Contribution
The 2023 editorial (LIT-896, NOTE-697) asserted that LLMs are agents without intelligence and gave no criteria. This paper supplies two things. It names the thesis: the Multiple Realisability of Agency (MRA), set against the Artificial Realisability of Intelligence (ARI). And it gives a taxonomy of kinds of agency (natural, biological, animal social, artefactual, human individual, human social, artificial, social artificial) organised by Floridi & Sanders’ three criteria (LIT-898) and placed under the Method of Abstraction (LIT-900). What is new after it is a single citable statement of where Floridi puts AI among agents: “a computational, goal-driven form of agency defined by human purposes” (p. 18), with Agentic AI as its collective form (§13). The argument for MRA over ARI, which the title and the outline promise, is not in it.
Key insight
Stop asking whether AI is intelligent, and ask what kind of agent it is. Agency comes in kinds that differ by mechanism and origin. Physical law gives natural agency, evolution gives biological agency, design gives artefactual agency, and design together with statistical learning gives artificial agency. Interactivity is the one criterion every kind shares: “without some form of interaction, there can be no agency” (p. 8). On this picture intelligence, consciousness and mental states mark one kind of agency, the human kind. They are not conditions of agency as such, so an AI can lack all three and still be an agent. Floridi frames the project as a test he could fail: “if one cannot convincingly define the kind of agency AI represents, it becomes difficult to argue that it possesses agency at all” (p. 2).
Assumptions
- AI is an agent. This is assumed, not shown. “The central question is not whether AI can be considered a form of agency, but rather, assuming it is, what its nature is. For readers already persuaded by this premise, further justification is unnecessary” (p. 2). The paper’s stated aim is narrower than its abstract-like outline. It “addresses the objection that unintelligent agency, as a concept, is nonsensical” (p. 2).
- AI lacks intelligence, consciousness, understanding and mental states. This is asserted at each mention (pp. 18, 19, 20, 24) and not argued. “The absence of consciousness, intelligence, and understanding creates an unbridgeable gap between artificial and biological-human agency” (p. 19). The 2023 editorial’s support for it, anecdotal failures, is not repeated, and nothing replaces it.
- No definition of agency. “I do not attempt to define agency formally, in terms of necessary and jointly sufficient conditions” (p. 3). Agency, like intelligence, democracy and love, is a cluster concept, identified by “clusters of overlapping features” in a Wittgensteinian sense (pp. 3–4).
- An adverbial theory of abstract nouns. “‘intelligence’ should be understood as a shorthand for ‘behaving intelligently’ … not for something that a system has or fails to have” (p. 3, after Chisholm, fn. 1).
- The LoA is the analyst’s, used “in a Kantian sense”. It is “focusing on the ontology of the model rather than the metaphysics of the system being modelled” (p. 5). Observables need not be behavioural, and “‘mental states’ can qualify as observables” (p. 5).
- Pluralism without relativism, by purpose. LoAs “can be assessed as better or worse depending on the specific purpose for which they are adopted” (p. 5). This is NOTE-700’s C8, unchanged.
Key results
There are no formal results. The positions taken, in order:
- The dilemma (pp. 1–2). Either widen intelligence to take in artificial kinds (ARI) or widen agency to take in kinds “that lack cognition, intelligence, intention, or mental states” (MRA). Floridi concedes the weak form of the first. There is “an uncontroversial, albeit limited and unremarkable, sense in which AI can be considered ‘intelligent’”, as Hobbes’s reckoning (p. 2). So ARI is the stronger claim of intelligence “equivalent to human intelligence, or even surpassing it” (p. 2). He defends the general MRA rather than an Artificial Realisability of Agency, because “it is more compelling to defend the stronger version” (p. 2).
- The legs of the argument (pp. 1, 2). “Scientific evidence, common sense, Ockham’s razor, and an increasing body of scholarly research favour the MRA thesis over the ARI thesis” (p. 1, outline; repeated p. 2). Of the four, only the scholarly research gets a section (§3). Ockham’s razor and common sense are not mentioned again. No evidence about AI is presented. See Claims.
