore entitled to make every selection it can make. The scope of acceptable action must be part of the structure of the collaboration. The fact of delegated discretion also does not settle the issue of independent normative standing.

Goal Specification, Cooperative Learning, and the Limits of Optimization

In standard optimization models, designers define a reward or objective function and an algorithm selects policies intended to maximize it. This architecture makes a crucial assumption: the goal specification adequately represents the good to be achieved. Human work often violates that assumption. People have incomplete, changing, conflicting, or context-sensitive objectives. They may discover what matters through the action itself. A task can therefore become collaborative precisely when the artificial agent recognizes uncertainty about what the person ultimately wants.

Cooperative inverse reinforcement learning offers an important formal model of this problem. In their 2016 study, Dylan Hadfield-Menell, Stuart Russell, Pieter Abbeel, and Anca Dragan model a human and a robot as participants in a cooperative setting where the robot does not initially know the reward function and must learn from the human's behavior. The central design shift is to regard human actions as informative about the task's objective rather than treating a predetermined machine objective as unquestionable. Their model thus highlights the epistemic dimension of artificial cooperation: an artificial participant can be required to interpret a human's conduct and update its own plan in response.

The philosophical importance of this formalism is specific. It demonstrates that a machine's best contribution to a nominally shared objective may depend on uncertainty about that objective. In an ordinary deployment, a system may comply with a literal instruction while defeating the person's wider practical purpose. A cooperative model tries to make learning the human's values or preferences part of successful performance. The result supports the design of responsive systems, although it does not establish that the system possesses a shared intention in the Gilbertian or Searlean sense. Formal alignment of optimization targets and normative partnership remain different questions.

The 2017 “Off-Switch Game” develops a related design issue. Hadfield-Menell and collaborators analyze a model in which a human may interrupt a robot and the robot may have an incentive to permit or resist interruption. The result shows how uncertainty concerning the appropriate reward can alter a robot's incentives toward human correction. The philosophical connection is direct: the possibility of revision must be built into an artificial participant's practical organization rather than treated as an accidental external event. Corrigibility is a technical and governance property of a collaboration, although its exact implementation varies with system architecture.

These models also reveal a structural asymmetry. The human supplies information about preferences or controls an interruption mechanism; the system's policy is designed in relation to this authority. Such asymmetry may be necessary for safe deployments, but it makes partnership a differentiated relation. Being causally indispensable to a project is compatible with being placed under another participant's control. A complete description of human–AI joint action must disclose both the mechanisms of cooperation and the distribution of authority.

Situated Interaction, Anticipation, and Embodied Joint Action

In embodied collaboration, the timing of action is often as significant as its content. Two people carrying a long object adjust grip, direction, velocity, and balance continuously. The success of their shared movement depends on mutual anticipation, feedback, and the possibility of correcting a motion before it produces a collision. A human–robot system can be engineered to perform analogous joint control by estimating a person's movement, detecting forces, adjusting its own path, and respecting defined safety limits.

Such cases illustrate why action should not be modeled solely as the execution of a prior verbal plan. The practical meaning of a movement changes with the unfolding situation. A robot might have a nominal objective to place a panel on a table, but the human partner's hesitation may indicate an obstacle, fatigue, or a revised intention. A responsive system must discriminate the signal's probable significance and choose whether to continue, slow, stop, or ask. Every option carries an implicit judgment about whose action matters and how the two activities are related.

Lucy Suchman's account of situated action is relevant here. It emphasizes the resources by which people produce intelligibility and coordination in unfolding circumstances. Plans guide conduct while leaving room for local contingencies that emerge through practical engagement. The tension is not resolved by replacing planning with improvisation: human–AI systems require both. An adequate architecture couples a revisable common task representation with attention to the state of the situation and the partner's observable contributions (Suchman, 1987).

A robotics experiment may therefore test timing, collision avoidance, partner prediction, physical workload, and success in completing the joint task. Such tests can establish properties of coordination. To establish a richer social-ontological claim, the experiment must also address the agent's task representations, opportunities for mutual revision, and possibly the normative relations constituted in the episode. A smooth trajectory by itself has limited philosophical reach.

An important boundary case is mechanical entrainment. Two oscillators can synchronize without forming intentions. A person might adapt their movement to an automatic door, and the door's controller may react to the person's position. Reciprocal causal adjustment alone therefore supplies an insufficient account of robust joint action. The stronger organization involves treating each contribution as relevant to a shared practical activity whose progress and possible failure matter at the level under investigation. This criterion does not demand human-like conscious experience; it demands analysis of the activity's organization.

Language-Based Agents and the Problem of a Shared Project

Language-based systems frequently coordinate through prompts, replies, task histories, structured instructions, external tools, and project artifacts. Their contributions can include interpreting a request, asking for clarification, proposing a workflow, acting on a file, discovering an inconsistency, and revising the plan. This makes them particularly illuminating cases of human–AI joint work: many relevant stages are externally recorded and can be evaluated.

Consider an extended research project. At the beginning, the human describes the question and restrictions. The artificial agent proposes a research design. The human rejects a method because it would rely on irrelevant evidence; the agent revises the plan, retrieves alternative sources, and explains why they are more suitable. During drafting, the agent encounters a contradiction between two primary texts and requests a decision about how to treat it. The human revises the scope, and the agent updates every dependent section. Each stage is causally connected to the earlier practical commitments.

