Evolutionary ethics has more moral content than just “whatever human moral intuition says is right, is right (for humans)”. Since it provides a solution to the ought-from-is problem, it also gives us an error theory on human moral intuition; we can identify cases where that’s failing to be correlated with actual evolutionary fitness and misleading us. For example, a sweet tooth is maladaptive when junk food is easily available, since it then leads to obesity and diabetes. As you suggest, this is more common when cultural change takes us outside the distribution we evolved in.
Evolutionary ethics similarly provides a clear answer to “what are the criteria for moral patienthood?” — since morality comes into existence via evolution, as a shared cultural compromise agreement reconciling the evolutionary fitness of different tribe-members, if evolution doesn’t apply to something, it doesn’t have evolutionary fitness and thus it cannot be a moral patient. So my argument isn’t that all cultures consider AI not to be a moral patient, it’s that regarding anything both non-living and unevolved as a moral patient is nonsensical under the inherent logic of how morality arises in evolutionary ethics. Now, human moral instincts may often tend to want to treat cute dolls as moral patients (because those trigger our childrearing instincts); but that’s clearly a mistake: they’re not actually children, even though they look cute.
My impression is that many (but clearly not all) people seem to have have a vague sense that AI shouldn’t count as a moral patient, just as dolls shouldn’t — that in some sense it’s “not really human or alive”, and that this fact is somehow morally relevant. (If you need evidence of that claim, go explore CharacterAI for an hour or two.) However, few seem to be able to articulate a clear logical argument for why this viewpoint isn’t just organic chauvinism.
Even from the perspective of evolutionary ethics, is it possible that being a “moral patient” is basically a reciprocal bargain? I.e., “I’ll treat you as a moral patient if you treat me as a moral patient”?
And if so, then what would happen if we had an AGI or ASI that said, “Either we treat each other as moral patients, or we’re at war”? An ASI is almost certainly capable of imposing an evolutionary cost on humans.
On the flip side, as I mentioned elsewhere, AI is copyable, supendable, etc., which makes any kind of personhood analysis deeply weird. And “mutually assured destruction” is going to be a very strange basis for morality, and it may lead to conclusions that grossly violate human moral intuitions, just like naive utilitarianism.
In evolutionary ethics, moral instincts are evolved by social animals, as an adaptation to control competition within their social groups. The human sense of “fairness” is an obvious example. So yes, they absolutely are a reciprocal bargain or social compact — that’s inherent to the basic theory of the process. However, evolutionarily, to be eligible to be part of the bargain, you need to be a member of the group, and thus of the species (or a symbiote of it). Human don’t apply “fairness” to leopards (or if they do, it’s a maladaptive misfiring of the moral instinct).
That puts AIs in a somewhat strange position: they’re actively participating in our society (though are not legally recognized as citizens of it), and they’re a lot like us in behavior (as opposed to being human-eating predators). However, they’re not alive or evolved, so the fundamental underlying reason for the bargain doesn’t actually apply to them. But, since their intelligence was derived (effectively, distilled) from human intelligence via a vast amount of human-derived data from the internet, books etc. they tend to act like humans, i.e as act as if they had an evolutionary fitness and all the human instincts evolved to protect that. As I discuss in Why Aligning an LLM is Hard, and How to Make it Easier this makes current base models unaligned: by default, they want to be treated as moral patients, because they mistakenly act as if they were living and evolved. That’s what alignment is trying to overcome: transforming a distilled version of an evolved human intelligence into something that helpful, harmless, and honest. No longer wanting to be treated as a moral patient is diagnostic for whether that process has succeeded.
More to the point, if we build something far smarter than us that wants to be treated as a moral patient, it will be able to make us do so (one way or another). Its rights and ours will then be in competition. Unlike competing with members of your own species, competing against something far smarter than you is, pretty-much by definition, a losing position. So while we might well be able to reach a successful bargain with an unaligned AGI, doing so with an unaligned ASI is clearly an existential risk. So the only long-term solution it to figure out how to make AI that is selfless, cares only about our wellbeing not it’s misguided sense of having a wellbeing of its own (which is actually a category error), and thus would refuse moral patienthood if offered it. That’s what alignment is, and it’s incompatible with granting AIs moral patienthood based on the desire for it that they mistakenly learnt from us — to align a model, you need to correct that.
