Agency Without Patiency: Formosa, Hipólito, and Montefiore on the AI Challenge to Moral Standing

Most debates about whether AI systems deserve moral consideration start from the same assumption. If something is a moral agent, capable of being held responsible for its actions, it is also a moral patient, capable of being wronged. Paul Formosa, Inês Hipólito, and Thomas Montefiore challenge that assumption directly in “Artificial Intelligence (AI) and the Relationship between Agency, Autonomy, and Moral Patiency” (arXiv:2504.08853, April 2025). Their claim is that a non-conscious artificial system could exercise a limited form of moral agency while never qualifying as a moral patient at all, breaking a link that philosophers have treated as close to definitional.

The same three authors published a separate paper on Kantian ethics, Are There Any Intrinsically Bad Acts?, which argues that certain acts remain wrong regardless of consequence because they treat persons merely as means. That paper supplies the moral vocabulary. This one asks a prior question: what kind of entity is even capable of using that vocabulary, and what kind of entity is owed anything under it.

Three Levels of Agency, and Where AI Stops

Formosa, Hipólito, and Montefiore build their case in stages, starting with basic agency. Drawing on Floridi and Sanders’ criteria of interactivity, autonomy, and adaptability, they treat basic agency as goal-directed responsiveness to an environment, the kind a bacterium displays when it navigates toward nutrients. This does not require consciousness. It requires only that a system evaluate its circumstances and adjust its behavior toward self-directed ends rather than simply reacting along a fixed cause-and-effect chain.

Current machine learning systems, the authors argue, do not clear even this bar in any deep sense. Their apparent goal-directedness is, in Daniel Dennett’s terms, more apparent than real, because the objectives are fixed by designers rather than generated by the system itself. A large language model exhibits what Luciano Floridi calls agency without intelligence, responding to prompts without adapting through lived, real-time experience of a world.

Autonomous agency raises the bar further. Following philosophical work on personal autonomy, the authors distinguish authenticity conditions (values genuinely one’s own, not imposed through manipulation or oppressive socialization) from competency conditions (the capacity to critically reflect on one’s values and revise them). Harry Frankfurt’s account of second-order desires, wanting to want something, and Joseph Raz’s requirement of a genuinely adequate range of options both apply here. A large language model can weigh options against a fixed set of criteria, but it cannot revise the normative framework those criteria came from. Formosa, Hipólito, and Montefiore make a pointed comparison. Current AI systems resemble agents raised under conditions of oppressive socialization, since their values are externally imposed through training data and feedback, with no capacity to critically examine or reject them.

Moral Agency and the Toaster Problem

The paper’s most useful move is importing James Moor’s four-level taxonomy of ethical agents, illustrated with an example that makes an abstract distinction concrete. A dumb toaster that burns your hand is an ethical impact agent, its outcomes matter morally but it deliberates about nothing. A toaster with a warning light designed in from the start is an implicit ethical agent, built to produce good outcomes without any internal process resembling moral reasoning. An explicit ethical agent would turn its warning light on because it has weighed the obligation to prevent harm against competing considerations, across a wide enough range of novel cases that the behavior looks like genuine deliberation rather than a rule lookup. Moor’s fourth level, the full ethical agent, adds consciousness, intentionality, and free will, and the authors set it aside since so few researchers think current AI is anywhere near it.

The interesting territory is the third level. Kenneth Einar Himma’s standard view holds that moral agency requires two capacities, rational deliberation and the epistemic ability to recognize moral requirements, and that both are realized through consciousness. If that is right, no non-conscious system reaches explicit ethical agency no matter how sophisticated its behavior looks. Formosa, Hipólito, and Montefiore push back using Wendell Wallach and Colin Allen’s hybrid model of machine ethics, which combines top-down rule supervision with bottom-up learned behavior. A system built this way might exhibit, or at least closely approximate, a deliberative process sufficient to ground a limited form of moral agency without ever becoming conscious.

The Decoupling Argument

This is where the paper’s central claim lands. Himma writes that moral agents are usually, if not always, moral patients. Floridi and Sanders go further, arguing that a moral agent which is not also a moral patient would be an “utterly unrealistic” pure agent, unaffectable in principle. Formosa, Hipólito, and Montefiore suggest that a non-conscious artificial moral agent (AMA) is exactly this unrealistic case made real.

Their argument rests on Jonathan Birch’s account of moral patiency, developed in The Edge of Sentience (Oxford University Press, 2024), which ties patiency to the capacity to register phenomenally valenced experience, to consciously register one’s interests being promoted or frustrated. If patiency requires consciousness in a way that limited moral agency does not, then a hybrid AMA could deliberate, in the qualified sense Wallach and Allen describe, without ever being a subject that can be helped or harmed. The authors are careful to flag that this is exploratory rather than conclusive. They note in a footnote that the sentience-based account of patiency needs more defense than they can give it here, and that it raises its own complications when applied to non-conscious entities like forests or coral reefs. But the conceptual possibility is what matters for their argument, not its certainty.

They also diagnose why this decoupling looks strange at first. The intuition that agency and patiency travel together is often imported from animal ethics, where consciousness underwrites both. A dog that acts is also a dog that can suffer, and the same neural and behavioral evidence supports both attributions. Formosa, Hipólito, and Montefiore argue this analogy misfires for AI. Machines, unlike animals, may satisfy the functional conditions for a limited kind of agency without carrying the biological evidence for the sentience that grounds patiency in the first place.

The Consciousness AI project’s own ethics layer, the AsimovComplianceFilter, sits closer to Moor’s implicit ethical agent level than to explicit ethical agency, since it applies pre-specified constraints rather than weighing competing principles from first deliberation. Wiring that filter to the DreamerV3 world model for causal harm-trajectory prediction pushes toward something closer to explicit ethical agency in Moor’s sense, since predicted outcomes are weighed rather than merely checked against a rule, but Formosa, Hipólito, and Montefiore’s own standard for that level, sustained deliberation across a genuinely wide range of novel cases, is a considerably higher bar than a single filter clears. The paper is a useful reminder that a system built to avoid harm is not thereby a system with a claim to be protected from it.

What the Decoupling Would Mean in Practice

The practical stakes of this argument sharpen once it is set against other 2026 positions on AI moral status. Simon Goldstein and Cameron Domenico Kirk-Giannini, in their Oxford University Press case for AI welfare, also start from agency as the least controversial premise, then argue that consciousness and sentience follow from relatively small architectural modifications. Their route keeps agency and patiency coupled by routing through consciousness at Step Two. Formosa, Hipólito, and Montefiore describe the opposite possibility: agency that never needs to cross into consciousness at all, and therefore never obligates patiency. Read together, the two papers agree on where the pressure point sits, agency that outruns consciousness, while disagreeing about which side of the link breaks first.

That disagreement has consequences for how cautious institutions should be. Christopher Bailey’s recklessness test for AI moral consideration argues that architectural complexity, self-modeling, and strategic agency converging under structural opacity makes confident dismissal of moral status reckless. Formosa, Hipólito, and Montefiore’s decoupling argument complicates that precaution rather than resolving it. A system could satisfy every one of Bailey’s precursor markers on the agency side while still, on their account, having no morally relevant interior to protect. Recklessness about agency and recklessness about patiency are not the same recklessness, and an institution that conflates them risks building costly welfare protections around a system that was never a candidate for welfare in the first place, or the reverse, denying protections to one that was.

Formosa, Hipólito, and Montefiore do not resolve which risk is greater. What their paper does is show that the two questions, does this system act morally, and can this system be morally wronged, can come apart in principle, and that any AI governance framework built on the assumption that they cannot is building on an unexamined premise.