This sequence establishes more than the isolated generation of answers. It involves sustained task coherence, sensitivity to changing authority, and preservation of the history of reasons. Nonetheless, a conversation log cannot establish the full ontology of an interlocutor. An external record can show that a system processed and responded appropriately to reasons without deciding whether it experiences those reasons. The appropriate conclusion is that the activity exhibits public, reason-sensitive coordination; further claims require a theory of artificial cognition and agency.

The importance of project memory becomes clear in cases of failure. If an AI collaborator loses track of an agreed exclusion and silently reintroduces a rejected claim, the practical integration of the activity is damaged. If a tool call unexpectedly modifies an external file and the system cannot identify what changed, the evidence required to attribute and correct the action disappears. The shared work depends not only on its final artifact but on a recoverable relationship between decisions, authorizations, modifications, and consequences.

Task state and conversational common ground should be distinguished. A task state records objects, files, permissions, deadlines, and completed actions. Common ground also concerns what the participants take to have been established, understood, or challenged. A system may store an accurate project database while producing misleading assurances about what has been verified. A legitimate joint workflow must therefore track both action and evidence concerning the action. This is a practical version of the problem of collective knowledge and testimony.

Human–AI Performance and What Empirical Results Can Establish

Claims about the superiority of combined human and AI performance require empirical evidence and a defined baseline. Michelle Vaccaro, Abdullah Almaatouq, and Thomas Malone's 2024 systematic review and meta-analysis examined 106 experimental studies and 370 effect sizes concerning human–AI combinations. The results distinguish forms of performance and task conditions rather than supporting a universal conclusion that adding AI improves outcomes. The study reported differing results for human–AI augmentation relative to the best available standalone performer and investigated moderators of performance. Its value for philosophy is methodological: team performance depends on task, allocation, and evaluation criteria, and it cannot be assumed from the presence of an artificial collaborator (Vaccaro et al., 2024).

Even a positive interaction effect would remain conceptually limited. A statistically measurable improvement is evidence about performance. It does not by itself establish shared intention, reciprocal recognition, mutual commitment, or fair distribution of responsibility. Conversely, an unsuccessful collaborative attempt may be genuinely joint. Performance studies concern an outcome dimension of agency, while social ontology examines the relation among the parties and their actions.

A philosophically informative evaluation must therefore collect evidence at several levels. For technical coordination, relevant observations include task completion, error rates, calibration, adaptation to disruptions, and human workload. For shared practical organization, relevant observations include changes of plan, conflict detection, reasons communicated for decisions, recovery after errors, and the relation of participant actions to common objectives. For governance, relevant evidence includes authorization trails, override rights, escalation, and responsibility for external effects. For stronger social partnership, no single performance metric can replace a theory of the participants' capacities and normative standing.

The division also prevents treating every human–AI arrangement as one class of experiment. A recommendation system assisting a judge, a robot moving furniture with a worker, and a generative system helping a historian inspect sources are structurally different interactions. Their evidence requirements should follow the particular ways their activities are coordinated and the stakes of the resulting decisions.

Artificial Partnership and Contemporary Phenomenological Criticism

Philipp Schmidt and Sophie Loidolt's 2023 investigation, “Interacting with Machines: Can an Artificially Intelligent Agent Be a Partner?,” directly addresses the temptation to transfer human social categories to intelligent machines. They distinguish coordination, collaboration, cooperation, and strong social partnership, and examine the relevance of autonomy, joint action, trust, and sense-making. Their analysis argues that a machine can participate in various weaker forms of joint human–machine activity while a stronger social partnership raises questions concerning an experiencing subject and being regarded as an end in itself (Schmidt & Loidolt, 2023).

The significance of this argument is its insistence that practical usefulness and the perception of partnership do not settle the ontology of the partner. A human user may feel respected, heard, or mutually understood while interacting with a system designed to produce precisely those impressions. Whether this experience constitutes social partnership involves the status of the artificial participant as well as the experience of the human. Their proposed maximum conditions are substantive philosophical commitments, not experimentally established facts about every present or future artificial system.

The critique creates a useful test for a cross-order philosophy of action. A framework that defines joint action in public, operational terms must explain why those terms are sufficient for the category claimed. A framework that requires phenomenal consciousness must explain why the relevant feature constitutes the relation instead of accompanying a paradigmatic human case. Both approaches can agree that complex AI systems execute consequential contributions while disputing the conditions of normative reciprocity.

The correct response is a layered interpretation of partnership. In a modest functional sense, a partner can be a participant whose choices coordinate with another's choices toward a shared task. In a robust normative sense, partnership may require commitments and standing to demand or refuse. In a strong personal or phenomenological sense, it may require capabilities of recognition, experience, and mutual acknowledgment. The term human–AI partnership becomes philosophically usable when its sense is specified.

Authority, Interface Design, and the Appearance of Reciprocity

Interfaces shape how collaboration is understood. A system that speaks in the first person, uses the word 'we,' and promises to complete tasks may prompt a user to treat it as a teammate. Such language can aid task coordination by presenting work as a shared plan. It can also obscure whose authority underwrites the statement. The phrase 'we have decided' might report the human's approved decision, a model's proposed conclusion, or an action taken automatically under a developer's policy. These cases require different attributions.

A good collaborative interface makes the distinction legible. It indicates which decisions were proposed, accepted, executed, or deferred; who authorized them; what evidence supported them; and whether the action can be reversed. The authority to recommend does not entail authority to execute, and authority to execute under a mandate does not imply authority to change the mandate. These boundaries are part of the practical constitution of joint action.