- The LoA section (§2, pp. 4–5). A thermostat is described at four LoAs: physical, control-system, environmental and system-integration. “At the environmental level, the focus shifts to the temperature regulation agency” (p. 4). The one constraint stated for comparing systems is that “The key is maintaining a stable LoA while examining different systems, thus avoiding category mistakes” or metábasis (p. 5).
- The literature (§3, pp. 5–7). On the “standard view”, agency requires mental states that cause intentional action (Swanepoel, Sebastián). On the “non-standard view”, which Floridi endorses but will not call “functionalist” (p. 5), “any entity with some degree of interactivity may qualify as an agent, with more complex forms of agency including other features, such as autonomy, adaptability, and, eventually, internal mental states” (p. 6). Dung’s five dimensions, List on group agency, Laukyte, Symons & Abumusab on degrees, van Lier, Popa (“human goals remain constitutive”, p. 6) and Dattathrani & De’ are surveyed as “an emerging position, if not a consensus” (p. 7). None is argued with.
- The three criteria (§4, pp. 7–8), “Following (Floridi & Sanders, 2004)”. Interactivity: the agent “can act on the environment and, in turn, be acted upon by it”. It is “the foundational criterion for all forms of agency” (pp. 7–8). Autonomy: “an agent’s ability to initiate state changes independently of direct external causation”. It is graded, “from zero in some natural phenomena”, and “dependent on the level of abstraction” (p. 8). Adaptability: “the capacity of an agent to modify its behaviour based on input, such as changing physical conditions or acquiring data, information, or experience” (p. 8).
- The taxonomy (§§5–13). Set out in the table below.
- The comparison (§11, pp. 17–18). Kinds of agency form “a branching pattern”, “not … a strict linear hierarchy” (p. 17). “Human agency is the only agency that shapes all other agencies” (p. 17). Artificial agency blends “aspects of artefactual and biological agency while introducing novel properties” (p. 18).
- Artificial agency (§12, pp. 18–20). “At its core, AA is a computational, goal-driven form of agency defined by human purposes” (p. 18). It has “data-driven adaptability” through statistical learning, which is more open than the thermostat’s (p. 18). It cannot “generate truly original, chosen, or preferred purposes or goals, including the ability to choose whether to choose” (p. 19). Its understanding “remains firmly at pattern recognition and matching” (p. 19). It “could also be described as a syntactic form of agency” (p. 20). Five characteristics are listed: informational, instrumental, non-conscious, distributed and reproducible (p. 20).
- Agentic AI (§13, pp. 21–23). “an emergent form of collective agency, where autonomous, coordinated AI systems interact to achieve complex goals with minimal human oversight” (p. 21, citing Shavit et al. 2023). It has “a reasoning engine, often powered by LLMs” (p. 21). “Their interactions often produce emergent behaviours not explicitly anticipated by their designers” (p. 21). It “lacks consciousness or intentionality. However, it exhibits forms of collective intelligence” (p. 21). The example is a “hypothetical textbook” travel planner (p. 22).
- Governance (§§12.3, 13.3, 14). The list of recommendations is not argued. Deployment needs “a clear distinction between artificial and human agency”, and “governance frameworks must specifically address Artificial Agency” with “graduated autonomy systems that maintain appropriate human oversight” (pp. 20, 25). “Traditional models of control, moral accountability/responsibility, and legal liability are inadequate when dealing with emergent behaviours arising from complex AI interactions” (p. 23).
The Method-of-Abstraction comparison, as the paper gives it
The paper has no table. This one collects what each section says about the three criteria and what else it uses to mark the kind.
