A self-driving car swerves and someone gets hurt. A loan model denies an application on grounds that turn out to be discriminatory. The instinct is to ask who is to blame, and the awkward part is that the immediate cause was a machine that no one thinks has a mind. So a prior question has to be settled first: is the machine even the kind of thing that can be blamed, or is it just a complicated instrument whose blame belongs entirely to the people around it? That is the question of whether an AI can be a moral agent, and it turns out to split cleanly along the same fault line that runs through the consciousness debate. One camp says agency can be read off behavior alone. The other says it requires having felt something, which no algorithm has.
The idea
A moral agent is an entity that performs actions and is an appropriate target of praise or blame; moral agency is whatever bundle of capacities makes an entity that kind of thing, and the whole disagreement is over what belongs in the bundle. Luciano Floridi and J.W. Sanders argue the bundle is thin and fully external: an entity is a moral agent if, at the right level of abstraction, it is interactive, autonomous, and adaptive, none of which requires a mind. This is their “mind-less morality”, and it separates being the causal source of an action (accountability) from having the intentions and free will behind it (responsibility), holding that accountability alone is enough. Carissa Véliz argues the opposite: moral agency requires sentience, the capacity to feel pleasure and pain, because you cannot understand moral concepts without having felt something. Algorithms are “moral zombies”, acting like agents while feeling nothing, so they cannot be moral agents at all. Where the mind question was ethics-adjacent before, here it becomes the ethics question directly.
Floridi and Sanders: agency you can read off behavior
The Floridi and Sanders move is to refuse the demand for a mind before the conversation even starts. In “On the Morality of Artificial Agents” they propose that agenthood rests on three criteria evaluated at a chosen level of abstraction, the interface through which you observe the system. Interactivity means the agent and its environment act on each other, responding to stimulus by changing state. Autonomy means the agent can change its own state without a triggering stimulus, so it is not a mere pass-through. Adaptability means the agent can change the very rules by which its states transition, which is to say it can learn. Every one of these is confirmable from the outside, by watching what the system does at the interface, with no appeal to what, if anything, it is like on the inside.
The payoff they draw is what they call mind-less morality. If those three observable properties are all agency requires, then you can evaluate a system’s actions as morally good or evil without ever settling whether it has free will, mental states, or an inner life. Morality becomes a threshold defined on the observables at that level of abstraction: an agent is good if all its actions respect the threshold and evil if some action crosses it. The engine of the whole view is a distinction they insist on between two things ordinarily fused together. Accountability is being the causal source of an action, the place the good or bad outcome actually came from. Responsibility is the richer notion that carries intention, belief, and free will. Their claim is that an artificial agent can be fully accountable while never being responsible, and that accountability is sufficient to count it as a moral agent, a genuine source of good and bad in the world, even though praise and blame in the full backward-looking sense may not attach. This is a deliberately deflationary answer, built so that the responsibility gap does not leave machine harms floating free of any moral vocabulary at all.
Véliz: no feeling, no understanding, no agency
Véliz attacks the bundle from the other end. In “Moral Zombies” she borrows the philosophical zombie, a being that behaves exactly like a conscious person while there is nothing it is like to be it, and asks whether moral agency could be one of those behaviors you can fake all the way down. Her answer is no, and the load-bearing capacity is sentience, the capacity to have subjective experiences of pleasure and pain, which is the phenomenal side of consciousness wearing an ethical name. Algorithms, she argues, plainly lack it. They are moral zombies: they can act indistinguishably from moral agents, yet there is nothing it is like to be them, and that absence is not a detail but the whole problem.
The argument runs through moral understanding. To be a moral agent, on her account, you have to understand moral concepts, and moral concepts are not the sort of thing you can grasp from descriptions alone. To understand what it means to inflict pain on someone, you need experiential knowledge of pain; a creature that has never felt anything can only process the words for suffering, the way a colorblind person can master every fact about red without ever seeing it. The felt experience is doing work no amount of information can replace. Since algorithms have no sentience, they have no experiential purchase on the concepts, so they cannot understand morality, so they cannot be moral agents. She turns this directly against Floridi and Sanders on two points. Their autonomy is too broad: a tornado also changes its state and goes about its business without external help, and that tells us nothing about whether it is an agent, much less a moral one. And she rejects the accountability-without-responsibility maneuver as distorting the word. Morality is intrinsically about normative evaluation, so if something really is a moral agent, we must be able to hold it responsible, which requires the sort of inner life, the capacity to feel guilt and to improve, that an algorithm does not have.