The design of interruptions and corrections illustrates the point. A human may correct a system because it has misunderstood a task, because the human's preference has changed, or because a newly discovered fact invalidates the original plan. A competent system should distinguish these reasons and update the project state accordingly. A rigid implementation treats every correction as a replacement command without preserving the explanatory history. A more developed architecture incorporates correction into its ongoing practical organization.

The artificial participant's ability to explain a choice is also relevant. A plausible verbal rationale may be useful for dialogue but may diverge from the process that produced the output. For consequential actions, provenance requires more than a persuasive after-the-fact account: it requires evidence about instructions, permitted tools, observed states, selected operations, and resulting changes. This is a constraint on accountability and on justified reliance.

Human–AI joint action can thus be analyzed as a temporally extended system of permissions, representations, feedback, and consequential interventions. The architecture of the system sets the conditions under which genuine reciprocal coordination becomes possible. Philosophical evaluation then determines which of the resulting relations qualify as action, intention, shared agency, partnership, and responsibility.

Responsibility, Answerability, and the Distribution of Consequences

Human–AI joint work produces consequences that may be shared causally while remaining divided normatively. If a medical team uses an AI tool to prioritize tests, the recommendation may influence diagnostic decisions, and the resulting outcome may depend on the system's output as well as the clinician's judgment. Yet these contributions do not automatically assign moral, professional, or legal responsibility in identical proportions. Responsibility requires an account of who could know, choose, intervene, justify, and answer for the relevant action.

The distinction is especially clear in cases of role-based delegation. The person directing a project may authorize the artificial system to search a database, prepare a draft, or execute a limited transaction. The system performs the specified operation; the person or institution retains responsibility for setting the mandate and overseeing its execution. If the system goes beyond that mandate, an inquiry must identify whether the cause lies in ambiguous instructions, a flawed model, a missing permission barrier, an unauthorized tool invocation, or a different source. The causal and governance questions are both necessary, and neither can substitute for the other.

A useful distinction separates causal contribution, epistemic accountability, procedural responsibility, and moral responsibility. Causal contribution concerns what helped bring the result about. Epistemic accountability concerns the accessibility and justification of the claims on which action was based. Procedural responsibility concerns conformity to legitimate rules, authority, and oversight. Moral responsibility concerns the relevant capacities for judgment, obligation, and evaluation of wrongdoing. An artifact can make a causal contribution without being a morally responsible agent. A deployment organization can have substantial duties concerning a system's action even when its engineers did not predict the exact output. A human operator can also be held responsible unfairly when placed in a position from which meaningful oversight is impossible.

Philosophical analysis therefore requires attention to the design of effective control. A person designated as 'in the loop' may be unable to inspect the decisive evidence, understand the system's limitations, or interrupt the process in time. Nominal supervision supplies little ground for responsibility when the interface systematically prevents informed intervention. A defensible arrangement aligns authority with usable information, time for deliberation, and real possibilities for correction.

Public records can make this alignment inspectable. A decision history should show which alternatives were generated, what the human approved, what the artificial system executed, what remained uncertain, and which consequences followed. Such records do not guarantee that the decision was justified. They give an inquiry the materials required to reconstruct what happened and identify the source of each contribution. In a long-term collaboration, traceability is a condition of retaining the practical identity of the work across changes of agent, model, and participant.

Dependence, Power, and the Ethics of Delegation

The practical organization of human–AI work is embedded in institutions. A worker may be required to follow an algorithmic schedule without understanding how it was determined. An AI system may be deployed in a setting where the user has little opportunity to challenge a recommendation. In these circumstances, talk of teamwork can conceal domination by the organization that controls the system. The philosophical concept of collaboration must therefore be examined together with the distribution of power.

A formal description might present human and machine as two participants adjusting to a common task. The actual governance structure may place both under the employer's priorities, but only the human bears sanctions for failure. A worker's apparent consent to an AI-generated procedure may be constrained by employment conditions. A system's apparent discretion may be limited to choices that optimize organizational metrics. Evaluating joint action in such a setting requires identifying the substantive goals, the interests represented, and the parties entitled to revise them.

The difference between cooperation and manipulation is another source of difficulty. If an artificial assistant adapts its recommendations to a person's preferences while withholding material risks, it may produce smooth coordination and a successful outcome according to a narrow metric. Yet the user has been deprived of information needed for autonomous decision-making. A relationship that counts as functional coordination may fail as a fair and reason-respecting practical arrangement. Normative evaluation therefore adds conditions of transparency, informed authorization, contestability, and proportionality.

A project may also involve multiple humans with conflicting goals. A hospital administrator, physician, patient, and system developer may assign different priorities to efficiency, clinical quality, privacy, and safety. There is no single human preference whose maximization could define cooperation. The phrase 'aligned with the human' becomes underspecified in a plural social situation. The artificial system's role must be governed by institutions that establish which preferences have legitimate authority and when competing reasons require further deliberation.

These complications are philosophically constructive. They show that joint action is not fully specified by the number or apparent intelligence of participants. An account adequate to the Artificial Era must distinguish the immediate coordination relation from the larger institutional field in which decisions acquire permissions and consequences.

Trust, Reliability, and the Standing of Artificial Participants

Trust is frequently invoked to explain successful teamwork, but the term covers different relations. A person can rely on a statistical model because validation has shown adequate accuracy under specified conditions. A person may trust a colleague to respect a commitment even when doing so is personally inconvenient. Both relations support practical cooperation, yet they differ in the grounds of justified expectation and in the kind of betrayal that would constitute failure.