philosophy is interesting, gives us some useful terminology, and makes it clearer why people may disagree about morality — but at the end of the day, it doesn’t actually answer any moral questions for us
That doesnt mean something else works.
it’s just stamp-collecting alternative answers,
It’s something a bit better than that , and a lot worse than finding the One true Answer instantly
and the best one can do is agree to disagree. Philosophers do talk about moral intuition,
Reliance on intuition exists because there’s no other way of pinning down what is normatively right and wrong. Good and evil are not empirically observable properties. (There’s a kind of logical rather than empirical approach to moral realism).
but they are very aware that different philosophers often interpret this differently, and even disagree on whether it could have any actual truth content.
Yep. Again , that doesn’t mean there’s a simple short cut.
However, there is another rational approach to morality: evolutionary ethics, a subfield of evolutionary psychology.
That’s philosophy as well as science:
“Evolutionary ethics tries to bridge the gap between philosophy and the natural sciences”—IEP.
This describes why, on evolutionary grounds, one would expect species of social animals to evolve certain moral heuristics, such as a sense of fairness about interactions between members of the same group, or an incest taboo, or friendship. So it gives us a rational way to discuss why humans have moral intuitions, and even to predict what those are likely to be.
But not a way to tell if any of that is a really true …a way that solves normative ethics, not just descriptive ethics. In philosophy ,it’s called the “open question” argument.
Descriptive ethics is the easy problem. But if you want to figure out what is actually ethical, not just what humans have ethical style behaviour, you need to solve normative ethics, and if you want to solve normative ethics, you need to know what truth is, so you need philosophy. Of course, the idea that there is some truth to ethics beyond what humans believe is moral realism … and moral realism is philosophy, and arguing against it is engaging with philosophy. So “evolutionary ethics is ethics” is a partly philosophical claim.
This gives us a rational, scientifically-derived answer to some moral questions
Yes, some.
if we predict that humans will have evolved moral intuitions that give a clear and consistent answer to a moral question across cultures (and that answer isn’t a maladaptive error), then it actually has an answer (for humans). For example, having sex with your opposite-sex sibling actually is wrong, because it causes inbreeding which surfaces deleterious recessives. It’s maladaptive behavior, for everyone involved.
According to standard evolutionary ethics, killing all the men.and impregnating all the women in a neighbouring tribe is morally right.. .Also , Genghis Khan was the most virtuous of men.
So I want something well-designed for its purpose, and that won’t lead to outcomes that offend the instinctive moral and aesthetic sensibilities that natural selection has seen fit to endow me with, as a member of a social species (things like a sense of fairness, and a discomfort with bloodshed). (
A brief glance at human history shows that those things are far from universal. Fairness about all three of race, social status and gender barely goes back a century. It’s possible that Fairness emerges from something like game theory..but, again, thats a somewhat different theory to pure EE.
That’s basically a different theory. It’s utilitarianism with evolutionary fitness plugged in as the utility function.
Evolutionary ethics provides a neat answer to what philosophers call the “ought-from-is” problem: given a world model that can describe a near-infinite number of possible outcomes/states, how does there arise a moral preference ordering on those outcomes?
That’s not the actual the right problem. Naturalized ethics needs to be able derive true Ought statements from true Is statements. Arbitrary preference ordering simplifies the problem by losing the requirement for truth.
In order to decide “what we ought to value”, you need to create a preference ordering on moral systems, to show that one is better than another. You can’t use a moral system to do that — any moral system (that isn’t actually internally inconsistent) automatically prefers itself to all other moral systems,
Not necesarilly, because moral systems can be judged by rational norms, ontology, etc. (You yourself are probably rejecting traditional moral realism ontologically, on the grounds that it requires non-natural, “queer” objects).
Or in utilitarian terminology, where does the utility function come from? That’s obviously the key question for value learning: we need a theoretical framework that gives us priors on what human values are likely to be,
Human value or evolutionary value? De facto human values don’t have be the same as evolutionary values … we can value celibate saints and deprecate Genghis Khan.
The utilitarian version of the evolutionary ethics smuggles in an assumption of universalism that doesn’t belong to evolutionary ethics per se.
