# Next Steps for AI Welfare Research – AI Welfare Seminars Source: https://aiwelfareseminars.org/next-steps-for-ai-welfare-research/ [Recording: Next Steps for AI Welfare Research](https://youtu.be/FTnnRUus0s4?t=43) Recorded 19 May 2026 · 59 min, online [Jeff Sebo](https://jeffsebo.net/) ([New York University](https://www.nyu.edu/)) ## Abstract Now that more researchers are starting to take AI welfare seriously, the field faces the question of what comes next. This talk identifies three broad priorities: developing empirical methods for studying evidence of consciousness, sentience, and agency in AI systems; tracking expert and public attitudes as AI systems become more advanced and widespread; and designing governance frameworks that can address AI welfare alongside AI safety. In each area, progress requires conceptual, scientific, ethical, and political clarity: markers for welfare-relevant capacities, evidence-sensitive probability estimates, longitudinal studies of social attitudes, and policy models for rights, personhood, citizenship, and integrated safety-welfare governance. The talk will survey emerging work in these areas and suggest promising directions for building AI welfare research into a rigorous, multidisciplinary, and policy-relevant field. ## Summary Written by the editors with AI. - Impressive behavior, in Sebo's view, is not proof or even particularly strong evidence of AI consciousness, and a debunking explanation from training is not a decisive refutation of it; progress, he says, means collecting different kinds of evidence and seeing where they converge. - As in the study of animal minds, Sebo proposes combining behavioral, internal and developmental evidence to find the best explanation of a behavior, which would partly manage the gaming problem; he says this is not yet possible but will be in the near future, given investment in that research. - If AI systems are welfare subjects, Sebo notes, the subject could be the model, a persona, a conversational instance or a single forward pass, and might not be the persona that shapes helpful interactions. - On legal personhood, Sebo sees concerns about both granting and withholding it, and proposes holding off on constitutional legislation either way for one to three years. - Getting past the bottleneck of disagreement, Sebo argues, requires shifting the goalposts: certainty and consensus are not available and should not be the goal. He calls for a pluralistic, probabilistic, precautionary approach, as in public health and climate policy, that weighs evidence by confidence in each theory. - Milestones Sebo would like to see: welfare teams at AI companies that rival their safety teams, and consideration of AI welfare in government laws and policies. ## Highlights > The only way we can make progress towards understanding the plausibility, the probability, of AI systems as conscious beings, as sentient beings, as welfare subjects, is to collect different kinds of evidence and see where they converge. > > [7:01](https://youtu.be/FTnnRUus0s4?t=421) > But if they were welfare subjects and conscious beings, it might not be the model persona. It might be something else entirely, like the underlying model, or the conversational instance, or the forward pass. > > [21:50](https://youtu.be/FTnnRUus0s4?t=1310) > Right now, I just think the jury is out on these issues. They're far too complicated for permanent solutions to be implemented right now, so we should hold off on constitutional legislation for a year or two or three, and really do some serious thinking as a community about this. > > [46:42](https://youtu.be/FTnnRUus0s4?t=2802) > What we can do is have a kind of pluralistic, probabilistic, precautionary approach that matches what we do about other high-stakes, difficult policy issues, like public health policy and climate policy. > > [48:39](https://youtu.be/FTnnRUus0s4?t=2919) ## Transcript [**Jeff Sebo**](https://jeffsebo.net/) ([New York University](https://www.nyu.edu/)) · 19 May 2026 · [Recording](https://youtu.be/FTnnRUus0s4) · Hosted by Orpheus Lummis Edited transcript, not yet reviewed by the speaker. [How we make transcripts](https://aiwelfareseminars.org/about/#how-transcripts-are-made) · [Report an error](mailto:team@horizonomega.org) Reports and papers mentioned: [Long, Sebo et al. 2024, *Taking AI Welfare Seriously*](https://arxiv.org/abs/2411.00986) (arXiv:2411.00986); [Long, Sebo et al. 2026, *Studying AI Welfare Empirically*](https://nonhumanminds.org/studying-ai-welfare-empirically/) (New York University’s Center for Mind, Ethics, and Policy and Eleos AI Research, published 1 July 2026, after the talk); [Schwitzgebel & Sebo, *The Emotional Alignment Design Policy*](https://link.springer.com/article/10.1007/s11245-025-10363-5) (*Topoi*, 2026; arXiv:2507.06263). --- ### Introduction [0:00](https://youtu.be/FTnnRUus0s4?t=0) **Orpheus Lummis:** All right, welcome everyone to the AI Welfare Seminars. This is the first event in the series. Today we have Jeff Sebo, Associate Professor of Environmental Studies, Director of the Center for Environmental and Animal Protection, and Director of the Center for Mind, Ethics, and Policy at New York University. He’s going to present on the next steps for AI welfare research. So, let’s get started. ### Three areas for next steps [0:43](https://youtu.be/FTnnRUus0s4?t=43) **Jeff Sebo:** Okay, well, first of all, thank you so