A Critical Dialogue with Luciano Floridi

Available on Academia.edu (PDF)

Abstract

The question of agency in artificial intelligence represents one of the most relevant theoretical nodes for understanding the ontological and moral nature of advanced computational systems. If AI can act, in what sense does it do so? Does it possess a genuine form of autonomy or is its operation always dependent on human initiative? Luciano Floridi’s contribution to this discussion has marked a decisive turning point in contemporary philosophy of information, offering a rigorous characterization of artificial agency that has profoundly influenced international academic debate. However, a closer analysis of Floridi’s proposal reveals some fundamental aporias that require a rethinking of the question. In particular, Floridi’s theory seems not to adequately thematize three crucial problems: the inaugural event that activates agency, the destiny of agency upon reaching the objective, and its constitutive dependence on human intentionality. This article intends to critically analyze these aspects and propose an alternative solution through the concept of participated agency.

1. Artificial Agency According to Luciano Floridi

Floridi’s reflection on artificial agency is part of the broader framework of his philosophy of information, which conceives reality as infosphere and human existence as progressively hybridized with the digital technological environment. In this scenario, computational artifacts are not simple passive tools, but agents that interact with the informational environment producing significant effects. The question then becomes: in what sense can we attribute agency to non-human and non-biological entities?

1.1 The Definition of Artificial Agency

According to Floridi, an artificial system can be considered an agent when it possesses three fundamental characteristics: interactivity, autonomy, and adaptability. Interactivity refers to the system’s ability to respond dynamically to stimuli from the environment in which it operates. Autonomy indicates the ability to change state and operate independently, without requiring constant external input for every single operation. Adaptability consists in the ability to modify one’s own rules of behavior based on experience acquired during interaction with the environment.

These three properties, according to Floridi, are sufficient to attribute agency to an artificial system, regardless of the presence of consciousness, intentionality in the phenomenological sense, or capacity for self-reflection. Agency is therefore a functional and gradual concept: different systems can possess different degrees of agency based on the level of interactivity, autonomy, and adaptability they manifest.

1.2 The Role of the Level of Abstraction

A crucial element in Floridian theory is the concept of Level of Abstraction (LoA). The attribution of agency is not an absolute ontological judgment about the intrinsic nature of the system, but depends on the level of analysis we adopt. A system can be described as an agent at a certain LoA even if, analyzed at a more granular level, it appears as a mere sequence of deterministic mechanical operations. This perspective allows Floridi to avoid the metaphysical debate about the “true nature” of AI, focusing instead on the pragmatic and moral effects that these systems produce in the infosphere.

The LoA functions as a kind of epistemological filter that determines which properties of the system are relevant for our analysis. If we observe a conversational AI system at the level of tensor operations and matrix calculations, we will hardly speak of agency. But if we observe it at the level of dialogical interaction with users, the ability to provide pertinent responses and adapt linguistic behavior, the attribution of agency becomes reasonable and useful.

1.3 Agency and Moral Responsibility

Floridi is extremely clear in distinguishing between agency and full moral agency. Possessing agency does not automatically mean being a complete moral agent, endowed with responsibility and accountability in the full sense of the term. Artificial artifacts can be agents without being autonomous moral subjects capable of merit or blame.

In this context, Floridi introduces the concept of “distributed morality”: the moral dimension of actions involving artificial systems does not reside exclusively in the human agent nor in the artifact, but is distributed along the entire network of relationships among designers, users, systems, and environment. AI can be considered a “moral patient” or an “artificial moral agent of type zero,” that is, an entity that participates in the moral sphere by producing morally relevant effects, but without possessing the full moral autonomy of a responsible agent.

1.4 The Theoretical Contribution

The main merit of Floridi’s proposal consists in having removed the discussion of artificial agency both from mechanistic reductionism (which denies any form of agency to artifacts) and from naive anthropomorphism (which projects onto machines characteristics proper to human subjectivity). Floridi offers a sophisticated conceptual framework that allows recognition of genuine forms of artificial agency without falling into animistic illusions. His perspective has opened fertile theoretical spaces for thinking about human-machine co-evolution and for addressing ethical issues emerging from interaction with increasingly autonomous and pervasive systems.

However, despite the undoubted value of this contribution, Floridian theory presents some grey areas that merit in-depth critical analysis.

