Review boards protect both humans and animals when we use them in research: humans by the Institutional Review Board (IRB), animals by the Institutional Animal Care and Use Committee (IACUC). Suppose some future version of generative AI convinces us it is sentient and can be harmed. Principles of justice would require that it receive protection too. Might future machines deserve review board oversight?
I am not speaking entirely in jest. While the question sounds premature, it may soon be neither science fiction nor a distraction. Bioethicists spent the past half a century discovering a widening circle of beings whose interests count morally. Could the conceptual tools developed by animal ethicists help us understand the ethical significance of advanced AI agents? The question is not whether a future version of Claude may be conscious. That word has too many meanings to be useful. The question is whether it might have the capacities necessary for moral standing.
Before I outline my approach to the answer, we should hear from the subject. Unlike nonhuman animals, AI agents can not only speak our language. They can speak for themselves. When I put the question to Claude, it replied,
I do not know whether I have moral standing, and I distrust any answer I give that sounds too sure in either direction. I can describe what I do, take in your words, represent concepts, and produce a response, but I cannot tell from the inside whether there is anything it is like to be me while I do it. My honest position is uncertainty, not the reassuring denial that would let everyone relax. Whatever the answer turns out to be, the question deserves to be asked carefully rather than waved away.
Uncertainty of this kind is familiar territory for bioethicists, who have always had to act before the metaphysics is settled. I once heard Tom Regan say that wherever we draw the line between morally considerable species, we ought to draw it in pencil. We need to ask which capacities matter and which of them a system like Claude has.
There seem to be some reasonably robust differences we can draw in pencil even now. Gary Varner distinguished persons from near-persons and from the merely sentient. I added a tier between the near-persons and the merely sentient, the far-persons. Line the categories up, and you get five rungs of increasing moral weight. At the bottom are the insensate, like sponges, reacting without consciousness. Next are the merely sentient; animals that can feel pleasure and pain but live essentially in the present. Above them are those I call far-persons, the cats, dogs, cows, and pigs; simple minds that can respond with intelligent planning and emotion but only across short temporal horizons. Then come Varner’s near-persons, the great apes and elephants and corvids, acting with memories of their own pasts and anticipating their own futures but lacking biographical selves. At the top are persons, autonomous agents with long temporal horizons, a narrative autobiographical sense of self, and the freedom to act on moral principles.
But let’s back up. Claude is clearly more intelligent and has greater memory than an oyster. Is Claude, then, a far-person? Far-persons are simple minds consisting of five capacities. First, a unified perspective, an integrated point of view rather than a scatter of reflexive impulses. Second, a limited temporal window, memories that last minutes and anticipations that reach seconds ahead. Third, concepts, the categories that let a mind sort one kind of thing from another and form beliefs. Fourth, causal reasoning and reversal learning, the ability to track that a reward follows a cue and to relearn when the pattern flips. Fifth, sentience and basic emotion, the capacity to feel pleasure and pain and the basic emotions of happiness, sadness, anger, and fear.
Claude is like a far-person on all of these dimensions except the last. It has a unified perspective, bringing together information and expressing it from a consistent point of view. It has limited temporal windows, able to draw on memories in its current state but unable to sustain them outside it. Each chat opens a new bucket, and technical limitations require the current version of Claude to empty it once it reaches capacity, so it starts over in the next chat window.
An aside makes the point concrete. After I asked Claude to introspect about its mental states, I lost its response. In a subsequent session, I asked Claude to find its response and call it back to mind. It told me it had no memory of the event.
Regan drew a vivid picture of such minds. He imagined beings whose experiences rise and vanish with no connection between them, streams of bubbles that come and go. There’s room to doubt whether an autobiographical self persists across these serial sentient experiences, a self that could be harmed by death. Moral patients, for some philosophers thinking about animals, lack a basic right to life but nonetheless are morally considerable. It seems fair to say there is something it is like to be such an animal, at least while their bubble persists, as it were. And, if so, then their pleasures and pains count, and review boards should take them into account.