What the disagreement is really about
Strip away the vocabulary and the two positions are picking different jobs for the phrase “moral agent” to do. Floridi and Sanders want a forward-looking, systems-level notion fit for computer ethics, a way to say “this system is a source of harm and its behavior can be regulated against a moral threshold” without waiting on the unanswerable question of machine consciousness. On their level of abstraction, insisting on a mind is a category error that leaves too much of the moral landscape undescribed. Véliz wants the ordinary, backward-looking notion, the one that licenses blame, guilt, and moral growth, and she argues that stretching the term to cover mindless systems empties it of exactly the content that made it worth having. The clash is not really about what algorithms can do, since both sides agree on the behavior. It is about whether the felt interior is optional decoration on moral agency or its foundation.
That is why this debate is where the mind question stops being a curiosity and becomes decisive. If the biological-substrate objection is right and something about felt experience rides on machinery our silicon lacks, then Véliz’s requirement is one no current system can meet, and mind-less morality is the only kind of machine morality on offer. If instead functionalism holds and feeling follows from the right functional organization, then a future system might satisfy even Véliz’s demand, and the disagreement collapses into a timing question rather than a difference in principle. Either way the practical question of who answers for a machine’s harms, whether the machine can advise us as a moral advisor without being an agent, and whether tomorrow’s models cross into agency, all wait on the same unresolved fact about whether anything is felt inside.
The same act, judged by each account
- A person shoves someone off a ledge. Interactive, autonomous, adaptive, and sentient. Both accounts agree: a moral agent, accountable and responsible, an appropriate target of blame and capable of guilt.
- An algorithm denies a loan discriminatorily. For Floridi and Sanders it is interactive, autonomous, and adaptive at the right level of abstraction, so it is a moral agent, accountable as the source of a bad outcome even if no one is behind the wheel. For Véliz it is a moral zombie: no sentience, no understanding of the harm, so not a moral agent, only a conduit for the responsibility of its makers.
- A tornado destroys a house. Autonomous in Floridi and Sanders’s sense, changing state without a stimulus, yet not interactive or adaptive in the required way. Véliz uses it to show that autonomy alone proves nothing about agency, which is exactly why she thinks the three-criteria bundle is too thin.
- A hypothetical sentient AI. If a machine ever genuinely felt pleasure and pain, the two accounts would finally converge, and the argument would move from “can it ever” to “has this one.”
Related Notes
- Consciousness: Access vs Phenomenal, the distinction that separates the functional side Floridi and Sanders rely on from the felt side Véliz demands
- The Biological-Substrate Objection, the leading reason to think no current machine has the sentience Véliz requires
- The Responsibility Gap, what mind-less accountability is trying to close and what Véliz’s account leaves open
- AI as Moral Advisor, the alternative role for a system that can reason about morality without being a moral agent
- Could an LLM Be Conscious?, the live question that decides whether Véliz’s bar is permanently or only currently out of reach
Sources
- Floridi, Luciano and J.W. Sanders, “On the Morality of Artificial Agents,” Minds and Machines 14(3), 2004. https://link.springer.com/article/10.1023/B:MIND.0000035461.63578.9d . Supports the three criteria for agenthood (interactivity as response to stimulus by change of state, autonomy as ability to change state without stimulus, adaptability as ability to change the transition rules) evaluated at a given level of abstraction, and the mind-less morality conclusion that the concept of moral agent need not exhibit free will, mental states, or responsibility, with morality as a threshold on the observables at the interface.
- Véliz, Carissa, “Moral zombies: why algorithms are not moral agents,” AI & Society 36, pp. 487-497, 2021 (open access via PMC). https://pmc.ncbi.nlm.nih.gov/articles/PMC7613994/ . Supports that algorithms are functional moral zombies lacking sentience, that moral agency requires an agent to be autonomous and morally responsible via moral understanding derived from experiential knowledge of pleasure and pain (“to understand what it means to inflict pain on someone, it is necessary to have experiential knowledge of pain”), the criticism that Floridi and Sanders’s autonomy is too broad (the tornado example), and the criticism that a genuine moral agent must be evaluable as responsible, not merely accountable.
- “Machine ethics,” Wikipedia. https://en.wikipedia.org/wiki/Machine_ethics . Supports the framing of the artificial-moral-agent question, the split between consciousness/intrinsic-property accounts (robots lack consciousness and subjective experience and so cannot be harmed) and relational/behavioral accounts, and the governance emphasis on accountability, oversight, and redress rather than treating machines as responsible agents.
- “Sentience,” Wikipedia. https://en.wikipedia.org/wiki/Sentience . Supports sentience defined as the capacity to experience feelings and sensations, the narrower valenced definition (the capacity for positive or negative experiences such as pain and pleasure), and its connection to the subjective “what it is like” quality of conscious experience.