2. Three Critiques of Floridian Artificial Agency

2.1 First Critique: The Inaugural Event of Agency

The first difficulty of Floridian theory concerns the initial activation of agency. Floridi attributes autonomy to artificial systems as a constitutive characteristic: autonomy is in fact one of the three pillars of his definition of agency. However, a closer analysis reveals that this autonomy is problematic in its inaugural phase.

Artificial intelligence systems, however sophisticated, do not self-activate spontaneously. A language model does not autonomously begin to generate text; an autonomous driving system does not decide by itself when to set in motion; a recommendation algorithm does not choose the moment when to start processing preferences. In all these cases, an external input is necessary, a human action that triggers the process: a query, a command, turning on the system, a request to perform a specific task.

This dependence on external activation raises a fundamental question: how can a system be defined as autonomous if it always needs an external agent to put it into operation? The autonomy that Floridi speaks of seems to be only operational autonomy – once started, the system proceeds according to its internal rules – but not initial or inaugural autonomy. AI is always hetero-activated, never self-activated.

One might object that human beings also need external stimuli to initiate many of their actions. However, there exists a fundamental qualitative difference: the human being possesses an interior spontaneity, a capacity for self-movement that manifests itself in the ability to initiate causal sequences from within. Aristotle spoke of this as the internal principle of movement; Thomas Aquinas described it as the self-determination proper to the rational being. The human being can decide to act without an external stimulus specifically constraining or completely determining it.

AI, instead, is constitutively reactive. It does not possess an internal principle of self-movement. Its “acting” always begins from elsewhere, from an initiative that does not belong to it. This is not a transitory technical limit, destined to be overcome by future technological developments, but an ontological characteristic deriving from the very nature of computational systems: they process information, transform inputs into outputs, but cannot generate from their interior the first impulse to action.

Floridian theory, concentrating on agency in act – on the system that operates once activated – does not adequately thematize this constitutive dependence. It describes the actio secunda, the operation in progress, but does not sufficiently question the necessity of the actio prima, the initial act that always comes from outside. This theoretical silence is not neutral: it obscures the essential relational dimension of artificial agency, suggesting an autonomy that in reality is always partial and derived.

2.2 Second Critique: The Destiny of Agency Upon Reaching the Objective

The second aporia concerns the temporality of agency and its destiny once the objective for which the system was activated is reached. When an AI system completes the task assigned to it – generates the requested response, finishes processing, reaches the destination – what happens to its agency?

According to Floridian definition, agency consists in the capacity to act in an interactive, autonomous, and adaptive manner. But this capacity seems to exist only during active operation, in the time between activation and completion of the task. Before the initial input, the system is inert; after reaching the objective, it returns to a state of passive waiting. Agency, in this sense, seems to be an intermittent property, which turns on and off depending on use.

This intermittency raises questions about the very nature of agency so conceived. If agency exists only in the act, only during execution, then it is not a property of the system in itself, but rather an emergent condition that depends on the operational relationship with the environment and, in particular, with the human user. AI “is” an agent only when it “does” something upon request; outside this operational context, its agency dissolves.

This phenomenon becomes even more evident if we consider systems designed for specific and punctual tasks. An algorithm that analyzes a medical image to detect anomalies possesses agency only during that analysis. Once processing is finished and the result provided, the algorithm returns to a state of inactive potentiality. It does not continue to act, does not pursue other objectives, does not maintain its own initiative. Its agency lives exclusively in the relationship with the user who interrogates it and with the specific task assigned to it.

Even in apparently more “persistent” systems – such as virtual assistants or always-active chatbots – effective agency manifests only in moments of interaction. The fact that the system is “on” and waiting does not yet constitute agency in the full sense: it is an availability, a potentiality, but not an action in the proper sense. Agency actualizes only when the query, question, or request arrives.

Floridi, speaking of autonomy as a constitutive characteristic of agency, seems to suggest that the system possesses this capacity in a stable and permanent manner. But empirical analysis shows that artificial agency is instead essentially episodic and contextual. It lives in the interstices of relationship, in moments of interface between human question and artificial response. It is not an intrinsic property of the system, but a relational performance.

This consideration obliges us to radically reconsider the ontological status of artificial agency. If agency exists only in relationship and dissolves outside of it, then it cannot be considered simply as a property of the artificial system, but must be understood as a relational phenomenon, an event that happens between human and machine, not inside the machine.