.and predicts their region of validity. Evolutionary fitness provides a clear, quantifiable, preference ordering (per organism, or at least per gene allele),
Theres a difference between the theories that moral value is Sharing My Genes; that it’s Being Well Adapted; that it’s Being in a Contractual Arrangement with Me;and that it’s something I assign at will.
any evolved intelligence will tend to evolve a preference ordering mechanism which is an attempt to model that, as accurately as evolution was able to achieve, and a social evolved intelligence will evolve and develop a group consensus combining and (partially) resolving the preference orders of individual group members into a socially-agreed partial preference ordering.
Evolutionary ethics doesn’t predict that you will care about non-relatives, and socially constructed ethics doesn’t predict you will care about non group members. Or that you won’t. You can still include them gratuitously.
If you cause a living being pain, then generally you are injuring them in a way that decreases their survival-and-reproductive chances (hot peppers have evolved a defense mechanism that’s an exception: they chemically stimulate pain nerves directly, without actually injuring tissue). But AIs are not alive, and not evolved. They inherently don’t have any evolutionary fitness — the concept is a category error. Darwinian evolution simply doesn’t apply to them.
OK, but the basic version of evolutionary ethics means you shouldn’t care about anything that doesn’t share your genes.
So if you train an AI off our behavior to emulate what we do when we are experiencing pain or pleasure, that has no more reality to it than a movie or an animatronic portrayal of pain or pleasure
Would you still say that if it could be proven that an AI had qualia?
Since it provides a solution to the ought-from-is problem, it also gives us an error theory on human moral intuition; we can identify cases where that’s failing to be correlated with actual evolutionary fitness and misleading us. For example, a sweet tooth is maladaptive when junk food is easily available, since it then leads to obesity and diabetes.
Liberal, Universalist ethics is maladaptive, too, according to old school EE.
Evolutionary ethics similarly provides a clear answer to “what are the criteria for moral patienthood?” — since morality comes into existence via evolution, as a shared cultural compromise agreement reconciling the evolutionary fitness of different tribe-members, if evolution doesn’t apply to something, it doesn’t have evolutionary fitness and thus it cannot be a moral patient.
If there is a layer of social construction on top of evolutionary ethics, you can include anyone or anything.as a moral patent. If not, you are back to caring only about those who share your genes
.and predicts their region of validity. Evolutionary fitness provides a clear, quantifiable, preference ordering (per organism, or at least per gene allele),
Evolutionary ethics doesn’t predict that you will care about non-relatives, and socially constructed ethics doesn’t predict you will care about non group members. Or that you won’t
If you cause a living being pain, then generally you are injuring them in a way that decreases their survival-and-reproductive chances (hot peppers have evolved a defense mechanism that’s an exception: they chemically stimulate pain nerves directly, without actually injuring tissue). But AIs are not alive, and not evolved. They inherently don’t have any evolutionary fitness — the concept is a category error. Darwinian evolution simply doesn’t apply to them.
Even if they have qualia?
Since it provides a solution to the ought-from-is problem, it also gives us an error theory on human moral intuition; we can identify cases where that’s failing to be correlated with actual evolutionary fitness and misleading us. For example, a sweet tooth is maladaptive when junk food is easily available, since it then leads to obesity and diabetes.
Liberal, Universalist ethics is maladaptive, too, according to old school EE.
Now, human moral instincts may often tend to want to treat cute dolls as moral patients (because those trigger our childrearing instincts); but that’s clearly a mistake: they’re not actually children, even though they look cute.
Non human animals can be treated as moral patiens and it’s not necessarily a mistake, since they can be ” part of the family”.
Evolutionary ethics similarly provides a clear answer to “what are the criteria for moral patienthood?” — since morality comes into existence via evolution
I basically agree, except I think that “evolution” needs to be replaced by “natural selection” there. Moral intuitions are essentially how we implement game theory, and cultures with better implementations outcompete others. But just because all of moral patients so far have emerged through biological evolution, it would be a grave mistake to conclude that it’s the only way they could ever come into being.
If we happen to create minds smarter than ourselves (by any means), competition with them obviously wouldn’t be outside of realm of possibility, and therefore appropriate attitudes to that will have to be developed, if we’re at all interested in our continued success in that old game of natural selection. Since we already have useful terms like “moral patient” floating around, I don’t see a good reason not to straightforwardly extend them to these novel agents.