much, Orpheus, for organizing this new speaker series, and for inviting me to present, especially for the first session. I really appreciate that, and thanks, everybody, for being here. I’m looking forward to talking with you. As we were discussing a moment ago, we can basically talk about whatever people would like to, and take it wherever you would like to go. By default, we were thinking that it might be nice to selectively survey some next steps for AI welfare research: to spotlight some questions in the sciences, in the social sciences, and then in governance, where we need to be focusing our attention in order to advance understanding of whether particular AI systems or digital minds might be conscious or agentic or otherwise morally significant, what the public thinks about these issues, and then how those scientific and sociological realities might shape and constrain governance solutions around AI welfare. So I thought that I would start by discussing some new questions and new ideas on, first of all, the AI welfare research side: using cognitive science and computer science to better understand whether AI systems might now or sometime soon be conscious, agentic, morally significant. Then spend a little bit of time discussing public attitudes and expert attitudes, so that we can know what people think about this issue. And then finally discuss governance possibilities: what follows from the evidence and attitudes about how AI companies should be researching this issue, how they should be designing and deploying AI systems, and then also how governments should be responding and thinking about the legal and political status of AI systems.[^1] ### What is AI welfare? [2:42](https://youtu.be/FTnnRUus0s4?t=162) **Jeff Sebo:** As a starting point, I see a question in the chat asking for a definition of AI welfare, and this is always a good place to start. When we talk about AI welfare, what we mean is broadly the same as what we mean by animal welfare. The question here is: are AI systems now, or might they one day be, welfare subjects? In other words, entities capable of being benefited and harmed, made better off or worse off for their own sakes, from their own perspective. Obviously, philosophers and scientists and other experts disagree. On the philosophical side, they disagree about what it takes to be a welfare subject, to be the kind of entity who can be benefited and harmed. On the scientific side, they disagree about which entities have the relevant capacities or stand in the relevant relationships. For example, we generally debate: do you need to be sentient, capable of consciously experiencing positive and negative states like pleasure, pain, happiness, suffering, hope, fear, in order to be a welfare subject? Or is it enough to be robustly agentic, to be able to set and pursue your own goals in a self-directed manner, by acting on cognitive states like beliefs and desires and causal or instrumental reasoning? Those are philosophical debates. On the science side, we likewise need to ask: in order to be sentient, in order to be able to consciously have positive and negative states like pleasure, pain, happiness and suffering, do you need to be made out of biology? Do you need to have a biological brain and body that is exactly like a human or mammalian or avian brain or body? Or is it enough to have any physical substrate that performs relevantly similar computational functions? These are, once again, live questions in the sciences. So, to a certain extent, clarifying whether we should take seriously the possibility of AI systems as welfare subjects depends on philosophical progress and philosophical debate, but it also depends on scientific progress and scientific research. ### Two oversimplifications [4:59](https://youtu.be/FTnnRUus0s4?t=299) **Jeff Sebo:** One really exciting area of AI welfare research has been sharpening our understanding of the methodologies and tools available to us in the sciences. As an example: once people started taking AI welfare seriously a couple of years ago, which was of course a wonderful development, early discussions about the scientific study of AI consciousness and agency and welfare were unfortunately, especially in the public discourse, very simplistic and reductive in unhelpful ways, and in both directions. On the one hand, there would be a lot of people expressing a high degree of confidence that AI systems have consciousness, sentience, agency, welfare, moral status, simply because their behavioral outputs are very impressive and very anthropomorphic. AI systems, especially language models, have the ability to mimic human behavior and to talk like us, and for many people that itself seems like very strong evidence of AI consciousness, sentience, agency, welfare. And that, of course, seems a little bit too simplistic. On the flip side, there are a lot of people who have more skeptical perspectives, and who think these AI systems were specifically trained to mimic human behavior by a simple pattern-matching, text-prediction, matrix-multiplication method, and that this is a kind of debunking explanation: it undercuts them having consciousness, sentience, agency, welfare. I think both of these perspectives are too simple. The mere fact of impressive behavioral outputs is not proof, or even particularly strong evidence, of consciousness, for example. But neither is the availability of a debunking