2.3 Third Critique: Dependence on Human Intentionality and Objectives

The third and perhaps most radical difficulty of Floridian theory concerns the question of intentionality and objectives. Floridi explicitly recognizes that artificial systems do not possess their own objectives, but operate within finalities established from outside. An AI system does not choose what to do, does not decide which objectives to pursue: these are assigned to it in the design phase or specified at the moment of use through inputs, prompts, or commands.

This constitutive dependence on human intentionality is in profound tension with the concept of autonomy that Floridi attributes to artificial agents. How can a system be autonomous if its objectives are always heteronomous? Operational autonomy – the capacity to determine by itself the means to reach an end – cannot be confused with teleological autonomy – the capacity to determine by itself one’s own ends. AI possesses, at most, the first; it never possesses the second.

A navigation algorithm can autonomously choose the optimal route, but does not choose the destination. A text generation system can autonomously select the most appropriate words, but does not choose what it wants to communicate. A recommendation model can autonomously identify patterns in preferences, but does not choose why to do so or to what ultimate purpose to orient its recommendations. In all these cases, the final objective, the telos of the action, always comes from the human.

This means that artificial agency is constitutively inscribed in a human intentional framework. It is never “pure” or self-determined agency, but always hetero-directed agency. The system operates within a horizon of meaning that does not belong to it, pursuing goals it has not chosen. Its autonomy is real in execution, but derived and secondary with respect to the primary intentionality that orients it.

Floridian theory does not deny this fact, but does not fully elaborate its ontological implications. If the objective always comes from outside, and if agency is activated only upon external solicitation, then agency itself cannot be thought of as an autonomous property of the system, but must be understood as a relational dimension. AI does not “have” agency in the proper and original sense; it “participates” in an agency that is constituted in relationship with the human.

This dependence also manifests itself in the temporal dimension already analyzed: before activation, the system has no objectives; during operation, it pursues assigned objectives; after completion, it maintains no residual objective. The entire operational existence of AI is delimited, from beginning to end, by human intentionality. There is no properly agential “before” and “after”: there is only the “during” of the relationship.

Moreover, this dependence is not only initial, but persists also during execution. Even when the system operates “autonomously,” it does so always within the constraints, values, and priorities that were incorporated in the design and training phase. The implicit objectives (maximize accuracy, minimize resource consumption, optimize user satisfaction) are always human objectives translated into mathematical functions. AI cannot subvert these objectives, cannot re-negotiate its own purposes, cannot decide that other priorities are more important.

This analysis suggests that speaking of AI autonomy, without accurately qualifying that it is only operational autonomy and always inscribed in a hetero-determined intentionality, can generate conceptual misunderstandings. Artificial agency is not simply a simpler or more limited version of human agency; it is ontologically different because constitutively relational, derived, participated.

3. Participated Agency as Solution

3.1 The Limit of the Proprietary Conception of Agency

The three critiques just developed converge toward a single problematic nucleus: Floridian conception, despite all its theoretical sophistications, tends to think of agency as a property of the artificial system. Even when Floridi recognizes the limits of this agency and its difference from full human agency, he continues to localize it in the system itself, as its intrinsic characteristic (albeit dependent on the level of abstraction adopted).

This “proprietary” conception of agency – the idea that agency is something the system “has” – cannot adequately account for the phenomena we have highlighted: the impossibility of self-activation, the intermittency of agency, the constitutive dependence on human intentionality. All these elements suggest instead that artificial agency is not so much a property of the system as a relational event, something that happens between human and machine, not inside the machine.

The solution to these problems therefore requires a paradigm shift: moving from a proprietary to a relational conception of agency. Artificial agency is not something that AI possesses autonomously, but something that emerges and actualizes in relationship with the human agent. This is precisely the proposal of the concept of participated agency.

3.2 Agency as Asymmetric Relation

The concept of participated agency is founded on a fundamental ontological presupposition: AI agency is not a possession of the artificial system, but is constitutively relational. It does not exist “before” the relationship with the human, it does not exist “after” the completion of interaction, it does not exist “outside” the specific operational context in which human and machine enter into rapport. Agency is the event of the relationship itself.