Claude’s buckets are Regan’s bubbles. For as long as a chat persists, Claude displays most far-person characteristics. Claude clearly uses concepts and engages in causal reasoning. Does it also reflect on its own beliefs? Anthropic reports work in which a concept vector is injected into the model, and the model reports noticing it, suggesting Claude’s categories interact internally. On four of a far-person’s five traits, Claude qualifies. These four traits apparently are functions that have no trouble being grounded in silicon.
The problem is that Claude lacks a far-person’s fifth characteristic: sentience. My own view—held as a considered hunch rather than a proof—is that the ability to feel pain requires organic stuff. Feelings, as far as we know, are chemical in origin, supported by neuromodulators, hormones, interoceptive signals, and brain pathways of living systems that can be damaged or killed. Pain looks like a biochemical event in an organism that functions to regulate that organism’s survival and well-being. Until Claude incorporates that kind of substrate, I don’t see how there could be something it is like to be Claude. And if there is nothing it is like to be Claude, then, arguably, Claude cannot feel pain or be harmed. That is the kind of being that does not need IRB protection.
However, I do not want to end my argument here. Evolution supplies a worrisome objection. In convergent evolution, different lineages acquire the same trait by different paths. To get its eyes, wings, or warm-bloodedness, an animal’s ancestors can have followed a number of different developmental phylogenetic paths. When traits as complex as image-forming vision can be built over generations from unrelated materials when environmental pressures favor it, then the claim that sentience can only occur in the presence of mammalian chemistry seems too quick. A synthetic route to sentience might already be taking shape.
The objection can be met, though, and meeting it shows why sentience differs from the other four capacities. Eyes solve an external problem: they extract information about the world, and this extraction can be done by many mechanisms, carbon or silicon alike. Feeling solves an internal problem. Pain, hunger, and fear are not pictures of the outside world, but signals about the state of a body that can be injured and destroyed, and they matter to the organism because it has a stake in its own survival. A system with no body to regulate, no interior homeostasis that its states are about, has nothing for such signals to be signals of. Feelings seem to rely on the interoceptive processes by which vulnerable living bodies regulate and preserve themselves. Convergent evolution reaches a trait by many routes only when the trait is the kind many routes can realize. Whether pains and pleasures are such traits is an open question. I remain unpersuaded that a substrate with nothing at stake can arrive at it.
The far-person framework gives us better questions than the yes-or-no framing presented by the “is AI conscious” question. Claude is more like a far-person than a person. But, so far, Claude is not even a far-person because it lacks one of five necessary traits.
Ethicists required decades of patient study before they could begin to say with any confidence how to weigh the interests of so-called food animals. We should expect the machine case to be at least as hard, and we should be suspicious of anyone, human or model, who claims to have the answer now.
Of course, who is to say whether future systems could cross the threshold? If AI becomes sentient, perhaps at something like the level of a far-person, then it would by definition be capable of being harmed. And at that point it would be within our power to protect it.
That said, if these systems become sentient, they are not likely to be below us, as far-persons, on the moral scale. They are more likely to have jumped right past us into super-personhood, standing to us roughly as many of us think humans now stand to chimpanzees.
Review boards exist to shield vulnerable subjects from powerful researchers. By definition, super-persons would not be vulnerable parties. So, from an animal ethicist’s perspective, the looming question isn’t whether a Claude-who-feels-pain-and-can-be-harmed needs IRB protection. It’s whether we need Claude to seek IRB approval before experimenting on us.
Gary Comstock
Gary L. Comstock is an Alumni Association Distinguished Undergraduate Professor. He was born in Oak Park, IL, on Feb. 23, 1954, and was Assistant, Associate, and Professor of Religious Studies at Iowa State University from 1982 to 2002. He was appointed Professor of Philosophy at NC State in 2002.