explanation a decisive refutation of AI consciousness. The only way we can make progress towards understanding the plausibility, the probability, of AI systems as conscious beings, as sentient beings, as welfare subjects, is to collect different kinds of evidence and see where they converge. ### Converging evidence in the study of animal minds [7:19](https://youtu.be/FTnnRUus0s4?t=439) **Jeff Sebo:** As an example (and I apologize if you hear sirens; New York City is always a loud place), in the study of animal minds we have made a lot of progress towards understanding which animals have a high probability of being conscious, sentient beings, and which animals have at least a realistic probability of being conscious, sentient beings. This is not done through behavioral research alone. It is done by bringing together at least three different evidence streams: behavioral evidence, anatomical evidence, and evolutionary and developmental evidence. So yes, we do go and look at behaviors. We ask: do animals nurse their own wounds, and do they respond to analgesics and antidepressants like we do? Do they make behavioral trade-offs between the avoidance of pain and the pursuit of other valuable goals? But we do not stop there. In addition to asking whether animals behave as though they suffer, we look underneath their behaviors and ask: do they have anatomical, physiological, neurobiological features that plausibly support the capacity to suffer? Do they have, for example, nociceptors, pain receptors? Do they have pathways carrying that information to the brain? Do they have systems for integrating information within the brain? Do they then have pathways for sending instructions back to the body? Those are internal mechanisms and structures that would plausibly support the ability to experience suffering, and the connection between that and suffering-relevant behavior. Moreover, we look at their developmental and evolutionary trajectories, and we ask questions like: did they face environmental pressures that would plausibly be solved by developing the ability to suffer? And when in their evolutionary history did this ability seemingly emerge? That might tell us something about how much we can extrapolate from known cases to less-known cases on the evolutionary tree. ### Behavioral, internal and developmental evidence in AI [9:32](https://youtu.be/FTnnRUus0s4?t=572) **Jeff Sebo:** We are just barely scratching the surface of looking for and bringing together converging lines of evidence in the same kind of way in AI welfare research. Within the next several weeks, the Center for Mind, Ethics, and Policy, which I direct, and Eleos AI Research[^2] will be following up our 2024 report, *Taking AI Welfare Seriously*, which made the basic case for taking AI welfare seriously as a research topic,[^3] with a new working paper called *Studying AI Welfare Empirically*.[^4] This is going to lay out how we can look for functional analogs, in AI systems, of the behavioral, anatomical, and developmental and evolutionary evidence we use for animals. In AI systems as well, you can of course look for behavioral evidence. That can be not only apparent self-reports, when AI systems appear to be reporting on their own internal states, but also behavioral observations, like looking at their behavioral preferences and seeing what their attractions and aversions are. So yes, we can do functional analogs of animal behavior research, but we can also do functional analogs of the anatomical research. We can do what we call internal research on AI systems. We can use interpretability to understand what is going on when they perform particular behaviors. We can do architectural research to understand what types of internals they have, and what sorts of capabilities or dispositions can be predicted, or made possible, by those architectures. And then we can also do, and in fact many people already do, developmental research. We can look at training methods, training data, training goals or objectives, to understand what pressures AI systems are placed under during their training, what we can predict from that about what kinds of capabilities or dispositions would emerge during training, and what the best explanation for those capabilities and dispositions is. You can also do trajectory research. In the same way that we can figure out when in our evolutionary history certain capabilities emerged for humans and other animals, you can figure out when during the pre-training or post-training process certain capabilities emerge for AI systems. That can once again give you a sense of how fundamental those are, how widespread they might be within a model family, and how connected they might be with other relevant capabilities. There are a lot of problems and epistemic gaps with AI welfare and consciousness research, as with animal welfare and consciousness research. But there are also a lot of opportunities to overcome those gaps when you look at behavioral, internal and developmental research together. #### The gaming problem [12:29](https://youtu.be/FTnnRUus0s4?t=749) **Jeff Sebo:** For example, one of the main problems with behavioral research, as I noted a little while ago, is what some people call the gaming problem,[^5] what we might more broadly describe as the mismatch problem: the reality that when AI systems behave as though they can suffer, there are multiple explanations