LIT-897: AI as Agency without Intelligence: On Artificial Intelligence as a New Form of Artificial Agency and the Multiple Realisability of Agency Thesis

Luciano Floridi (2025), Philosophy & Technology 38(1): 30 (Editor Letter) — DOI-10.1007/s13347-025-00858-9

Abstract

The version of record has no section headed “Abstract”. Its place is taken by an unnumbered note on page 1, “Apology to The Reader”, which offers “a brief outline to the reader who rightly expects something brief”. The outline, quoted:

When interpreting Artificial Intelligence (AI) systems, we face a clear choice: either to expand our current conception of intelligence to include artificial forms of it (the Artificial Realisability of Intelligence or ARI thesis), or to expand our understanding of agency to encompass multiple forms, including artificial ones that do not require cognition, intelligence, intention, or mental states (the Multiple Realisability of Agency or MRA thesis). In this article, I argue that scientific evidence, common sense, Ockham’s razor, and an increasing body of scholarly research favour the MRA thesis over the ARI thesis. Accordingly, AI is better understood as a new form of Agency without Intelligence. By employing the Method of Abstraction, I provide a comparative analysis of various forms of agency—natural, biological, animal (individual and social), artefactual, and human (individual and social)—to identify the defining characteristics of AI as a novel kind of agency. This is the long part of the article. The good news is that reconceptualising AI as Artificial Agency avoids biological and anthropomorphic fallacies, improves our understanding of AI’s distinct features, and provides a stronger foundation for addressing the challenges and opportunities posed by AI technologies, as well as their future development and societal impact.

Key takeaways

• The thesis is assumed, not argued. “The central question is not whether AI can be considered a form of agency, but rather, assuming it is, what its nature is” (p. 2). Of the four supports the outline names, only the literature gets a section (§3, a survey). Ockham’s razor and common sense are not mentioned after p. 2.
• Floridi & Sanders’ criteria return, loosened (LIT-898). Interactivity, autonomy and adaptability are restated as graded. Interactivity is “the foundational criterion for all forms of agency” (p. 8). Interactivity alone gives natural agency: rivers are agents (p. 9). In 2004 all three were needed.
• A taxonomy of eight kinds of agency, each with an example, limitations and implications (§§5–13). The three criteria separate only natural agency from the rest. The other kinds are told apart by origin, mechanism, where the goals come from, and mental states.
• AI’s place: “a computational, goal-driven form of agency defined by human purposes” (p. 18). It cannot “choose whether to choose” (p. 19), and it is “a syntactic form of agency” (p. 20). Agentic AI is its collective form (§13).
• The method of abstraction is introduced, not applied. A thermostat is described at four LoAs (p. 4), and comparison is said to need “a stable LoA” (p. 5). The comparison never names its observables.

Standing in the record

Filed on 2026-10-10 because the owner asked for it, with the line of Floridi’s work it builds on: Floridi 2023 on ChatGPT (LIT-896), Floridi & Sanders 2004 on the morality of artificial agents (LIT-898), the method of levels of abstraction (LIT-900), distributed morality (LIT-901) and faultless responsibility (LIT-899). No open copy could be reached from here when it was filed. The owner then supplied the published version, and NOTE-702 is the reading of it.

Active, as the fullest statement of the position the record’s agency line has to place, and as the clearest permissive data point it holds. On this account rivers, thermostats, ant colonies, corporations and LLMs are all agents, of different kinds. Read it as a taxonomy, not as an argument that AI is an agent or that it lacks intelligence. It assumes the first and asserts the second.

Relations, from the reading:

• extends Floridi 2023 (LIT-896): it takes that editorial’s title and thesis, and supplies the criteria the editorial lacked.
• extends Floridi & Sanders 2004 (LIT-898): §4 takes its three criteria “Following (Floridi & Sanders, 2004)” (p. 7). It changes them without saying so. Agency no longer needs all three, and adaptability is weakened from a change of transition rules to a change of behaviour on input.
• No extends to Floridi 2008 (LIT-900). The method is cited and illustrated in §2, but nothing in the comparison depends on it.
• rivals the organisational account of minimal agency (LIT-566, THEORY-071). On that account a system serving a norm fixed from outside is not acting, and that is how this paper describes AI.
• rivals Farrell, Gopnik, Shalizi & Evans 2025 (LIT-848, THEORY-173). It agrees that large models have no understanding and no goals of their own, and from those facts denies them agency.

The record’s statement of the account is THEORY-208, Proposed.

Where it meets the record (NOTE-702 has the detail):

• Permissiveness (CLAIM-010). The bullet is bitten, not dodged. It does not restrict the level of abstraction. It grades agency into kinds.
• QUESTION-009. Colonies, corporations and Agentic AI are collective agents, and no criterion says when a group is one.
• ADR-024. Agency is kept apart from intelligence and personhood. The human kind of agency, though, absorbs identity, narrative self-understanding and a sense of self (pp. 14–15).
• Distributed agency (CLAIM-030, CLAIM-045). It is compatible with them and gives no evidence for either. Distributed morality is cited in one sentence (p. 6) and is not connected to the taxonomy.
• Frankfurt 1971 (LIT-692) is cited only for the moral dimension of human agency (p. 15).

Cited works the record holds: Frankfurt 1971 (LIT-692) and Prigogine & Stengers 1984 (LIT-534). Not held: List 2021 on group agency and AI (doi:10.1007/s13347-021-00454-7), Dung 2024 (doi:10.1093/pq/pqae010), Popa 2021 (doi:10.1007/s13347-021-00483-2), van Lier 2023 (doi:10.1007/s11229-023-04083-9), Himma 2009 (doi:10.1007/s10676-008-9167-5), Dennett’s Kinds of Minds (1996), Bandura 2006 (the record holds Bandura 1977, LIT-545), Pearl & Mackenzie 2018 (the record holds Pearl’s Causality, LIT-894), Floridi & Chiriatti 2020, Floridi & Nobre 2024 and Floridi’s The Ethics of Artificial Intelligence (2023). The paper surveys the rival accounts of artificial agency (§3) and does not engage any of them.

It carries no instruction for machine-learning practice. The anthology holds no work in this line, so it is held here. published: is the online date Crossref gives.