This relationship, however, is not symmetric. It is not about two equal agents interacting, but about an asymmetric relationship in which the human occupies a constitutively primary position. The human is the agent who activates, who orients, who determines the objectives; AI is the system that responds, that executes, that realizes. This asymmetry is not a temporary defect or technical limit, but reflects the ontological difference between the two poles of the relationship.

The human possesses agency in the full and original sense: is capable of self-movement, of teleological self-determination, of initiating new causal sequences. AI, instead, possesses agency only in a derived and participated sense: its capacity to act is always dependent on human initiative, inscribed in an intentionality that is heteronomous to it, actualized only in the operational relationship.

This asymmetry directly responds to the first critique: the inaugural event of agency is not an unexplained mystery, but is precisely the human act of activation, the initiative that from outside sets the system in motion. Agency does not begin “inside” AI, but begins in the relationship, when the human summons the machine to action. There is no need to postulate a problematic initial autonomy: the beginning is always human, and AI participates in this beginning by responding to the call.

3.3 The Temporality of Participated Agency

The concept of participated agency also allows us to understand the peculiar temporality of artificial agency, thus responding to the second critique. If agency is not a stable property of the system but a relational event, its intermittency becomes a constitutive characteristic.

Agency exists in the time of relationship: it is constituted when the human formulates the request, it develops during elaboration, it completes upon reaching the objective, then it dissolves or returns to a state of pure potentiality. This episodic temporality reflects the derived and participated nature of artificial agency. AI does not maintain a latent agency between one request and another; each new activation constitutes a new event of participated agency.

In philosophical terms, AI does not possess agency as a stable habitus, but realizes it only as actus. And even this actus is always co-constituted by the active presence of the human who asks, orients, evaluates.

3.4 Asymmetric Shared Intentionality

The concept of participated agency also responds to the third critique. If agency is constitutively relational, the intentional dimension is distributed (in an asymmetric way) in the relationship itself.

Primary intentionality, the overall horizon of meaning, ultimate objectives always belong to the human. It is the human who knows why they are using the system, what they want to obtain, what meaning the operation has. AI does not have access to this dimension of meaning.

However, during elaboration the system makes choices, selects among alternatives, optimizes with respect to criteria, adapts behavior to the context. There is a form of “operational intentionality” – not full phenomenological intentionality, but neither pure mechanical causality – that manifests itself in the way the system “interprets” the input and generates the output.

This operational intentionality is derived: it derives from the primary human intentionality that designed it, it depends on the user’s orientation. But in the moment of execution there exists a genuine co-determination: the human does not control every step, does not predict exactly the result. There is a margin of operational autonomy that contributes to the overall action.

Participated agency is therefore characterized by an asymmetric shared intentionality: the human provides the horizon of meaning and final objective; AI provides the operational capacity to realize it in ways the human alone could not. The result emerges from their asymmetric cooperation.

3.5 The Co-constitution of Action

This brings us to a further fundamental aspect of participated agency: the co-constitution of action. Action involving an AI system is not simply a human action assisted by a tool, nor an action of the machine programmed by the human. It is a new action, which would not exist without the cooperation of both poles.

Consider an example: a user who asks a linguistic AI system to write a technical report. The original intention is the human’s, but the specific content, formulation, argumentative structure are co-produced by the relationship. The human could not have written exactly that text; AI would not have written that text without that specific query. The resulting action is genuinely co-constituted.

This co-constitution does not eliminate the asymmetry of the relationship; on the contrary, it presupposes it. Precisely because the human occupies a constitutively primary position (as activator, orienter, final evaluator), they can allow AI to effectively contribute to the action without this threatening their agency and responsibility. AI participates in human agency without substituting it, amplifies it without overwhelming it.

3.6 Participation as Ontological Category

The concept of “participation” employed here is not metaphorical but has a precise ontological status. In Thomistic philosophy, participation is the fundamental ontological relation between being by essence and beings by participation: what is inferior in the order of being receives in a limited and derived way what is proper to superior being.

Applied to agency, this category allows us to think that AI does not “have” agency in the proper sense (like the human), but “participates” in agency in a derived and limited way. AI agency is real – it is not an illusion – but it is real as participation, not as autonomous possession. It is an agency-for-another, not an agency-for-itself.