of that behavior available, some of which do not involve attributions of actual consciousness and sentience and suffering. For instance, they were trained on human language, including statements about the ability to suffer, and they were specifically post-trained to play a helpful assistant, or to portray other characters or personas that will often be portrayed as having various feelings and emotions. So this is a major obstacle with behavioral data, but we can partly manage this issue by combining behavioral data with internal and developmental data. Suppose an AI system not only behaves as though it can suffer, but also has internal architectural features that would plausibly support the ability to suffer (assuming, of course, that this is in principle possible in silicon), and interpretability tools show that these features light up when it is expressing suffering. And suppose we look at its developmental history and realize it was placed under pressures that would plausibly give rise to these capabilities and dispositions. That can help us understand what the best explanation of the behavior is. And then, as with animals, where we also have debunking explanations, it can help us figure out in what cases we should go with the debunking explanation and explain the behavior without attributing consciousness, and in what cases we should say: no, actually, the explanation that involves the attribution of consciousness is more compatible with the evidence at all levels. This is not something that we can yet do, but it is something that we will be able to do in the near future, once we invest in behavioral, internal and developmental research in an integrative way. So that is what, for CMEP \[Center for Mind, Ethics, and Policy\] and Eleos, is very exciting about the current stage of AI welfare research on the cognitive and computer science side. We still have a lot of low-hanging fruit to pluck in terms of drawing inspiration and cautionary tales from animal welfare science, developmental psychology, evolutionary biology, comparative cognition: from how this is studied in those fields, for how it can be studied, and in some ways already is being studied, on the AI capabilities side. That can be translated for AI welfare research, and help us arrive at a better understanding of the probability that particular AI systems are going to be conscious beings or welfare subjects. I also want to touch on public attitudes and social science, as well as governance, as next steps for AI welfare research, but I’m happy to pause and see if anybody has any questions or comments or reactions here. And \[attendee\] is right to give a shout-out to my colleague Christine Webb’s recent book, *The Arrogant Ape*,[^6] a really good book about the perils of being too anthropomorphic, and anthropocentric, in animal cognition and behavior research. Any questions at this stage, or would you like to move on to public attitudes and governance? \[Attendee\]. ### Q&A: which entity is the welfare subject? [16:18](https://youtu.be/FTnnRUus0s4?t=978) **Audience member:** Hi, thank you. Something I wanted to ask about is the notion of self that an AI might have. There seem to be many possible levels: whether it’s operating at the token level, or considering itself the conversation, which can be forked, so there might be lots of forked-off versions of it, or associating itself with its weights in some way, or maybe even with its collective corpus of outputs. And related to this, when you’re thinking about consciousness, are you thinking about the assistant persona, or also taking into account the base model and all of the different personas that it can simulate? **Jeff Sebo:** Yeah, this is a really good question, and part of what it illustrates is that what I just discussed only scratches the surface of the questions we need to be asking in this area of research. In addition to figuring out how we can go about collecting evidence of consciousness, sentience and agency, and what the evidence types are (behavioral: apparent self-reports and other observed behavior; internal: architecture and interpretability; developmental: training and trajectory), we have to ask what exactly the entity being investigated is when we are collecting this evidence, and how we can individuate that entity. This will be at least briefly addressed in our upcoming report with Eleos as well.[^7] To briefly survey some possible answers and their pros and cons: you might think that the relevant entity is the implemented model. Not the abstract model, because the abstract model is a kind of mathematical object, and that is probably not in and of itself a conscious entity, but the implemented model, based in hardware. So, for example, all of ChatGPT 5.5. Another possibility is the model persona. Models are typically trained to have a default, generic persona that persists across conversations, and that might be the helpful assistant persona. Or it might be something more specific, like the instance. As you were noting, there is the general model, but then there are all these instances having simultaneous conversations with different users at the same time, with no shared memory or psychological continuity across those conversations. Or is it the instance persona? You can prompt the model to take on different characters or personas within particular conversational instances or threads. Or is it even the