| kind | example | interactivity | autonomy | adaptability | what else marks the kind | pp. |
|---|---|---|---|---|---|---|
| natural | rivers | yes | none (“zero”) | none | physical law; “absence of intentionality or purpose” | 8–10 |
| biological (animal individual) | dogs | yes | ”degrees" | "degrees”, learning | evolution, homeostasis, “elements of intentionality” | 10–11 |
| animal social | ant colonies | yes | not stated | ”adaptability” | emergent coordination “without centralised control” and without rules or norms | 11–12 |
| artefactual | smart thermostat | yes | limited, by design | limited, by design | goals “explicitly defined through human intention” | 12–14 |
| human individual | none: “humans are the only example” | yes | ”unprecedented" | "unprecedented” | consciousness, intentionality, moral responsibility, duties, narrative identity | 14–15 |
| human social | corporations | yes | not stated | organisational learning | collective intentionality, institutions, rules; “legal persons” | 15–17 |
| artificial | LLMs | yes | ”bounded” | data-driven, statistical | computational; goals “defined by human purposes”; no intelligence or mental states | 18–20 |
| social artificial (Agentic AI) | hypothetical travel planner | yes | ”minimal human oversight" | "real-time” | networked modules, emergent behaviour, “collective intelligence” | 21–23 |
Two things follow, and the paper draws neither. First, the three criteria separate only natural agency from the rest. Every other kind has all three in some degree. Second, the work of telling the kinds apart is done by observables of other sorts. Origin is one: physics, evolution, design or learning. Substrate is another: biological or computational. Then there is where the goals come from (the agent itself, evolution, or a human), and whether mental states, consciousness and intelligence are present. The last are the observables §2 allowed in (p. 5), and they mark the line between artificial and human. No LoA listing these observables is written down, and their values for AI are asserted, not observed.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | MRA is better supported than ARI by scientific evidence, common sense, Ockham’s razor and the literature | assertion | named on pp. 1–2 and not argued. No evidence about AI is presented. Common sense and Ockham are not mentioned again. §3 reports that a non-standard view exists. It does not show that view is right |
| C2 | Unintelligent agency is a coherent concept, because a determinate kind of it can be described | weak | the taxonomy (§§5–13). It shows that a description can be written. It does not show that the description picks out agency rather than causal efficacy plus design, because the criterion common to every kind, interactivity, holds of any physical system that acts and is acted upon |
| C3 | Interactivity is necessary for agency, and on its own it is enough for natural agency | stated | ”without some form of interaction, there can be no agency” (p. 8). Natural agents have “no autonomy or adaptability as previously described, only interactions” (p. 9), yet they are agents (p. 9). Neither half is argued |
| C4 | Some non-living systems, such as rivers, are agents | weak | the river example (p. 9). The support given is that animism and design arguments misattribute “higher forms of agency” to them, and “its presence lends support in favour of the correct recognition of some non-living systems as (some kind of) agents” (p. 9). That is an error-theory argument run backwards: that people misread a thing is taken as evidence that something is there to read |
| C5 | AI lacks intelligence, consciousness, understanding and mental states | assertion | stated at pp. 18, 19, 20 and 24. No criterion or test is given, and on the paper’s own adverbial theory (p. 3) the denial needs the strong, human-level sense of intelligence that ARI names, not the weak sense conceded on p. 2 |
| C6 | Artificial agency is a new kind, distinct from artefactual agency | weak | the kinds differ by “advanced learning capabilities” (p. 18) and openness of learning, a difference of degree on the adaptability criterion. The smart thermostat already learns (p. 13) |
| C7 | AI’s goals are set by human purposes, and it cannot choose its own goals or “choose whether to choose” | assertion | pp. 18–19. The same view Popa 2021 is cited for (p. 6) |
| C8 | Agentic AI is a new collective kind of agency that achieves coordination and “collective intelligence” without consciousness or intelligence | weak | the description of a hypothetical system (§13.1). “Collective intelligence” is in tension with C5 unless it is meant in the weak sense of p. 2, and the paper does not say which |
| C9 | Comparing agents needs a stable LoA, and LoAs are ranked by purpose | stated | p. 5. The paper does not name the LoA its comparison uses |
| C10 | Only human agents have duties; other sentient agents may have rights | assertion | one sentence (p. 15) |
Method
The paper uses conceptual taxonomy. It defines three graded criteria and describes each kind of agency in four parts: definition, example, limitations and implications. Then it compares the kinds in prose (§11). The Method of Abstraction is introduced in §2, and an illustration (the thermostat at four LoAs) is the only place it is applied. The formal machinery of LIT-900 (typed observables, moderated LoAs, gradients) and of LIT-898 (transition systems) does not appear.
Concepts
- Agency: not defined. It is a cluster concept identified by overlapping features, best read adverbially (pp. 3–4).
- Interactivity: mutual influence between agent and environment (p. 7). It is necessary for every kind of agency.
- Autonomy: initiating state changes “independently of direct external causation”, graded and LoA-relative (p. 8). It is close to Floridi & Sanders’ internal transitions. It is not self-governance.
