Luciano Floridi & J. W. Sanders (2004), Minds and Machines 14(3), 349–379 (August 2004 issue) — DOI-10.1023/B:MIND.0000035461.63578.9d
Key takeaways
- Agenthood is relative to a level of abstraction. At the LoA the paper proposes, an agent is a system that is interactive, autonomous and adaptable. At a video-camera LoA a rock is none of the three, a thermostat is only interactive, and a human is all three (Figure 2). Learning software, webbots and organisations are agents at a suitable LoA (§2.6).
- The same system can be an agent at one LoA and not at a lower one. MENACE, Michie’s matchbox noughts-and-crosses learner, is adaptable when only its games are observed. When its matchboxes are observed, the “learning” shows up as a deterministic update of state, and adaptability fails (pp. 11–13). Whether closed-source software is an agent therefore depends on whether its code can be seen.
- A moral agent is any agent capable of an action that “can cause moral good or evil” (criterion O, p. 15). This criterion is neither consequentialist nor intentionalist.
- Accountability and responsibility come apart. An agent is accountable for what it is the source of. To be responsible as well, it needs “the right intentional states” (p. 21). Search-and-rescue dogs and Oedipus are the paper’s moral agents without responsibility. An artificial agent can be “accountable — though not responsible” (p. 26).
- Moral goodness at an LoA is a threshold function on the observables that stays within a tolerance. Human ethical judgement sets the tolerance, so the formalism supplies none of the normative content.
- For Computer Ethics, the paper argues that most of the ACM Code applies to artificial agents. Censure of an immoral artificial agent is maintenance, disconnection or deletion (§4).
Standing in the record
Filed on 2026-10-10 at the owner’s request, as part of the line of work that Floridi’s “AI as Agency without Intelligence” (2025, LIT-897) builds on. The anthology holds none of the works in this line. This is the earliest paper in the line: it is where Floridi first separates moral agency from intelligence and mental states (“AAs, though not intelligent and fully responsible, can be fully accountable sources of moral action”, p. 3), and the 2023 and 2025 papers carry it into their title, “agency without intelligence” (LIT-896). Its levels-of-abstraction method is the one used in Floridi’s informational structural realism (LIT-152, NOTE-099) and his account of personal identity (LIT-136). Here the method is only outlined, and its definition is deferred to a 2003 Floridi & Sanders chapter, “The Method of Abstraction”, which the record does not hold. The paper promises that its LoA will make “distributed morality” intelligible (pp. 3, 16), and takes this no further than naming it in the conclusion. Floridi 2013 (LIT-901) is the paper that takes it up.
For the record’s agency line, the paper gives an answer to QUESTION-009: a pattern is an agent when, at some LoA, it is interactive, autonomous and adaptable. That answer is as permissive as the objection the line already carries (CLAIM-010), and NOTE-701 sets out why.
NOTE-701 is the close reading of 2026-10-10, and it placed the work: Active — worth reading as the founding statement of the line. It argues by cases and by rebutting objections, and its criteria are LoA-relative in a way the paper embraces and does not limit.