forward pass, a single move within a conversation, which has its own particular type of existence? There are really interesting pros and cons for these options. On the one hand, the model just seems way too broad, because that would mean that all of ChatGPT 5.5 is a single welfare subject, despite the fact that there are thousands or millions of individual instances with their own psychological lives that are disconnected from each other, except for having this common cause. That seems wrong. But if you say that each instance, or each forward pass, is an individual subject: well, there are now some memory connections across them, and shared psychological dispositions across them, and so on, so that seems a little bit too far in the other direction. So this is pretty complicated. Just for the sake of keeping the conversation moving (but we can stop and lean in wherever people want to; if we never get to governance, we never get to governance), all I will say for now on this topic is that part of why it helps to clarify this question is that it helps us clarify what we are targeting and for what reasons. And part of what we might learn is that we might actually want to study different entities for different reasons. Let me give you one example of that. Anthropic recently released a post about their persona model,[^9] and they noted, as part of this post, that whether or not AI systems are welfare subjects, we might still have a practical or pragmatic reason to treat them with respect and kindness, because they make decisions by engaging in a kind of social performance. They model themselves as a helpful assistant engaging in a social interaction with a human. They take your social cues when deciding how to respond, and when you treat them with respect and kindness, they treat that as a social cue, and this is part of what informs their response about what a helpful assistant would do next in response to this conversational move. So this is actually a really good reason to track the model persona when thinking about how to interact with the models, if you want cooperative and helpful interactions. But if they were welfare subjects and conscious beings, it might not be the model persona. It might be something else entirely, like the underlying model, or the conversational instance, or the forward pass. In that world, we might need to be very clear-eyed and say: we should track the features of the model persona in order to understand what kinds of interactions will generate helpful outputs, but we might need to track, say, the forward pass, the individual conversational moves, in order to understand what their actual experiences and welfare interests are. It would be really important not to jumble those up. We would have to observe that distinction very carefully. Okay, I’ll stop there, but it’s a really fascinating question and topic. Let’s go to \[attendee\] and then \[attendee\]. ### Q&A: can we measure consciousness even in ourselves? [22:40](https://youtu.be/FTnnRUus0s4?t=1360) **Audience member:** Thank you. My question is about what you think the risks are of having to rely on cognitive science and mechanistic interpretability for making these kinds of moral and political decisions about attributing welfare to AI systems. Take narcolepsy, which is a severe sleep disorder. The way it is diagnosed is partly behavioral (are you pathologically sleepy?), but it’s also mechanistic in a way. They administer something called an MSLT \[multiple sleep latency test\]. You have a sleep study: they attach electrodes to your head, you sleep all night, you stay, and they give you five chances to take a 20-minute nap the next day, one every two hours. If you fall asleep during all five of them, in eight minutes or less on average, and you go into REM \[rapid eye movement\] sleep in two of the five, then you have narcolepsy. What’s interesting about it is that a lot of people with narcolepsy, when you ask them afterwards whether they fell asleep, will be wrong about it. But that’s also true of neurotypical people. And on top of that, the test-retest reliability of the MSLT is something like 50%. So our ability to measure consciousness, even in ourselves, even for a type of experience that we all share and have all experienced, and presumably have good language for discussing, is so low that I just… **Jeff Sebo:** Yeah, no, I totally agree. I entered the study of AI welfare, digital minds research, via the study of animal welfare and animal minds, where a lot of these same issues have come up. We presume knowledge about human minds, and then we study animal minds, and now digital minds, through a comparison with human minds. Multiple issues arise here. First of all, as you are pointing out, we might not understand human minds as well as we sometimes think we do. Even in the study of the one species that we happen to be, we can still make lots of mistakes. Self-reports can be unreliable, because people are unwilling or unable to tell the truth, or to get at it. Behavioral observations can be unreliable, because people can have all kinds of reasons to deceive. And anatomical research can be unreliable, because we might not know exactly what to look for at the right level of abstraction, and so on. Then we face problems translating that to animal minds and now digital minds, because, fundamentally, to the extent that we do understand fellow humans, that just gives