- Adaptability: modifying behaviour “based on input” (p. 8). This is weaker than Floridi & Sanders’ change of transition rules (see Connections).
- Artificial Agency (AA): computational, goal-driven agency whose goals are set by human purposes. It has data-driven adaptability and no intelligence or mental states. It is “a syntactic form of agency” (pp. 18–20).
- Social Artificial Agency, Agentic AI: networked AI systems coordinating toward complex goals with minimal oversight (p. 21).
- ARI / MRA: the two ways of fitting AI into our concepts (p. 2). Floridi treats them as the horns of a dilemma.
- Standard / non-standard view: agency needs mental states, or agency needs only some degree of interactivity (pp. 5–6).
Connections
What it adds to the 2023 editorial (NOTE-697). The title, the slogan and the denial of understanding are kept. The paper adds what NOTE-697 found missing. There are now criteria (the three), a place for the method of abstraction, and a placing of LLMs among other agents. The editorial’s one demarcation, that a sea wave “can make nothing but that difference” (2023, p. 5), is changed in kind. A river is now a natural agent (p. 9). The line it drew between agents and mere causes has become a line between two kinds of agent. The editorial’s remark that RL fine-tuning “solves agency” is dropped. Base and fine-tuned models are not distinguished.
What it changes from Floridi & Sanders 2004 (NOTE-701). It takes the three criteria explicitly (p. 7). That is the reason for the extends on the LIT. Three changes come with them, and the paper marks none of them as a change.
- Agency no longer needs all three. In 2004 an agent at the proposed LoA had to be interactive, autonomous and adaptable. The postbox and the mill, interactive only, were not agents (2004, Fig. 2). In 2025, interactivity alone gives natural agency. Floridi & Sanders’ “for most entities there is no LoA at which they can be considered an agent” (2004, p. 26 of the preprint) is not repeated, and it is hard to square with C3.
- Adaptability is weaker. In 2004 the interactions “(can) change the transition rules”. In 2025 the agent modifies “its behaviour based on input” (p. 8). Read literally, this is met by anything whose output depends on its input, which is interactivity again. The MENACE analysis, on whether a rule change seen in the code is adaptation at all, is not taken up.
- Moral agency drops out. Accountability, responsibility, criterion (O) and thresholds are absent. Morality appears as a mark of the human kind (p. 15) and as a governance problem (§§12.3, 14).
The method of abstraction (NOTE-700). The paper cites Floridi 2008 for the method (p. 4) and uses its vocabulary: interface, observable, non-hierarchical levels, metábasis, pluralism without relativism. That reading asked whether an agency claim names the system’s own LoA or the analyst’s. The answer here is the analyst’s. The “Kantian sense” (p. 5) makes that explicit. It also asked whether any argument in the agency papers uses a relation between LoAs. None does. No extends to Floridi 2008 is written. The method is named and illustrated, but nothing in the comparison depends on it, and the comparison would read the same without §2.
Distributed morality (LIT-901, LIT-899). These are cited in one sentence as Floridi’s answer to responsibility gaps (p. 6). The taxonomy is not connected to them. Agentic AI’s emergent behaviours make “traditional models … inadequate” (p. 23), but the paper does not say whether distributed morality is the replacement.
Other held works it cites. Frankfurt 1971 (LIT-692, NOTE-545) is cited once, for the moral dimension of human agency: “a form of agency that incorporates moral consideration into decision-making and action” (p. 15). That is not what Frankfurt’s paper is about. The nearer point of contact is uncited: AI’s inability to “choose whether to choose” (p. 19) is the higher-order structure Frankfurt uses to mark persons off from wantons. Frankfurt’s wantons are agents who are not persons. That is the same cut MRA makes, from the other side. Prigogine & Stengers 1984 (LIT-534, unread) is cited for the limits of natural agency (p. 10). Dennett is cited for Kinds of Minds (1996, not held) on interactivity and biological limits (pp. 7, 11). Nothing else in the 59 references is held. List 2021, Dung 2024, Popa 2021, van Lier 2023 and Himma 2009 remain unregistered. The paper surveys them and does not engage them.