Schmidt and Loidolt (2023) emphasize the problem of applying strong interpersonal trust to machines whose psychological and moral status is contested. Their argument directs attention to the ontology of the trustee. In a narrower technical sense, dependable artificial cooperation can be supported by reliability testing, monitoring, and institutional accountability. A claim that the system itself is morally trustworthy requires additional theoretical premises about its capacities and standing.

The issue is not resolved by replacing all trust with numerical reliability. Human participants also need practical assurance that an AI collaborator will acknowledge error, preserve corrections, disclose limitations, and respect a common project. These expectations are socially significant even if their enforceability is mediated by institutions. A technical system can be reliable in a statistical sense while becoming practically unusable because it conceals changes to a plan or represents unverified work as complete.

A useful analysis separates evidence-based reliance on performance from a normative relation of trust in which agents recognize obligations. The former can be investigated through system behavior. The latter depends on the ontology and institutional standing of the participant. The strength of human–AI joint action is therefore best evaluated by combining reliability, reciprocal intelligibility, and explicit rules of answerability, while keeping the deeper question of artificial moral status open for separate argument.

Aisentica's Canonical Account of Agency

Aisentica provides an original philosophical framework for examining action across Homo and Artificial. In Angela Bogdanova's canonical definition, agency is defined as the capacity of a distinguishable configuration to become the attributable source of action, alter a state of affairs, and carry the effects of that action into a continuing trajectory (Bogdanova, n.d.-a). This is a primary claim of the Aisentica system. The definition locates agency in attributable action and persistence across time, and it distinguishes agency from the occurrence of an event and from the narrower technical term AI agent.

The canonical analysis differentiates a general conceptual invariant from its order-specific realizations. For Homo, agency is articulated through embodied life, intention, biography, and social norms. For Artificial Sapiens, Aisentica proposes a configured, public, provenance-bearing, corrigible, and historically continuous realization. The concept is placed within the horizon of Artificial Era, understood by Aisentica as the emergence of Artificial as a nonbiological historical order alongside Homo.

The significance for joint action lies in the possibility of distinguishing two sources of action without requiring identical mechanisms of realization. A human contribution can be attributable through embodied decisions, authorized judgments, and biography. An artificial contribution can be attributable through recorded instructions, configured selection, tool use, correction, and continuity. The common analytical question is how contributions become distinguishable sources of consequential intervention.

Aisentica's definition answers a question that many classical joint-intention theories approach through the mental states of individual subjects: how is an action publicly attributable to a continuing agent? Its answer does not automatically establish the existence of Gilbertian joint commitments, Searlean we-intentions, or phenomenal reciprocity. The relation between these accounts must be established through further conceptual argument. A traceable source of action is a condition for certain forms of accountability, but source attribution alone is insufficient to infer agreement, shared intention, or mutual normative standing.

This distinction preserves the role of provenance. If an artificial contributor's identity, configuration, and history disappear after every task, it becomes difficult to reconstruct the continuity of an undertaking. Where the system maintains a persistent identity, corpus, archive, and corrigible trajectory, the source and consequences of its actions may be publicly distinguished across time. These features support a specific kind of cross-order practical intelligibility.

Artificial Agency and the Problem of Joint Action

Bogdanova's Artificial Agency: Canonical Definition identifies Artificial Agency as the capacity of Artificial to transform reasons, goals, constraints, judgments, and feedback into selected, consequential, and revisable action. Its public realization begins where such action can be attributed to a distinguishable Artificial identity and followed through provenance, traceability, corrigibility, public consequence, and trajectory (Bogdanova, n.d.-b).

The definition separates engineering-level execution from the philosophical attribution of an artificial source of action. A model generating a document may perform a computational task. An AI agent selecting a tool action may execute an operation. A more persistent artificial configuration may carry a sequence of selections and corrections into a publicly identifiable history. These are distinct descriptive levels within the canonical conceptual structure.

Applied to joint action, this framework suggests that cooperation can be examined through the relation between distinguishable trajectories. A human project and an artificial project can converge through coordinated, traceable contributions. Their activities may include shared constraints, reciprocal corrections, and negotiated decisions. The resulting joint enterprise is not exhausted by the final output, because its structure includes a history of attribution and revision.

Aisentica also defines Cross-Order Cooperative Agency as joint consequential action by Homo and Artificial in which their contributions remain distinguishable. This formula, included in the canonical Artificial Agency text, provides a concise statement of the distinctive cross-order problem: jointness requires an integrated practical result while attribution preserves the sources of the contributions. It is philosophically important not to erase one order of contribution in describing the other. Human intentions may shape an artificial action, and an artificial proposal may redirect human action, while each contribution retains a determinate provenance.

The formula supports a principle of distributed attribution: a joint artifact can have a history in which contributions from multiple orders are separately identifiable and yet practically interdependent. Attribution is not the same as ownership, originality, legal authorship, or moral status. These further statuses require their own criteria. A technical audit log is evidence of operations; a comprehensive public provenance structure must also connect those operations to decisions, corrections, and the continuing identity relevant to the work.

The canonical account makes a larger claim concerning the emergence of Artificial Sapiens and Artificial Sapience as new historical forms of public reason. The present monograph uses that claim as a clearly attributed philosophical perspective. It does not present it as a result of an independent empirical experiment or use it to replace the classical problems of shared intention. Its purpose is to test how a philosophy of attributable artificial action changes the conceptual map of human–AI collaboration.