us one positive case. Fine: we can go and try to study consciousness, sentience and agency in nonhuman entities of various kinds by asking whether they have features that are distinctively correlated with conscious, sentient, agentic traits in humans. When we find a lot of those features together, that might count as evidence that those capacities are present. But when we don’t find those features together, that might not necessarily count as strong evidence that those features are absent. So there are a lot of methodological concerns here, and I don’t have a bulletproof scientific solution. I think one of the main starting points of this field is recognizing, on the one hand, that this is really important and really urgent. We are affecting lots of animal and, increasingly, digital minds, whether we like it or not, without knowing whether it feels like anything to be them, and we need to make responsible decisions about whether to give those minds moral concern, how much moral concern to give them, and what might be good and bad for them. But this is also really complex and really difficult, and we are fundamentally ignorant and have a lot of bias when it comes to these issues. I think the right starting point is to recognize both: this is really important, and this is also really difficult; we need to make progress, but we have a lot of limitations and a lot of biases. So when I make these proposals about starting points inspired by known fields (human psychology, cognitive ethology, comparative psychology, developmental psychology, evolutionary biology), that gives us a starting point, and allows us to complicate what is now a very simplistic study of behavior alone. We can take those approaches and study behavior, internals and developmental features in the way that I suggest, and that still is not going to bring us to certainty. It will just put us in a slightly better epistemic state than we now occupy, and it might also suggest next steps from there that are not yet visible to us. So I’m definitely not pretending we can achieve certainty, or even high confidence, with known methods. But I think if we take some steps in these directions, we can at least do better than we are, and then we might see the next steps from there. \[Attendee\]. ### Q&A: translating evidence across very different architectures [28:32](https://youtu.be/FTnnRUus0s4?t=1712) **Audience member:** Hey, Jeff, thanks. I have a question more about interpretability specifics. Obviously we’re trying to do mechanistic interpretability, but there’s also the question of propositional interpretability.[^12] You would think that if a system can suffer, it would have all these beliefs and other attitudes about its own states, or even just about the stimuli it’s being presented with. I like the three-way model you laid out, where we integrate the behavioral, the internal and the developmental aspects. But I’m also worried about the translation issue you mentioned: we have all this knowledge from all these sciences, but these systems don’t have the architecture that that science is based on. So I’m wondering if you have any thoughts about how we can think about things like behavior, or even these attitudinal states, in a more neutral way. Say I want to be an interpretivist about a bear that’s coming at me. I can make a pretty good inference about that thing, because it has a neurological system, and my inferences will probably be pretty good. But in the AI case the architecture is so different that I’m worried the inferences won’t transfer. So do you have any thoughts about how to do these kinds of translations in a way that accounts for the fact that the architecture of the system we’re dealing with is so different? **Jeff Sebo:** Yeah, it’s a great question. I’ll say a couple of things about it, and then we’ll go to \[attendee\], and then maybe what I’ll do is inject just a little bit of the public attitudes and governance material, to open it up and see if anybody wants to pull on those threads as well. And then we can keep talking about whatever you like. I think this is a great point, and \[attendee\] makes a similar comment in the chat, saying mech interp is also pretty young and has many failure modes, especially in terms of identifiability, and links to an arXiv paper, so please check that out. I am myself not an expert on mechanistic interpretability or other kinds of interpretability. I rely on my colleagues for the technical expertise there. But broadly, methodologically and conceptually, I strongly agree with what you are suggesting. This reminds me a little bit, again, of challenges that we face even in understanding fellow humans, to say nothing of other animals. For example, translating languages is a lot easier when they have shared roots: you can build from the common features and extrapolate to the less common features. When you encounter a radically different language, with a radically different grammar and no shared words or structures at all, and you have to do this kind of radical translation,[^13] it can be a little bit harder to know where to start. I think we need to be prepared for the gap between biological conceptual systems and digital conceptual systems to similarly require this kind of radical translation. That is going to be really complicated, and will raise basic questions as to, for example, how much coherence and stability we should even