Luciano Floridi (2016), Philosophical Transactions of the Royal Society A 374(2083): 20160112, in the themed issue “The ethical impact of data science” — DOI-10.1098/rsta.2016.0112
Key takeaways
- A distributed moral action (DMA) is a morally good or evil outcome caused by local interactions that are each morally neutral, meaning below a threshold of moral relevance (fn 11, p. 3). Intention is not closed under causal implication, so nobody intends a DMA (p. 4). An ethics for which intention is necessary therefore either ignores DMAs or reduces them to individual intentional acts (p. 4).
- The proposal (p. 7) has five steps. (a) Identify the DMA. (b) Identify the network causally accountable for it. (c) Back-propagate responsibility so that each node is “prima facie equally and maximally responsible”. (d) Correct the outcome. (e) Repeat. A node escapes by showing “no involvement” (p. 8). Responsibility stays with the nodes, not the network: “‘theirs’ (agents’) not ‘its’ (network’s)” (p. 8).
- Each of the three borrowings is “more an inspiration than a template” (p. 3). Back propagation supplies the picture of correction flowing backward to learners, but the equal, maximal default is the opposite of back propagation’s proportional credit assignment. Strict liability supplies the precedent of liability without fault. Common knowledge, reached by public announcement, is what makes the rule preventive. It is also what partly answers the charge of unfairness (p. 10), and it does so by giving agents foreknowledge, so that responsibility is no longer quite faultless.
- Two Dutch Supreme Court rulings illustrate the rule (p. 9). Three cyclists riding abreast are all held fully responsible (1948). Of four boats moored abreast only the fourth is held responsible (1931). Both rulings go by how easily each party could have rectified the situation.
- The paper answers two objections, that the rule is unfair and that it is unrealistic, and raises two challenges for it: free-riding on the prudent and the chilling of innovation (pp. 10–11). The remedies end in allocation “proportionally to the ability of the agents to avoid the negative outcome” (p. 11).
Abstract
The paper’s own abstract (p. 1):
The concept of distributed moral responsibility (DMR) has a long history. When it is understood as being entirely reducible to the sum of (some) human, individual and already morally loaded actions, then the allocation of DMR, and hence of praise and reward or blame and punishment, may be pragmatically difficult, but not conceptually problematic. However, in distributed environments, it is increasingly possible that a network of agents, some human, some artificial (e.g. a program) and some hybrid (e.g. a group of people working as a team thanks to a software platform), may cause distributed moral actions (DMAs). These are morally good or evil (i.e. morally loaded) actions caused by local interactions that are in themselves neither good nor evil (morally neutral). In this article, I analyse DMRs that are due to DMAs, and argue in favour of the allocation, by default and overridably, of full moral responsibility (faultless responsibility) to all the nodes/agents in the network causally relevant for bringing about the DMA in question, independently of intentionality. The mechanism proposed is inspired by, and adapts, three concepts: back propagation from network theory, strict liability from jurisprudence and common knowledge from epistemic logic.
Standing in the record
Filed on 2026-10-10 with the paper the owner asked for, Floridi 2025 (LIT-897), which cites it once as Floridi’s answer to responsibility gaps. The anthology holds none of the works in this line. It was registered Deferred, seeded from the abstract, because no copy could be reached. The owner supplied the published version, and NOTE-699 is the close reading of the same day. The reading placed it Active.
It extends Floridi 2013 (LIT-901, NOTE-698), which defined distributed moral actions and left responsibility for them open. The two papers are “a diptych, but they do not presuppose knowledge of each other” (fn 4, p. 2). It also extends Floridi & Sanders 2004 (LIT-898), whose conditions for agency (interactive, autonomous, adaptive) the nodes must meet, and whose “mindless morality” it carries over from agency to responsibility. It cites the method of levels of abstraction (LIT-900) once (p. 6), and ties causation to a level of abstraction in fn 15.
What the reading found, against the seed’s questions:
- Responsibility, not agency. The paper says nothing about agency beyond assuming the 2004 conditions and adding that nodes must learn (p. 6). Its claim is about allocation. What it allocates is causal accountability “and therefore … morally answerable” (p. 6), a step it gives no argument for. On p. 10 it disowns blame and punishment as the point. This is the 2004 paper’s accountability, which that paper kept apart from responsibility (NOTE-701). The 2016 paper neither revises nor mentions that split.
- Back propagation is a metaphor. It explains why responsibility is a corrective sent to learners. It does not fix “causally relevant”, and the theory of causation is deferred (fn 15).
- Common knowledge makes the rule preventive and partly answers the unfairness objection. The 2013 paper’s model, distributed knowledge held by a “supra-agent” (NOTE-698), is gone, and so is the supra-agent. Responsibility goes to the nodes. The change of operator is substantive, and the paper does not remark on it.
- Collective responsibility. It allocates to the nodes, not to the network, and so takes no side on QUESTION-009. It agrees with CLAIM-030 that the whole’s output is not what members intended. From that it draws the opposite allocation from French (LIT-727) and Pettit (LIT-718), who make the organised whole a bearer of responsibility. A THEORY states its account (THEORY-207).