Source
Floridi (2025), LIT-897, read in NOTE-702: §1 for the thesis, §4 for the criteria, §§5–13 for the kinds, and §12 for artificial agency. The criteria are taken from Floridi & Sanders (2004), LIT-898, read in NOTE-701, which the 2025 paper follows and loosens.
What was actually shown
Nothing was tested. The paper is a conceptual taxonomy, and it says outright that it assumes its premise: “assuming it is [a form of agency], what its nature is” (p. 2).
- The criteria. Interactivity is necessary for every kind: “without some form of interaction, there can be no agency” (p. 8). Autonomy and adaptability are graded, and are absent from natural agency, where there is “no autonomy or adaptability … only interactions” (p. 9). Agency is not defined by necessary and sufficient conditions (p. 3).
- The kinds. These are natural (rivers), biological (dogs), animal social (ant colonies), artefactual (smart thermostats), human individual, human social (corporations), artificial (LLMs) and social artificial (Agentic AI). They form “a branching pattern” (p. 17), not a ladder.
- AI’s place. “A computational, goal-driven form of agency defined by human purposes” (p. 18). It learns statistically and openly, and it cannot choose its own goals (p. 19). It has no consciousness, intelligence or understanding (p. 19).
What could have come out otherwise: a stated level of abstraction under which artificial agency failed to differ from artefactual agency except in degree. The paper writes none down, so nothing could.
What this does not say
- Not that AI is an agent because it is intelligent, or that it is not intelligent because it is an agent. The denial of intelligence is a separate premise, asserted (pp. 18–20), and the account is only as strong as that premise. In a weak sense the paper concedes AI is intelligent (p. 2). On its adverbial reading, intelligence is “behaving intelligently” (p. 3). The denial therefore needs the strong, human-level sense.
- Not that the three criteria sort the kinds. They separate natural agency from the rest and nothing else. The other distinctions use origin, substrate, goal source and mental states.
- Not an answer to the permissiveness objection (CLAIM-010). The account accepts it: rivers and thermostats are agents. It answers by subdividing agency into kinds, not by restricting what counts.
- Not a criterion for collective agency. Colonies, corporations and Agentic AI are described as collective agents, and no test says when a group is one (QUESTION-009).
- Not about selfhood or personhood. Those enter only as features of the human kind (pp. 14–15).
- No instruction for machine-learning practice.
Connections
- THEORY-071 (organisational minimal agency) is a rival. On it, a system serving an externally fixed norm is not acting, and that is exactly how this account describes AI.
- THEORY-173 (large models are a social technology, not agents) is a rival. It agrees that large models have no understanding and no goals of their own, and denies agency because its criterion requires self-set goals.
- THEORY-043. The agency here is the thin kind that account treats as uncontested.
- The 2023 editorial (LIT-896) states the slogan. This account is the 2025 paper’s development of it.