Exploring Sentience, AI Ethics, and a Just Future with Jeff Sebo

Not as much as I should. Yeah. [snorts]

Yes. So different people do describe sentience in different ways. For me, sentience is the capacity to consciously experience positive and negative states like pleasure, pain, happiness, suffering, satisfaction, frustration. Different sentient beings are going to experience different types of states. Some might have more intense bodily pleasures and pains than others. Some might have more intense intellectual, emotional pleasures and pains than others, or different kinds, of course. But for all sentient beings, if they can consciously experience states with a positive and negative valence, states that they like or dislike, states that they want more or fewer of, then they really have a stake in what happens to them. It matters to them what happens to them. And for me and for many philosophers, for many other people, that is a key condition, a key criterion for when they really have a kind of intrinsic value and deserve consideration for their own sakes in decisions that affect them. Now, I am open to the possibility that other types of capacities or relationships or features are sufficient for moral consideration, too. But sentience to me definitely, or at least very likely, suffices for moral consideration. And so I really do want to both conduct in my, my own case and support for other people work that can advance understanding both of which beings are or might be sentient and what they want and need, and then what we owe them and what follows for our practices, priorities and, and our, our policies. So that is why I emphasize that kind of slogan when, when I talk about what unifies a lot of my work.

No such thing as...

That totally does make sense. And this is an area of active research and debate right now, and, and a relatively new emerging area of, of research and debate. So one way to put an increasingly commonly held view is that whether you are sentient is all or nothing, binary. Either you have feelings with a positive and negative valence or not. But then how sentient you are, that might come in degrees. In other words, how intense and prolonged your experiences are, how good or bad they feel to you, how good or bad they are for you, that might come in degrees. So, for example, we might think that if both an elephant and an ant are sentient, then they are equally sentient in the basic sense that they can both have experiences that feel good or bad to them. But it may be that because the elephant has a more complex brain and nervous system and a longer lifespan, that the elephant is capable of experiencing more intense, prolonged states of pleasure, pain, happiness, and suffering. And, and for that reason, if a house is burning down and I have to choose between saving one elephant and saving one ant, then maybe the stakes are higher for the elephant because I would be preventing more suffering in the moment and allowing for more happiness in the future. And, and so they are both sentient, full stop. But the elephant might have a greater welfare capacity and, and in that sense be more sentient.

Yeah, good question. And also just to respond to your previous point, I think we may find for many species, if not all species, there are both explanations, right? So, so the evolutionary explanation of sex might be that it supports procreation and the extension of the species, but then the way that it plays that role is by using pleasure as an incentive for engaging in the activity. And so, so it might be that an individual organism has sex because it feels good for them. But then the reason it feels good for them is that it has this evolutionary advantage it was selected for. Right? So, so it, in, in many cases it might be that what seem like two competing explanations are actually two complementary explanations, and both of them are correct at different levels of explanation. But to answer, answer your current question, the difference between consciousness and sentience, as I understand it and many other philosophers understand it, is that consciousness is part of sentience, a precondition of sentience. So consciousness is the capacity for subjective awareness, subjective experience. Basically, if it feels like something to be you, then you are a conscious being. This is what philosophers sometimes call phenomenal consciousness. And then sentience is consciousness plus valence. Consciousness plus it feeling good or bad in particular. So, so we can imagine conscious states that are purely neutral, like pure color perception, that it feels like something to see the color red, but it might not feel good or bad, pleasurable or painful per se. It just is. Right? So you become sentient when you not only have feelings, conscious experiences, but they start to feel good or bad. You start to want more or fewer of, of particular ones. And, and that for many philosophers is when you start to care about your experiences. They, they start to matter to you. And it starts to matter whether I give you more of the good ones or more of the bad ones, etc.

Well, one question is how to explain this, and then another question is how to justify it. I think there are good explanations of why we divide animals into these categories. In my view, there are not particularly good justifications of why we divide them into these categories. But as far as the explanations are concerned, there are different factors involved and a little bit of a spectrum based on, on these factors. So, for example, we tend to extend more attributions of consciousness and moral status to animals when they look and act like us. So, for example, animals with large bodies, large heads, large eyes, symmetrical features, furry skin instead of scaly skin, four legs instead of six or eight legs, etc., etc., etc. We, we tend to see them as conscious more, we tend to see them as valuable more, in part because they look like us and they act like us. Also, of course, if animals are near us and we interact with them day-to-day, like cats or dogs in our homes, to a lesser extent, urban animals in our cities, then they might be more top of mind and, and we might be a little bit more likely to see them as fellow creatures or co-citizens in some sense. And then, of course, a lot depends on how we socially classify animals. When, when we socially classify them as companions, then the role we give them in our lives is to be a friend or a family member. And, and for that relational reason, we, we might, we might care about them more, attribute more consciousness and, and moral status to them. And, and then, of course, the reverse is true for animals with the reverse features. When animals look very different, for example, small bodies, small heads, small eyes, scaly skin, multiple, multiple legs, or when animals are really far away, or when animals are socially classified as commodities, for example, if, if the, the use we give them in our lives instead of being a friend and family member is being our food or being a source of new knowledge or being a source of new meds or ingredients, then that all creates an incentive to look the other way, to see them as less conscious, to see them as less morally significant. And so generally our intuitions about animals are going to be shaped by all of these cognitive and social and political and economic and, and ecological factors. And this is why we do tend often to attribute more significance to a cat or a dog than to a cow or a pig, even though they are all mammals, often, you know, with comparable cognitive and behavioral abilities and familial bonds and, and so on.