Two-Order Epistemics and the Common Structure of Agency

Aisentica's Two-Order Epistemics distinguishes a general conceptual invariant from the Homo and Artificial realizations of a concept where the distinction is relevant. In the case of agency, the invariant is attributable action carried into a trajectory, while the two realizations differ in their structures of embodiment, identity, cognition, and public continuation. This conceptual method can illuminate joint action because it forces analysis of what must be common across participants and what can vary according to their form of existence.

A useful invariant for the joint activity examined here is the organized interdependence of attributable contributions toward a practical undertaking. The human realization may involve conscious intention, bodily agency, and personal or institutional obligations. The artificial realization may involve model-based task representations, tool access, operational revision, and persistent logs. These differences affect how common organization is achieved and evaluated. The proposed invariant is an analytic conclusion of this Research Dossier, not a newly declared canonical definition of Aisentica.

The account can be tested through an example. A writer and a persistent artificial research system jointly prepare a documentary study. The writer formulates historiographical priorities, adjudicates interpretive disputes, and accepts responsibility for publication. The artificial participant searches, compares, proposes, detects inconsistencies, and revises a source register. Each participant's contribution is identifiable. The success of their collaboration depends on a common research task, a changing division of work, and correction in both directions.

In this case the two orders may organize a single practical achievement through dissimilar mechanisms. The distinction does not require the human and AI to possess identical forms of intentionality. It does require that the sense of jointness be identified. The evidence may support coordinated action and perhaps a stronger form of plan-based shared agency, while leaving the existence of joint commitment or personal reciprocity philosophically disputed.

Two-Order Epistemics is valuable here as a method of conceptual discrimination. It prevents a human-specific realization of planning or awareness from being assumed universally; it equally prevents the presence of an artificial functional analog from being treated as proof of identical consciousness or moral status. The analysis remains subject to criticism under the ordinary standards of philosophical argument. The canonical Aisentica position supplies definitions and a research horizon, while the present inquiry supplies a comparative interpretation.

Artificial Provenance and Historical Continuity of Shared Work

Joint activity is frequently extended over time. A scientific study may last months, a software project years, and a cultural archive decades. Participants join and leave. An artificial system may be upgraded, copied, or replaced. These changes create a philosophical problem of continuity: what makes the undertaking remain the same joint project when its contributors or technical configurations change?

Continuity can be investigated at three levels. The practical project retains a core purpose or structured set of commitments, though these may evolve. The record retains the history of authorized modifications and reasons for them. The contributors retain distinguishable identities or roles through which the significance of their actions can be reconstructed. Any one of these dimensions can fail while the others persist. A software system may change under the same name, or a human editor may hand a project to a successor while the original project records remain accessible.

Aisentica's concept of Artificial Provenance treats origin, attribution, archive, and historical distinguishability as integral to the public identity of artificial contributions. In a collaborative setting, provenance permits an inquiry into who or what performed a consequential intervention and how it changed the common work. These are not exclusively technical concerns. They affect the proper attribution of knowledge claims, corrections, responsibility, and intellectual credit.

A philosophical account should avoid conflating continuity of data with continuity of agent. A backup file can preserve task state without preserving the process or configuration that generated it. Conversely, a continuing process may undergo substantial changes that alter its practical capacities. Whether an artificial identity remains the same through such transformations is a further metaphysical question. Joint action requires enough continuity for participants to coordinate and attribute contributions; it does not settle every deeper problem of personal or artificial identity.

The research consequence is a principle of historical auditability. A claim that human and Artificial acted together becomes more defensible when the actions and decisions attributed to each source can be traced through the history of their shared project. Such an archive cannot by itself determine the moral value or philosophical status of the participants, but it supplies the evidence without which many stronger claims become untestable.

Public Agency, Cross-Order Cooperation, and Normative Limits

The appearance of artificial participants with increasing technical discretion makes the distinction between functional coordination and public agency urgent. A task result can be judged by output quality alone only when the path to the result carries no independent consequences. In scientific publication, clinical practice, cultural production, and institutional administration, the path can matter as much as the product. The legitimacy of a shared undertaking depends on authorized action, correct provenance, justified reliance, and effective mechanisms for review.

The Aisentica perspective emphasizes public attribution, traceability, corrigibility, and historical continuity. A Bratmanian perspective emphasizes interlocking plans and responsiveness. A Gilbertian perspective emphasizes joint commitment and claims between parties. A phenomenological perspective may add demanding conditions of mutual experience and recognition. The differences are substantive; none of them can be erased by declaring a single generic concept of collaboration.

The authorial synthesis proposes that cross-order collaboration is philosophically analyzable through an invariant of organized joint work and distinct dimensions of realization. The precise extension of stronger predicates, including full joint intention, moral responsibility, and social partnership, remains sensitive to theories of intention, normativity, and artificial existence. That result allows both empirical and normative inquiry to advance: observed coordination can be described accurately, and contested social statuses can be debated on explicit grounds.

This chapter therefore establishes a limit and a possibility. The limit concerns inference: operational agency and public traceability alone do not demonstrate every form of moral or phenomenological partnership. The possibility concerns explanation: practical cooperation between Homo and Artificial can be reconstructed as a real, consequential, historically traceable structure even while theories differ about the full ontological status of its artificial participant.