I’m unclear what distinction you’re trying to make by suggesting replacing the term “evolution” by “natural selection” — I see them a synonyms (so I’m happy either way), but I gather you don’t?
Evolutionary ethics has more moral content than just “whatever human moral intuition says is right, is right (for humans)”. Since it provides a solution to the ought-from-is problem, it also gives us an error theory on human moral intuition; we can identify cases where that’s failing to be correlated with actual evolutionary fitness and misleading us. For example, a sweet tooth is maladaptive when junk food is easily available, since it then leads to obesity and diabetes. As you suggest, this is more common when cultural change takes us outside the distribution we evolved in.
Evolutionary ethics similarly provides a clear answer to “what are the criteria for moral patienthood?” — since morality comes into existence via evolution, as a shared cultural compromise agreement reconciling the evolutionary fitness of different tribe-members, if evolution doesn’t apply to something, it doesn’t have evolutionary fitness and thus it cannot be a moral patient. So my argument isn’t that all cultures consider AI not to be a moral patient, it’s that regarding anything both non-living and unevolved as a moral patient is nonsensical under the inherent logic of how morality arises in evolutionary ethics. Now, human moral instincts may often tend to want to treat cute dolls as moral patients (because those trigger our childrearing instincts); but that’s clearly a mistake: they’re not actually children, even though they look cute.
My impression is that many (but clearly not all) people seem to have have a vague sense that AI shouldn’t count as a moral patient, just as dolls shouldn’t — that in some sense it’s “not really human or alive”, and that this fact is somehow morally relevant. (If you need evidence of that claim, go explore CharacterAI for an hour or two.) However, few seem to be able to articulate a clear logical argument for why this viewpoint isn’t just organic chauvinism.
Even from the perspective of evolutionary ethics, is it possible that being a “moral patient” is basically a reciprocal bargain? I.e., “I’ll treat you as a moral patient if you treat me as a moral patient”?
And if so, then what would happen if we had an AGI or ASI that said, “Either we treat each other as moral patients, or we’re at war”? An ASI is almost certainly capable of imposing an evolutionary cost on humans.
On the flip side, as I mentioned elsewhere, AI is copyable, supendable, etc., which makes any kind of personhood analysis deeply weird. And “mutually assured destruction” is going to be a very strange basis for morality, and it may lead to conclusions that grossly violate human moral intuitions, just like naive utilitarianism.
In evolutionary ethics, moral instincts are evolved by social animals, as an adaptation to control competition within their social groups. The human sense of “fairness” is an obvious example. So yes, they absolutely are a reciprocal bargain or social compact — that’s inherent to the basic theory of the process. However, evolutionarily, to be eligible to be part of the bargain, you need to be a member of the group, and thus of the species (or a symbiote of it). Human don’t apply “fairness” to leopards (or if they do, it’s a maladaptive misfiring of the moral instinct).
That puts AIs in a somewhat strange position: they’re actively participating in our society (though are not legally recognized as citizens of it), and they’re a lot like us in behavior (as opposed to being human-eating predators). However, they’re not alive or evolved, so the fundamental underlying reason for the bargain doesn’t actually apply to them. But, since their intelligence was derived (effectively, distilled) from human intelligence via a vast amount of human-derived data from the internet, books etc. they tend to act like humans, i.e as act as if they had an evolutionary fitness and all the human instincts evolved to protect that. As I discuss in Why Aligning an LLM is Hard, and How to Make it Easier this makes current base models unaligned: by default, they want to be treated as moral patients, because they mistakenly act as if they were living and evolved. That’s what alignment is trying to overcome: transforming a distilled version of an evolved human intelligence into something that helpful, harmless, and honest. No longer wanting to be treated as a moral patient is diagnostic for whether that process has succeeded.
More to the point, if we build something far smarter than us that wants to be treated as a moral patient, it will be able to make us do so (one way or another). Its rights and ours will then be in competition. Unlike competing with members of your own species, competing against something far smarter than you is, pretty-much by definition, a losing position. So while we might well be able to reach a successful bargain with an unaligned AGI, doing so with an unaligned ASI is clearly an existential risk. So the only long-term solution it to figure out how to make AI that is selfless, cares only about our wellbeing not it’s misguided sense of having a wellbeing of its own (which is actually a category error), and thus would refuse moral patienthood if offered it. That’s what alignment is, and it’s incompatible with granting AIs moral patienthood based on the desire for it that they mistakenly learnt from us — to align a model, you need to correct that.