That's something I, I, I've been vegan for getting on about 14 years now, I guess. I have no idea why. I have no idea why I, why it took me so long to, to sort of figure out, hey, this doesn't make sense. And, and for me, for me it was a lot of it was around environmental considerations and things. But then lastly I sort of learned about it. Do you, do you, do you see a day where we look back and think, wow, people used to eat meat and fish, that's crazy? Do you think we, think we'll get to that point, or is that just wishful thinking?

Well, I, I really think that we do and we must get to that point. And, and this is in fact one of the most common examples that people point to when they try to imagine what our successors will look back at, at our behavior and, and feel horror about, right? They, they will look back on the fact that we factory farmed animals, that we killed more than 100 billion invertebrates, more than a trillion invertebrates in captive settings per year, and, and even larger numbers of, of animals in the wild per year, even when plant-based alternatives were available to us. They, they will look back on that behavior with, with horror. Now, I think that is not a foregone conclusion. That is a, a possible future and, and one that we can and should work towards. And I think the more that we work towards it, the more likely we make that future. And, and that is a good goal to have. But, but nothing is guaranteed here. So, so I think imagining that future can be a, a possible motivation for doing the work that will make the future more likely, but it should not be a source of reassurance that this future is somehow inevitable and we can just sit back and wait for it to happen.

Oh, definitely. So I think if the question is how can we learn from our history and ongoing relationships with animals to make predictions about our future relationships with AI systems, then I think there definitely are lessons that we can learn. Now, all of them have to be taken with a grain of salt because AI systems are not animals. This is a very different context, but there are certain broad similarities that, that might be the case. So, for example, there are conversations developing now about whether AI systems are conscious, sentient, agentic, otherwise morally significant. And there is a knee-jerk reaction that some people have to think, of course they are not, because look how different they are and, and look how different their behavior is. Look how different their internal structures are. Look how different their development is than, than ours. And this is, of course, a reassuring thought as well because it makes us feel more comfortable with our current development, deployment of them, our current uses of them. But we know from our history of animals that there was a time not too long ago that many experts had exactly those reactions about non-human animals, even evolutionarily very close ones like other primates and mammals and, and vertebrates. And then in the fullness of time, we came to appreciate that the right kind of careful scientific and philosophical research could actually give us information about this and pointed towards different pathways towards their own distinctive kinds of consciousness and sentience and agency and moral significance. And, and so we've learned over time that we should give them the benefit of the doubt. We should actively try to understand their capabilities and their interests and, and then we, we should treat them with an appropriate amount of protection. And unfortunately, because it took us so long to learn that lesson, we along the way also scaled up this global system of factory farming and industrial animal agriculture and other forms of animal use. And, and now these industries will take decades to dismantle even if we decide, decided collectively to start today because we waited so long to, to arrive at these conclusions. And, and so as we look ahead to our relationships with AI systems, I think we can anticipate that we likely will to some extent have that same tendency to assume that they are not conscious, sentient, agentic, morally significant because of their differences with us, to go ahead and scale up global uses of them, and then to gradually wake up to the reality that there could once again be different pathways to, to consciousness, sentience, agency, moral significance. And, and then, uh-oh, now that we realize this is a real possibility and that we owe them some consideration in decisions that affect them, it will once again take decades to, to dismantle these, these industries. And the stakes could be even higher in this case because AI systems will be more capable in, in some respects than other animals, more powerful in some respects than other animals. And they will understand our relationships with them and will to some extent be learning from and, and reflecting back our, our treatment of them. And, and so the, the stakes of, of these decisions could be significant not only for them but also for us, for their treatment of us or, or our relationships and how they affect both of us. And, and we can expect the same biases that have been operating with animals to operate once again with AI systems. When they look like us and act like us, when they use human language like chatbots do, we might be inclined to overattribute consciousness and sentience and agency and moral significance to them. But then when they look different and they act different and, and we use them for different purposes like just running algorithms in a blank box, then we might be likely to underattribute these, these features to them. So, yeah, I think a lot can, can be learned.

And just, just to emphasize, these are not just past problems, these are ongoing problems. There, there has been a certain amount of progress, but there still is a lot of racism, a lot of sexism, a lot of other types of human prejudice, discrimination, oppression. And even when we make a certain amount of progress, we are always one election away, one economic recession away, one generational shift away from regression to, to some past state. So, so progress is partial, progress is delicate, and, and yet as, as you say, we have at least made more progress with human oppression than, than we have done with, for example, our treatment of non-human animals. And, and we, we might, we might worry that the same will be the case with these, these novel systems, too.