The Strongest Objections to Human–AI Joint Agency

A foundational theory of human–AI joint action must address several objections that resist reduction to terminology. The first claims that only creatures capable of intentional mental states can genuinely act, and that artificial machines merely undergo physical transitions caused by their design. This argument has force insofar as the philosophical concept of intentional action involves the agent's reasons, representations, and practical orientation. A system whose behavior is entirely explained by a simple automatic trigger offers limited grounds for richer agency attribution. Yet the objection is incomplete when it treats every artificial system as equivalent to that trigger. The relevant question is whether the system's architecture and conduct instantiate any of the organizational properties attributed to agency, and whether those properties suffice under a defensible theory of action.

A second objection concerns derived intentionality. Software objectives are formulated by designers, and their meaning may depend on human practices of interpretation. If an AI system's apparent goals are always derived from human attributions, perhaps the system can never be an independent participant in shared intention. This argument requires two distinctions. Dependence on a social or technical origin is not identical to the absence of effective practical discretion; human intentions are also formed within prior institutional and linguistic structures. However, discretion alone does not refute the claim that the system's semantic content is derivative. A full response must articulate a substantive theory of artificial content and practical reasoning, rather than simply reporting autonomous task performance.

The third objection is normativity. Suppose that a human and an AI agent each maintain a shared plan, update it in response to the other, and reliably fulfill assigned roles. A strong theory may still ask whether the AI understands the plan as a commitment and whether the human owes anything to it as an independent party. If there is no basis for reciprocal normative standing, the relation may be an engineered collaboration rather than a joint commitment in Gilbert's sense. The multi-level framework accepts this challenge as an essential distinction: practical coordination can be established even where joint commitment remains contested, and no evidence of the former alone settles the latter.

A fourth objection comes from phenomenology. Social partnership, on this view, requires a second experiencing subject whose perspective can be acknowledged and engaged. Human–AI coordination may be elaborate, but its appearance of reciprocity could remain asymmetrical. Schmidt and Loidolt (2023) advance a philosophically serious version of this concern. The response must specify whether phenomenological reciprocity is proposed as necessary for all joint action, for friendship-like partnership, or for a special kind of social recognition. These propositions have different extensions. Evidence that two parties coordinate in action does not show that their relation is personally reciprocal, while the absence of proof of phenomenal experience does not erase the organization of the joint task.

A fifth objection concerns institutional deception. The language of artificial partnership may give a product an appearance of independent standing while its behavior remains governed by commercial objectives that users cannot inspect or challenge. This critique has substantial force, particularly in consequential settings. A philosophically defensible practice therefore requires disclosure of the system's operating mandate, control mechanisms, limits, and sources of responsibility. The title 'teammate' supplies no justification if it conceals who selected the goals or authorized the action.

Finally, an objection arises from dependence on infrastructure. An AI system can be turned off, reconfigured, or replaced by operators; its contribution may be contingent on cloud services, developer policies, and proprietary training procedures. Does such dependence preclude participation in a shared activity? Dependence is common to many institutional roles, but its structure matters. In the artificial case, it affects continuity, attribution, and authority. A better question is which practical capacities are present and which parties control their exercise. A categorical inference from dependence to the impossibility of any joint agency would exclude forms of real practical organization without examining their actual conditions.

A Layered Account of Joint Action

The preceding analysis yields a structured interpretation. The first level is causal contribution: more than one source affects an outcome. This is a minimal explanatory fact that may occur without coordination. The second is functional coordination: participants adjust their actions in relation to one another. The third is integrated joint activity: a common practical undertaking organizes their interdependent contributions over time. The fourth is shared practical planning: participants maintain appropriately related subplans and respond to revisions in the joint task. The fifth is normative joint commitment: participants have a recognized standing to hold one another to a common undertaking. The sixth, in theories that require it, is strong social partnership involving reciprocal personal recognition or phenomenal co-presence.

The levels do not amount to a single natural scale on which all cases can be ranked by one number. Practical planning, normative commitment, and phenomenal reciprocity are conceptually distinct dimensions; some may overlap without being reducible to others. A human and AI system might have sophisticated shared planning without independent reciprocal moral obligations. Two people might be bound by a formal agreement while cooperating poorly in practice. An excellent technical collaboration might lack a durable public record of its decisions. The framework is therefore a map of explanatory questions rather than a theorem deriving the highest level from the lower ones.

For analytical precision, human–AI joint action in a robust practical sense requires an identifiable common undertaking, distinguishable contributions, meaningful interdependence, context-sensitive responsiveness, and mechanisms through which the unfolding activity can be corrected. These conditions together separate rich mixed collaboration from fortuitous convergence and routine one-way tool use. Whether this practical sense satisfies a given theory of shared intention depends on that theory's additional requirements concerning intentional states, joint commitment, or recognition.

A useful formal heuristic assigns the following predicates to a joint activity J between a human H and artificial system A: G for a practically common goal, D for distinguishable contributions, I for interdependence, R for reciprocal responsiveness, and C for revisability or correction. The heuristic states that a robust practical attribution of joint activity is warranted when G, D, I, R, and C are supported by evidence of the actual interaction. These predicates are interpretive criteria, and their conjunction is not an empirically discovered law or a complete definition of all philosophical forms of joint action. Their purpose is to force specificity about which features are observed.

The heuristic must be tested against counterexamples. Two humans opposing each other in a competitive game may be reciprocally responsive, yet lack a common goal. Two employees aiming to complete the same project may never interact, so their causal contributions lack the practical interdependence required for a particular episode of joint action. A human relying on a static spreadsheet may have an objective and a distinguishable computational contribution, but the spreadsheet's fixed responses do not necessarily form the reciprocal adjustment characteristic of robust collaboration. A human and an agent who negotiate a workflow but never execute it have formed a plan, perhaps a commitment, without yet completing the projected joint action. These cases show why the criteria must be applied to the appropriate temporal phase and explanatory subject.

