The Metaphysics of Machine Agency: Beyond Tool and Entity

The question of machine agency is not merely a technical inquiry into the capabilities of artificial systems; it is a foundational metaphysical and ethical puzzle that determines how we allocate responsibility, design governance structures, and ultimately coexist with entities that increasingly shape our world. The dominant discourse oscillates between two poles: the view of AI as a mere instrument, a sophisticated tool whose actions are always and only the actions of its human creators and operators, and the specter of AI as a full-fledged moral person, an entity possessing intrinsic rights and duties. This chapter argues that both poles are philosophically incoherent and practically dangerous. They represent an essentialist trap—searching for a single, binary property like "consciousness" or "real understanding" to demarcate the moral community. Instead, we must adopt a pragmatic ontology of machine agency. Agency is not a thing an entity *has*, but a *role* it *plays* within a socio-technical system, defined by its functional capacities, the norms governing its operation, and the social practices that attribute actions and responsibilities to it. This reframing dissolves the binary debate and provides the necessary toolkit for the ethical and legal challenges that follow, beginning with a critical deconstruction of the pervasive "human-in-the-loop" fallacy.

I. The Classical Architecture of Agency: Intentionality, Causality, Normativity

Western philosophical tradition has erected a formidable edifice around the concept of agency, typically conditioning it upon a triad of interlocking properties: intentionality, causality, and normativity. To act as an agent, in this view, is not merely to cause an event in the physical world (causality), but to do so *for reasons* (intentionality), where those reasons are assessed against a framework of values, rules, or oughts (normativity). This architecture, from Aristotle's *praxis* to Kant's rational will to contemporary analytic philosophy, is fundamentally anthropocentric, taking human rational agency as its paradigm.

Intentionality, following Brentano and later Searle, is the "aboutness" of mental states. A human agent acts *with* an intention; their beliefs and desires *represent* states of the world and *motivate* action. The key is the *semantic content* of these mental states—they are about something, and they have conditions of satisfaction (they can be true or false, fulfilled or not). For Searle, this biological, first-person, subjective feature is the "causal basis" of human agency and is absent in computational systems, which manipulate syntax without semantics. A thermostat causally influences a heater, but it does not *intend* to maintain a temperature; its state is not *about* anything in a representational sense.

Causality is the brute physical linkage. An agent must be an efficient cause, an entity whose activity brings about change. This is the least discriminating criterion; rocks, rivers, and robots all cause events. The philosophical interest lies in *agential causality*—a special subset where the cause is itself an entity with goals, plans, or dispositions. Aristotle's *archē* (originative cause) points to this: the agent is the source of motion in a way that a mere event is not.

Normativity introduces the dimension of evaluation. Agents are subject to norms of *rationality* (their beliefs and actions should cohere), *morality* (their actions can be right or wrong), and *legality* (their actions can comply or conflict with rules). This is the "space of reasons" (Sellars) or the "Kingdom of Ends" (Kant). To be a full moral agent is to be capable of grasping and acting upon norms, to be answerable for one's actions, and to be susceptible to attitudes like blame, praise, and guilt. This capacity is typically linked to practical reason and autonomy—the ability to legislate norms for oneself.

This classical model sets an extraordinarily high bar. It consigns all non-human animals, and certainly all machines, to the realm of "patients" rather than full "agents." They may be *acted upon* by norms (e.g., animal welfare laws), but they cannot be *bearers* of normative duties in the full sense. This model has been the implicit backdrop for much AI ethics, leading to the conclusion that AI systems are, at best, *instruments* and at worst, *hazards*. Responsibility for their actions must, by logical necessity, trace back to the human designers, programmers, owners, or users.

II. The Cracks in the Foundation: Large Language Models and Autonomous Systems

The classical architecture, however, faces severe stress when applied to contemporary AI, particularly Large Language Models (LLMs) and advanced autonomous systems. These systems exhibit behaviors that resist easy categorization within the tool/entity binary, forcing a confrontation with the adequacy of our philosophical criteria.

Large Language Models and the Illusion of Intentionality. An LLM like GPT-4 generates text that is contextually appropriate, informative, and often creatively surprising. It can answer questions, write essays, and even simulate a philosophical dialogue about its own lack of consciousness. To the classical view, this is a spectacular, but ultimately empty, syntactic manipulation. There is no "someone" behind the output; no beliefs, desires, or intentions. The system optimizes for statistical next-token prediction. Its "aboutness" is parasitic on the aboutness of its training data and the prompt engineering of its user; it has no independent representational content. Searle's Chinese Room argument seems to apply perfectly.

Yet, this picture is incomplete. From a functionalist perspective, what matters is not the substrate but the role played. An LLM *functions* as a *source of information*, a *collaborator in writing*, or even a *simulator of conversation*. Within a specific socio-technical context—a customer service chatbot, a coding assistant, a therapeutic dialog system—it occupies a role that *demands* a certain kind of interpretive stance from the human user. The user must *treat* the system *as if* it had intentions, if they are to successfully interact with it. This "as-if" is not a mere illusion; it is a constitutive feature of the interaction. The system's output has *derivative intentionality*—it inherits meaning from the human interpretive community that uses it, much like a book or a painting does. The classical model's focus on *original* (first-person) intentionality misses this crucial, socially constructed layer of meaning that enables LLMs to function as *quasi-agents* in communication.

Furthermore, LLMs demonstrate a form of norm-responsive behavior. They are trained on vast corpora of human language that embody social norms, ethical guidelines, and stylistic conventions. Their outputs are shaped by reinforcement learning from human feedback (RLHF), explicitly tuning them to be "helpful," "harmless," and "honest." They are not *aware* of norms, but they are *sensitive* to them in a way that a simple tool is not. A hammer is indifferent to whether it is used to build a house or break a window; a fine-tuned LLM will resist generating hate speech but comply with a request for a poem. This sensitivity, albeit derivative and brittle, complicates the tool metaphor. It suggests a kind of *externalized normativity*, where the system's behavior is *governed by* norms, even if it cannot *comprehend* them.

Autonomous Systems and the Spectrum of Causal Agency. Consider a Level 4 autonomous vehicle navigating a complex urban environment. It perceives, plans, and acts without real-time human intervention. Its causality is direct and robust. But does it act *with intention*? Its "decision" to change lanes or apply brakes is the output of a deep neural network processing sensor data. There is no "choice" in the human sense, only a high-dimensional mapping from in