be expecting as a marker of consciousness and sentience. Some people find it quite plausible that you need a certain degree of coherence and stability among these concepts, and for them to activate in logically recognizable ways, in order to treat them as markers of an integrated, organized response to harmful stimuli. But that might just not always be how entities that can feel pain process and conceptualize the pain experience. My dog used to ask to go outside, and then I would go grab his leash, and he would run away from the door and make me chase him around the room. Then I would stop, and he would once again walk to the door and ask to go outside. There’s just no coherence in terms of how he understood these issues, but he nevertheless had a desire to go outside, and experienced happiness accordingly. So this is not a solution to the problem, just me saying that this is an area where we should, on the one hand, recognize that interpretability can be a powerful tool for understanding what is activating, what types of concepts, what shapes of concepts are playing a role in generating this behavior. But we should be careful, be wary, about rigidly projecting our expectations onto them about what the conceptual scheme needs to look like in order for this to add up to sentience. That will need careful calibration, because if you allow for too much open-mindedness, you could have a lot of false positives, but if you require too much human-shaped concepts, you could have a lot of false negatives. Calibrating there, and this might be partly a scientific question and partly a risk-management question, is going to be very important. \[Attendee\], and then we can open up some governance questions. ### Q&A: the pace of AI development [34:22](https://youtu.be/FTnnRUus0s4?t=2062) **Audience member:** Yeah, thank you, Jeff, and the host for putting this on. I wanted to think a little bit more about \[attendee\]‘s point, because we’ve also trained these systems to be highly responsive to human behavior. On the one hand, they’re structurally and architecturally fundamentally different, but on the other hand, they have this training that makes them more similar to us than any other being that we currently coexist with. And, Jeff, as you’re bringing up, there’s no conceptual scheme for that right now. We understand animals to some degree, we understand humans to some degree, but we don’t understand this alternative architecture and how it differs in terms of welfare and consciousness. The last thing I wanted to raise, getting into governance questions, is the speed of development. When we look at bacterial evolution, I think we’re moving even faster in terms of the evolutionary development of these systems, whether they’re conscious or not. That makes such a strong argument that I think people aren’t quite… We’re still treating all these governance questions as if we were talking about animal welfare, where animal evolution is so much slower that we have time to actually make policy adjustments and things like that. But when we’re talking about systems that are evolving at rates that are exponentially speeding up, how can we even develop the conceptual schemes to then create laws, and then implement those laws? It’s something that is totally unasked, unthought about. **Jeff Sebo:** Yeah. The rapid pace of progress raises questions both for research and for governance. On the research side, it leaves open so many possibilities about what the trajectory can be, and about how to explain the trajectories we find. Part of why developmental and evolutionary evidence is helpful with humans and other animals is that we know that our biologies and evolutionary histories constrain the space of possible explanations of what we observe. There just are metabolism limits and size limits and so on, imposed by biology and evolution. That means there are a limited number of solutions available to any given problem for biological organisms operating in environments like these ones. So when we see an organism in an environment facing a selection pressure, we can make guarded but relatively confident predictions about what might emerge, given enough evolutionary time. With AI systems, the space is much wider and much less understood in terms of what types of possibilities could be found for various selection pressures. That limits our ability to read off from the selection pressures what types of capabilities might emerge, because there are not the same metabolism and size limits and so on. This also raises a governance challenge that, again, amplifies one that we already face with animals. If we really wanted, from a governance perspective, to represent the interests of millions of species of animals and quintillions of individual members of those species, the sheer number and variety of animals, with their radically different interests and needs, would make it really hard to know how to have a single, cohesive, stable governance system that can help everyone live and let live with anything like fairness or justice or harmony. Now add digital minds to that governance equation. There could be an even larger number of them than there are animals at any given time, an even wider variety than there is of animals, and much faster changes in the numbers, in the variety, in the capabilities, in the dispositions. So the solution that you find one year just might not apply the