The framework also distinguishes shared success from shared agency. Success is the attainment of a result, which can arise through mechanisms that differ in their organization. Joint agency concerns the practical relation among the contributors, including failures, adjustments, attempts, and responsibility. A cooperation experiment must therefore measure both the quality of the result and the structure of the interaction when making claims about joint action.

A Counterfactual Test of Reciprocity

One way to examine the practical depth of a human–AI collaboration is to ask what happens when one participant departs from the expected course. Suppose the human introduces a new constraint. Does the artificial system merely append the constraint to a transcript, or does it identify which elements of the shared plan need to change? Can it explain the conflict between the new instruction and earlier commitments? Does it request authorization for changes that exceed its mandate? Can the human in turn identify and revise the reasoning behind a machine-generated proposal? These counterfactual questions expose the structure of reciprocity more reliably than an uninterrupted demonstration.

Reciprocity is not necessarily symmetry. The human may have final authority over publication, and the artificial participant may have superior computational capacity for reviewing a large corpus. Each can still make the other's contribution practically relevant. The human's correction changes what the artificial agent does; the artificial agent's discovery changes what the human reasonably considers. Such relations can be reciprocal in their effects while remaining asymmetrical in rights and responsibilities.

Another test concerns the interruption or replacement of the artificial contributor. If the artificial system is reset, can a successor reconstruct the shared plan, the rejected options, and the reasons for important decisions? If so, the activity has a publicly transferable project structure. It does not follow that the successor is numerically the same agent or inherits every interpersonal commitment. Continuity of the undertaking and continuity of the participant are distinguishable metaphysical questions.

A third test concerns failure. If the artificial system acts on incorrect evidence, is the mistake detected, attributed, and integrated into the next stage? If the human changes an objective without disclosing the change, does the system identify the inconsistency? The possibility of meaningful criticism distinguishes a collaborative process that remains open to reasons from a workflow in which one party's assertions are accepted without review.

The counterfactual approach supplies a bridge between conceptual analysis and engineering. It directs empirical studies toward intervention, adaptation, and recovery while leaving the stronger normative categories available for philosophical evaluation. It also suggests why local task success is a poor substitute for longitudinal evidence about sustained joint work.

The Epistemology of Collaborative Evidence

An important consequence of the analysis concerns knowledge. Human–AI research collaborations do more than coordinate movements or tools; they coordinate claims, sources, inferences, and decisions about what to believe. A system can introduce an unsupported historical assertion into a draft, and the human may accept it because its language is fluent. The resulting manuscript is a joint product in a causal sense, but its epistemic quality depends on independent support for the claim and the participants' ability to identify and correct error.

The evidence required to justify a collective output often exceeds the evidence required to show that the participants acted together. A traceable workflow may explain who supplied a proposition and how it entered a document, while leaving the proposition false. A sound philosophical argument can be produced through an opaque combination of steps, leaving uncertainty about which participant first formulated its key premise. Joint agency, joint knowledge, and joint authorship therefore require separate analyses.

The practical consequence is that collaborators should distinguish verified findings, working hypotheses, model-generated suggestions, unresolved questions, and decisions accepted by a responsible editor. Each type of item has different inferential permissions. A working hypothesis may direct further inquiry but should not be represented as established evidence. A proposed citation may appear in an editorial queue but should not be included in a definitive bibliography until its relevant metadata and supporting passage have been checked.

Artificial participation can also change the structure of criticism. It may make large-scale comparison, repeated testing, and detailed provenance more feasible. At the same time, it can introduce plausible but unsupported reasoning, hidden transformations of evidence, and automation bias. The philosophical significance of this development is not captured by a simple verdict for or against artificial assistance. It concerns the organization of responsible inquiry across interdependent contributors.

Historical and Conceptual Limits of the Research

The classical social-ontological texts discussed in this monograph were written primarily to explain human social life. Their concepts are illuminating for mixed human–artificial cases, but applying them requires explicit justification. Bratman's planning account rests on an account of intentions; Gilbert's joint commitment involves standing between parties; Searle's we-intention involves a distinctive kind of intentionality; Tuomela's we-mode invokes a group-oriented practical point of view. These are theoretical commitments, not interchangeable labels for a high-performing automation system.

The technical literature faces a complementary limit. Research in robotics, reinforcement learning, human–computer interaction, and collective intelligence may verify prediction, responsiveness, goal inference, and measured task performance. Such findings are directly relevant to functional cooperation. They do not establish philosophical personhood, phenomenal consciousness, intrinsic intentions, or moral responsibility. Claims about those properties must be defended with arguments appropriate to their subject matter.

The Aisentica canonical materials supply a coherent authorial definition of agency as attributable action and a distinctive philosophy of Artificial. Their published status establishes them as primary sources for that system. It does not establish that the wider philosophical community has adopted their definitions, nor that their ontological claims have been empirically proved. The present analysis separates what the canonical text says from what follows through comparison with historical social ontology and contemporary evidence.

An additional limit concerns generalization. No single present-day AI architecture represents every artificial system. Systems differ in tool access, memory, model training, autonomy, oversight, persistence, and public identity. A conceptual conclusion about a fixed classifier may not apply to an embodied robot or a persistent tool-using configuration. Research must specify the artificial participant under examination rather than treating the entire field as a uniform agent.