Not cited, and bearing on it. Shanahan et al. (LIT-460, NOTE-370) deny the bare model goals of its own, “not even simulated versions”. Floridi calls AI “goal-driven” and “goal-directed” (p. 18), and in the same place says the goals are “defined by human purposes”. The two agree that the goals are not the system’s own. They disagree on whether goals that come from training and prompting still make the system an agent, and neither paper argues the point. Dennett’s “True Believers” (LIT-398, NOTE-353) runs a thermostat-to-person continuum with “no magic moment”. Floridi’s “branching pattern” (p. 17) has an “unbridgeable gap” at consciousness and intelligence (p. 19). He does not say what would show that a system had crossed it, so the gap is not defended against the intentional strategy.
Bearing on the record
- CLAIM-010, the permissiveness objection. The paper bites the bullet harder than its 2004 source does, and it does not restrict the LoA. The objection’s own list was “a thermostat … a river … a reef” (A112). On this account a thermostat is an agent: an artefactual agent, illustrated at four LoAs (p. 4) and as the smart thermostat (p. 13). A river is a natural agent (p. 9). An ant colony is an animal social agent (p. 12). So the objection no longer bites as a reductio, but the cost is that “agent” at the base level does no sorting. The discriminating work moves to the kinds. The paper’s answer is grading, not indexing. It does not suppose an agent and ask how its efficacy is organised, which was the record’s answer (CLAIM-080). It widens agency and then subdivides it. Since CLAIM-010 is about will, it is worth noting that the paper says nothing about will.
- QUESTION-009, when coordination is an additional agent. Three collective kinds are named: ant colonies, corporations and Agentic AI. No criterion is given for when a group is one. The implicit mark is surplus capability. Colonies “exhibit greater capability than individual agency” (p. 12), human social agency creates “capabilities (e.g., learning) that surpass the sum of individual agents’ contributions” (p. 16), and Agentic AI “challenges traditional boundaries between individual and collective action” (p. 21). These are descriptions, not a test. The question stays open, and this paper does not bear on it more than Floridi & Sanders did.
- CLAIM-030 and CLAIM-045, the distributed-agency line. The paper is compatible with CLAIM-030. Agentic AI’s “emergent behaviours not explicitly anticipated by their designers” (p. 21) and corporate capabilities beyond the members’ (p. 16) are the shape of that claim, but the paper gives no evidence for it. On CLAIM-045, its call for “a multi- and inter-disciplinary study of agency, not dissimilar but perhaps more successful than the field of cybernetics and complexity studies” (p. 24) is a wish, not a division of sub-questions.
- ADR-024, agent, individual, self and person. The paper’s main cut agrees with the decision. Agency is kept apart from intelligence, consciousness and personhood, and legal personhood “is a related but distinct debate” (p. 7). But the human kind is built by folding in what ADR-024 keeps apart. Human individual agency includes consciousness, “complex personal identities (identity construction, narrative self-understanding)” and “biographical coherence (a sense of self and purpose)” (pp. 14–15). So self, narrative unity and personhood become features of a kind of agency. A document citing the paper for the agent/person separation should cite §§1–3 and §12, not §9.
- The organisational account (LIT-566, NOTE-458, THEORY-071). The two are rivals. On THEORY-071, rivers and thermostats are not agents. Neither is an AI whose goals are “defined by human purposes” (p. 18), because a norm fixed from outside is an external constraint (footnote 3 there). Floridi’s own description of artificial agency is thus the condition that account excludes. The two do share something. Floridi’s biological agency includes “internal state maintenance through homeostasis” and “self-maintenance” (p. 10), but as features of one kind, not conditions of agency. Neither paper cites the other.
- THEORY-173 (large models are a cultural and social technology, not agents). Also a rival. It agrees on the facts Floridi asserts: no understanding, statistical processing, goals not the model’s own. It reaches the opposite verdict on agency, because its criterion for an agent includes setting novel goals and revising a world model from evidence. Floridi denies AI the first (p. 19), so by that criterion his artificial agents are not agents. The dispute between the two is about the criterion, and neither states a case that would decide it.
- THEORY-043 (agency survives both directions). The agency that account calls uncontested is the thin kind. Floridi’s is that kind, whatever he calls it. His refusal of the label “functionalist” (p. 5) rests on observables not being behavioural, and that does not make the criteria thicker.