Right. Right. Right. Yeah. That is, that is very big. But, but I like the big questions. This is, this is what we should be talking about. So I guess we can break this down into two questions. One is what are the responsibilities of AI companies around safety and welfare? And then another is how should they navigate the tensions that might arise between safety and welfare. So regarding what their responsibilities are, I mean, in the first instance we might say their responsibility is to effectively coordinate around a global pause on further AI development because this is just too dangerous and too fast moving and we should collectively stop here and adjust to what we have already created and then only in the fullness of time, slowly, when we are ready, if we are ready, continue. Now, that just should always be said. Now, with that said, since global coordination around a global pause is relatively unlikely, we can also talk about what their responsibilities are assuming that they do continue to scale up and, and develop and deploy these systems. And a general point to make here is that we should not wait until we are certain that a certain kind of harm will occur before taking reasonable, proportionate measures to consider and mitigate that risk. As various kinds of harms become more likely and, and more severe, then more significant precautionary measures could be warranted. And this is true both for safety and for welfare. So AI safety is broadly understood as the project of making sure AI can be safe and beneficial for humans and other animals. And then AI welfare is broadly understood as the project of making sure AI systems can also be safe and beneficial for the AI systems themselves, that, that if and when they have their own welfare capacities, their own welfare interests, that they can have good lives too. And, and that brings us to your second question because as you noted, if humans and animals and AI systems themselves are all stakeholders for AI governance, then there could be tensions and trade-offs between our different interests and needs and vulnerabilities. So, for example, a lot of what we currently do in AI safety involves interactions with AI systems that at least would raise moral questions if we interacted with fellow humans or other animals in those same ways. So, for example, standard techniques involve boxing. This, this means putting AI systems in constrained environments to limit their freedom to, to protect us from what they might do if they had more freedom. That would be seen as a kind of like constraint or imprisonment if, if they had interests like ours. We also have interpretability. This means basically reading their behaviors and reading their minds to understand their thoughts and feelings and, and behaviors and inclinations. But this could be seen as a kind of very intimate surveillance if, if they had interests like ours, right? Other, other examples include often limiting their situational awareness or actively deceiving them in order to test how they would behave in different circumstances. This would be kind of like a, a form of systemic deprivation of information or mass misinformation if, if they had interests like ours. And then, of course, various training techniques could cause suffering if they had the ability to suffer. A willingness to unplug them if they start to seem not aligned with us, that would be like killing them prematurely. And making them have thoughts and feelings that align with ours, that could be seen as a kind of brainwashing, right? In, in general we think, no, if, if we matter, if we have our own interests and, and moral significance, then we should have a certain amount of agency and autonomy over what kind of thoughts and feelings and goals we can have, when we share those with others, what types of goals we can set and pursue given our, our, our thoughts and feelings. And, and if we deprive AI systems of those same privileges, then it at least raises moral questions. So, so right now my main recommendation to AI companies is simply take these issues seriously. Acknowledge that these are real questions. Start assessing AI systems for safety-relevant issues and welfare-relevant issues. And start preparing policies and procedures for treating humans and animals and AI systems with an appropriate amount of moral concern. Because if we do all of that together, acknowledge the issues, assess the issues, and prepare the policies and procedures, that at least gives us an opportunity to, to try to find co-beneficial solutions that work for everyone, humans, animals, AI, where those are available, and then to navigate tensions and trade-offs thoughtfully where co-beneficial solutions are not available. So, so that right now is the main advice I have.

Well, I definitely think that there are downsides to, to doing it, and both intellectual honesty and practical rationality requires acknowledging those downsides, not because they should cause us to not consider it, but rather because they should cause us to consider it thoughtfully. So, so one downside just is what you mentioned. It, it does slow companies down. It, it gives them a disadvantage in, in the global arms race to develop artificial general intelligence. And there are real practical and political stakes to that because some companies are in the race because they fear that unless they win, some other more reckless company with worse values and priorities will win instead. And, and there, there is, I think, a reasonable concern that absent effective global coordination, it would be a mistake for the companies with a safety and welfare mindset to unnecessarily hinder their ability to really compete in, in this global race. So, so that is one consideration. Another related consideration is that while there likely are many co-beneficial solutions available, or, or at least possibly co-beneficial solutions available, it is a reality that everything takes money, everything takes time, everything takes personnel. And to whatever extent you invest money and time and personnel in, for example, AI welfare research and AI welfare policies, that at least potentially is money, time, personnel you are not allocating to AI safety, biosecurity, other issues that really matter too. And, and I do think that we should acknowledge that these are issues and, and that there are other issues too. As I say, I think the benefits of doing this work thoughtfully very significantly outweigh the, the benefits of not doing it at all. But I think what, what empowers us to do it thoughtfully is acknowledging those downsides and then thinking, okay, how can we undertake the work in, in a way that minimizes these downsides?