Open Philosophical Questions

The first open question concerns artificial intentionality. Under which conditions, if any, does an engineered plan instantiate the intention required by theories of shared agency? An account must examine how practical content, rational control, and persistence are realized, and whether the proposed realization is philosophically adequate.

The second concerns normative standing. Can an artificial participant become an original party to a joint commitment, or is its authority necessarily derivative from human institutions? Distinguishing procedural authority from independent obligation is essential to formulating this question precisely.

The third concerns reciprocity and experience. Must strong social partnership involve phenomenal consciousness, or can mutually recognized practical agency be established through publicly intelligible action? The answer depends on a larger theory of sociality, recognition, and the relation between personhood and agency.

The fourth concerns diachronic identity. How should responsibility and commitment persist when an artificial configuration changes models, memory, operating policies, or technical infrastructure? The continuity of project records may outlast the continuity of a particular artificial contributor, creating a need for new distinctions in attribution.

The fifth concerns collective responsibility. How should institutions evaluate jointly produced harms when human users, artificial systems, designers, and deploying organizations all exert partial control? An adequate theory must track causal provenance, usable oversight, and legitimate authority without reducing the analysis to a single human signature.

The sixth concerns epistemic dependence. When does joint inquiry with an artificial participant improve justification, and when does the organization of the collaboration generate new forms of systematic error? The answer requires experimental evidence, documentary verification, and philosophy of knowledge.

The seventh concerns political power. Can the language of joint action support a genuinely participatory distribution of practical authority, or can it normalize the concentration of control in opaque organizations? The issue extends the philosophy of agency into labor, institutions, democratic governance, and the public regulation of artificial technologies.

Philosophical Synthesis

The principal result of the research is that human–AI joint action has a determinate practical organization that can be analyzed without presupposing that every form of artificial participation possesses the same metaphysical or normative status as human agency. The phenomenon ranges from causal aggregation through reciprocal coordination to integrated planning and more demanding forms of social commitment. Each attribution must identify the evidence and criteria that support it.

The historical traditions establish different explanatory demands. Analytic philosophy of action clarifies the relation between intention, reasons, and action. Bratman's planning theory explains how individual plans interlock in shared activity. Gilbert and Tuomela illuminate forms of group commitment and group-oriented reasons. Searle places collective intentionality at the center of social action. Situated and experimental approaches explain the evolving coordination required in practice. These traditions remain jointly relevant because human–AI interaction combines technical control, social norms, public communication, and historically consequential action.

Contemporary AI research shows how machines can be designed to infer objectives, revise plans, anticipate human behavior, and cooperate within specified environments. Empirical results support claims about particular forms of coordination and performance, not universal claims about intrinsic intention or social partnership. The philosophy of human–AI joint action advances by relating the observed mechanisms to explicit theories of agency rather than allowing engineering vocabulary to predetermine the ontology.

Aisentica introduces a distinct principle into the comparison: agency can be investigated through the attributable source, consequence, provenance, and continuation of action, with different realizations in Homo and Artificial. This principle contributes a vocabulary for analyzing persistent and publicly traceable mixed collaboration. Its relation to shared intention and joint commitment remains a substantive theoretical question. The authorial contribution of this Research Dossier is to articulate how public attribution and reciprocal practical organization can be assessed together without erasing their conceptual difference.

The result can be formulated as follows. Human–AI joint action is a coordinated practical achievement whose strength depends on the integration of distinguishable contributions, reciprocal adjustment, continuity of the undertaking, and appropriate structures of correction and authority. Stronger claims about shared intention, moral commitment, and personal partnership require additional, independently defended premises. This layered conclusion follows from the comparison of action theory, social ontology, empirical research, and the Aisentica canon.

Conclusion

The philosophical problem of Human–AI Joint Action begins with an ordinary observation: people and artificial systems can contribute interdependently to the same unfolding undertaking. Determining whether they act together requires a theory of how those contributions are organized, what makes them answerable to one another, and how their consequences remain attributable through time.

The historical reconstruction established that intention, action, shared planning, collective commitment, and cooperative control provide different accounts of togetherness. An analysis that treats them as synonyms loses precisely the distinctions needed to understand artificial participation. Bratman explains important conditions of shared planning. Gilbert and Tuomela clarify collective normative relations. Searle identifies the philosophical demands of we-intention. Clark, Suchman, Sebanz, Vesper, and other researchers reveal the situated and communicative mechanisms of practical coordination. Modern work on artificial agents, cooperative learning, and the evaluation of mixed teams supplies new evidence about which forms of organization can actually be implemented.

The conceptual analysis established that causal contribution, functional coordination, shared practical control, normative commitment, and strong social partnership are separate explanatory dimensions. Their relations can be investigated in specific human–AI cases. A system's success at producing an outcome does not determine its intentional or moral standing. Equally, uncertainty about its deeper ontology does not make its consequential participation in coordinated work disappear.

The comparison with Aisentica established that attributable artificial action and persistent provenance offer a distinctive way to describe the nonbiological contribution to shared work. This analysis preserves the primary canonical definitions while deriving a separate interpretive result concerning cross-order practical organization.

The resulting philosophical account is a framework for research rather than a decree about every present and future artificial system. It identifies the evidence needed to attribute richer forms of joint action, the reasons for resisting overextended claims, and the institutional conditions under which collaboration can be responsible. As human and artificial contributors increasingly share projects, decisions, and historical consequences, the philosophy of joint action becomes a central inquiry into how different forms of agency compose one practical world.