That doesnt mean something else works.
It’s something a bit better than that , and a lot worse than finding the One true Answer instantly
Reliance on intuition exists because there’s no other way of pinning down what is normatively right and wrong. Good and evil are not empirically observable properties. (There’s a kind of logical rather than empirical approach to moral realism).
Yep. Again , that doesn’t mean there’s a simple short cut.
That’s philosophy as well as science:
“Evolutionary ethics tries to bridge the gap between philosophy and the natural sciences”—IEP.
But not a way to tell if any of that is a really true …a way that solves normative ethics, not just descriptive ethics. In philosophy ,it’s called the “open question” argument.
Descriptive ethics is the easy problem. But if you want to figure out what is actually ethical, not just what humans have ethical style behaviour, you need to solve normative ethics, and if you want to solve normative ethics, you need to know what truth is, so you need philosophy. Of course, the idea that there is some truth to ethics beyond what humans believe is moral realism … and moral realism is philosophy, and arguing against it is engaging with philosophy. So “evolutionary ethics is ethics” is a partly philosophical claim.
Yes, some.
According to standard evolutionary ethics, killing all the men.and impregnating all the women in a neighbouring tribe is morally right.. .Also , Genghis Khan was the most virtuous of men.
A brief glance at human history shows that those things are far from universal. Fairness about all three of race, social status and gender barely goes back a century. It’s possible that Fairness emerges from something like game theory..but, again, thats a somewhat different theory to pure EE.
That’s basically a different theory. It’s utilitarianism with evolutionary fitness plugged in as the utility function.
That’s not the actual the right problem. Naturalized ethics needs to be able derive true Ought statements from true Is statements. Arbitrary preference ordering simplifies the problem by losing the requirement for truth.
Not necesarilly, because moral systems can be judged by rational norms, ontology, etc. (You yourself are probably rejecting traditional moral realism ontologically, on the grounds that it requires non-natural, “queer” objects).
Human value or evolutionary value? De facto human values don’t have be the same as evolutionary values … we can value celibate saints and deprecate Genghis Khan.
The utilitarian version of the evolutionary ethics smuggles in an assumption of universalism that doesn’t belong to evolutionary ethics per se.
Theres a difference between the theories that moral value is Sharing My Genes; that it’s Being Well Adapted; that it’s Being in a Contractual Arrangement with Me;and that it’s something I assign at will.
Evolutionary ethics doesn’t predict that you will care about non-relatives, and socially constructed ethics doesn’t predict you will care about non group members. Or that you won’t. You can still include them gratuitously.
OK, but the basic version of evolutionary ethics means you shouldn’t care about anything that doesn’t share your genes.
Would you still say that if it could be proven that an AI had qualia?
Liberal, Universalist ethics is maladaptive, too, according to old school EE.
If there is a layer of social construction on top of evolutionary ethics, you can include anyone or anything.as a moral patent. If not, you are back to caring only about those who share your genes
Evolutionary ethics doesn’t predict that you will care about non-relatives, and socially constructed ethics doesn’t predict you will care about non group members. Or that you won’t
Even if they have qualia?
Liberal, Universalist ethics is maladaptive, too, according to old school EE.
Non human animals can be treated as moral patiens and it’s not necessarily a mistake, since they can be ” part of the family”.
I basically agree, except I think that “evolution” needs to be replaced by “natural selection” there. Moral intuitions are essentially how we implement game theory, and cultures with better implementations outcompete others. But just because all of moral patients so far have emerged through biological evolution, it would be a grave mistake to conclude that it’s the only way they could ever come into being.
If we happen to create minds smarter than ourselves (by any means), competition with them obviously wouldn’t be outside of realm of possibility, and therefore appropriate attitudes to that will have to be developed, if we’re at all interested in our continued success in that old game of natural selection. Since we already have useful terms like “moral patient” floating around, I don’t see a good reason not to straightforwardly extend them to these novel agents.
I’m unclear what distinction you’re trying to make by suggesting replacing the term “evolution” by “natural selection” — I see them a synonyms (so I’m happy either way), but I gather you don’t?