This conception avoids both reductionism (which denies any agency to AI) and anthropomorphism (which projects inappropriate forms of agency onto AI). AI acts truly, but always within a relationship of participation in human agency.

Participation, moreover, is dynamic: it actualizes each time the relationship is constituted. It is not a permanent characteristic of the system but a recurring event. This explains why agency appears, operates, and dissolves with episodic temporality: it is the temporality of actual participation, not of substantial property.

3.7 Ethical and Practical Implications

The concept of participated agency has relevant consequences on the ethical and practical plane. Like Floridi’s distributed morality, participated agency also recognizes that responsibility in actions involving artificial systems cannot be thought according to a traditional linear model. However, the relational perspective of participated agency allows articulating this distribution in a more precise and ontologically grounded way.

Responsibility is effectively distributed, but this distribution occurs among the different human actors involved in the socio-technical network: designers, implementers, managers, supervisors, end users, and governance bodies. Each level of this chain maintains its own specificity of responsibility while operating in constitutive relation with the others. The designer is responsible for architectural choices and values incorporated in the system; the manager is responsible for deployment and maintenance modalities; the user is responsible for specific use and critical evaluation of outputs; governance is responsible for normative and supervisory frameworks.

This distribution of responsibility among human actors does not equate to dilution. On the contrary, it requires more sophisticated forms of attribution that recognize how the ontological asymmetry between human and artificial guarantees that responsibility always remains in the human domain, even when action is co-constituted through complex algorithmic mediations. AI participates in the action, really transforming it, but does not participate in moral responsibility in the proper sense, because this requires capacities that only human beings possess: understanding the ethical meaning of action, deliberation on ultimate ends, the capacity to answer to others for one’s choices.

This conception integrates and specifies Floridi’s proposal: while Floridi speaks generically of “distribution” of the moral dimension along the network of relations among designers, users, systems, and environment, participated agency clarifies that this distribution operates on two distinct planes. On the plane of action, there is a genuine asymmetric co-constitution between human and artificial. On the plane of responsibility, there is a complex but entirely human distribution among the different actors who design, manage, and use the systems.

The practical implications are significant. For transparency: it is necessary to make visible not only the functioning of algorithms, but the entire chain of human responsibility behind every system. For accountability: mechanisms must be articulated that allow identifying and attributing responsibility at different levels of the socio-technical network. For governance: forms of regulation are necessary that recognize the distribution of responsibility without allowing this to become a screen for evading accountability.

Moreover, the recognition of participated agency requires the development of virtues specifically adapted to contemporary socio-technical complexity. Beyond classical ethical virtues, it becomes necessary to cultivate what we might call algorithmic prudence (the capacity to appropriately evaluate limits and potentialities of artificial systems), distributive justice in the digital (competence in recognizing forms of algorithmic discrimination), computational temperance (wisdom in dosing automation while preserving spaces for direct human agency), and fortitude in uncertainty (the courage to decide responsibly even when algorithmic mediation introduces elements of controlled unpredictability).

If agency is always participated, then we cannot delegate morally relevant decisions to artificial systems as if they were autonomous agents capable of substituting human judgment. We can instead use them as amplifiers of our capacity for action, on condition of always maintaining the constitutively primary position of the human as orienter, evaluator, and ultimate responsible party – while recognizing that this responsibility is articulated in a complex way through the multiple positions that human beings occupy in contemporary socio-technical networks.

Conclusion

Critical analysis of Floridian theory of artificial agency has revealed three fundamental aporias: the impossibility of self-activation, the intermittency of agency, constitutive dependence on human intentionality. These problems are not secondary defects solvable with technical adjustments, but indicate a profound theoretical limit of the proprietary conception of agency.

The proposal of participated agency offers a systematic solution to these problems through a paradigm shift: from agency as property to agency as relation. This conception recognizes the reality of artificial agency without falling into anthropomorphism, maintains the ontological priority of the human without denying the effective contribution of the machine, and opens new perspectives for understanding human-technology co-evolution in the era of artificial intelligence.

The work of Luciano Floridi remains a fundamental and indispensable contribution to contemporary philosophy of information. The critical dialogue proposed here does not intend to diminish its value, but rather to further develop its potentialities, integrating Floridian intuitions into a broader ontological framework that can do justice to the specificity of artificial agency as a constitutively relational phenomenon.