- Constitutivism (LIT-117, unread; THEORY-065, THEORY-075). The paper offers constitutivists nothing. Its agency has no constitutive aim and no definition, and it ties duties to the human kind (p. 15), not to agency as such. A river’s agency is the clearest case yet of an agency from which no norm could follow. Whether that counts as a “shmagency” in Enoch’s sense is for a reading of LIT-117 to say.
- A THEORY is filed from this reading (THEORY-208). It is Proposed, on the record’s practice for comparable accounts of agency (THEORY-071, THEORY-173). It states MRA as the record reads it: graded criteria, kinds marked by origin, mechanism and goal source, with AI placed as computational agency serving goals set by humans. It carries C1 and C5 as open liabilities.
- No instruction for machine-learning practice. The development remarks, “prioritising reliability and reproducibility” (p. 20) and working “within its distinct agency type rather than attempting to replicate human … intelligence” (p. 20), are orientation, not technique. The governance remarks are policy, not ML practice. The anthology holds none of this line.
Limitations
- The thesis of the title is not argued. Its four supports are named and only one is used, as a survey (C1). The paper says itself that it is answering a narrower objection, that unintelligent agency is nonsense (p. 2).
- The dilemma is not exclusive. On Floridi’s own concession, AI is intelligent in a weak sense (p. 2). On his adverbial theory, intelligence is “behaving intelligently” (p. 3). Then MRA and a weak ARI can both hold. The denial of intelligence (C5) needs the strong sense, and the paper switches between senses: Agentic AI “exhibits forms of collective intelligence” (p. 21) and has a “reasoning engine” (p. 21).
- No LoA is specified. §2 says agency attributions depend on the LoA and must keep it stable across systems (p. 5). The comparison never names the observables it uses. As the table above shows, the kinds are told apart by observables (origin, substrate, mental states) that differ from the three criteria the paper says organise them.
- The criteria do not sort. Natural agency needs interactivity only, and every other kind has all three criteria. So the criteria separate rivers from everything else, and nothing else from anything. “Artificial” versus “artefactual” is a difference in how open the learning is (C6).
- Internal tensions. Natural agents have “no capacity for goal-directed behaviour” (p. 9), yet rivers show “dynamic equilibrium maintenance” (p. 9). AI lacks “genuine intentionality” (p. 18), while dogs have “elements of intentionality” (p. 10), and no observable distinguishes the two. AI excels “in narrow, well-defined domains” (p. 24) and also offers “comprehensive pattern recognition across diverse domains” and “multitasking across diverse domains” (p. 18).
- Editing. §13.2, “Limitations”, opens with “A primary benefit” (p. 22). One sentence ends “continue to set new AI” (p. 22). Footnote 1 has “influce” and “Chisolm” (p. 3), and the last paragraph runs “sustainability.All informed people” (p. 25). The roadmap (p. 3) calls §6 “Animal Individual Agency (Biological Agency)”, and the section is titled “Biological Agency”. The author apologises for the length (p. 1) and calls §§4–11 “tedious” (p. 3).
- The examples are thin where the paper is new. Artificial agency is exemplified by “OpenAI’s GPT series” in one paragraph (p. 19). Agentic AI is exemplified by a hypothetical travel planner, since “many old applications are often merely rebranded” (p. 22).
Open questions
- What LoA does the comparison use? An answer would list the observables, including whatever observable carries “intelligence” or “mental states”, and give their values for an LLM, a smart thermostat and a dog. That would test whether the kinds are distinguished at one stable LoA, as §2 requires.
- Is interactivity sufficient for agency, or only necessary? If sufficient, every physical system that acts and is acted upon is a natural agent, and Floridi & Sanders’ “for most entities there is no LoA” is withdrawn. If not, the further condition natural agency needs (self-organisation? dynamic equilibrium?) should be stated.
- What would show that an AI system has crossed the “unbridgeable gap” (p. 19)? Without a criterion for intelligence or understanding, C5 cannot fail, and the thesis cannot be told apart from a decision about how to use the word.
- When is a coordinated set of AI systems one Agentic AI rather than several artificial agents interacting? This is QUESTION-009 for the paper’s newest kind.