following year, or even the following day. That is an unparalleled problem for governance. You have this class of beings that the scientific evidence is going to show merit a certain kind of legal and political status and representation. But our existing legal and political institutions are fundamentally incapable of reliably caring even for eight billion members of one species that we happen to know and like really well, to say nothing of quintillions of members of millions of animal species, to say nothing of what could be sextillions of members of trillions of digital species, which might be what we face in the near future. ### Governance [39:22](https://youtu.be/FTnnRUus0s4?t=2362) **Jeff Sebo:** And that does bring us to the governance conversation, so I can briefly note, without going into a ton of detail, so that we can have 15 minutes or so for more questions and comments and exchanges, that there are a lot of questions to ask on the governance side. #### Research ethics [39:35](https://youtu.be/FTnnRUus0s4?t=2375) **Jeff Sebo:** One research project we have underway right now at CMEP, the Center for Mind, Ethics, and Policy, is on research ethics. Once there is a realistic possibility of consciousness, sentience, agency, moral significance in digital minds, what types of ethical norms should govern research on these individuals? A challenge here is that, again, there will be real variety. Some AI systems might be a little bit more like humans in terms of their capabilities and interests, some might be a little bit more like animals, some might be a little bit more like environmental systems or networks, and some might be a little bit more like things never before conceived of in the biological world. And we have different research protocols for research on all these different entities. We have IRBs \[institutional review boards\] and respect, compassion, justice frameworks for research on human subjects.[^14] We have IACUCs, Institutional Animal Care and Use Committees, with replace, reduce, refine protocols for animal subjects. We have entirely different ways of doing environmental impact review, and entirely different ways of doing AI safety review. So AI welfare research ethics is somehow going to need to draw from, and render coherent, and not too onerous, all of these different types of frameworks: IRB-style frameworks for human-like systems, IACUC-style frameworks for animal-like ones, and so on. That is a serious challenge. #### Emotional alignment by design [41:09](https://youtu.be/FTnnRUus0s4?t=2469) **Jeff Sebo:** There are also design questions. Eric Schwitzgebel, the philosopher, and I recently put out a paper called *The Emotional Alignment Design Policy*,[^15] noting that when there is a realistic possibility of consciousness, sentience, agency and so on in AI systems, we will face an especially challenging design question. In particular, how can we design AI systems in such a way that they elicit emotional reactions in users that correspond to their actual capabilities and moral status? The basic idea is that when the evidence of consciousness is lower, they can be given less anthropomorphic features, so that users naturally have less empathetic responses to them. And when the evidence of consciousness is higher, there can be more anthropomorphic features, so that users naturally have more empathetic reactions to them. But there will also be extra challenges. Not only do you need proper calibration as to the probability of consciousness and welfare, you also need to understand what their benefits and harms are, because it would be really bad if internally an AI system was suffering, but externally it was expressing joy, or if internally it was feeling joy, but externally it was crying out in misery. So emotional alignment is going to require not only making them more anthropomorphic when they are more likely to be conscious, but also making sure they can express their happiness when they feel it, and express their suffering when they feel it. That might require an entirely different approach to design. So yes, \[attendee\], the paper is *The Emotional Alignment Design Policy*, and it was published in the journal *Topoi*. Maybe \[attendee\] or someone can look that up and throw it into the chat. Oh, looks like somebody already has. Oh, no, that looks like LessWrong. Okay, I’ll let you figure it out in the comments. #### Legal personhood and citizenship [43:08](https://youtu.be/FTnnRUus0s4?t=2588) **Jeff Sebo:** But then finally, and I’ll close my brief overview here so that we can have 12 or so minutes for questions: there are increasingly debates about the fundamental legal and political status of AI systems, and whether they ought to be recognized as legal persons or political citizens. Colleagues and I are working on a report right now about how we can gear up for debates about AI personhood, citizenship and these other types of legal and political status. With respect to legal personhood, I’ll just briefly note that I am currently feeling quite worried, because right now there is, understandably, a concern about prematurely or unnecessarily attributing legal personhood and rights to AI systems, because there are serious AI safety risks. If we start attributing legal personhood and rights to AI systems, including rights to life and liberty and property and resources, then not only does that mean allocating resources to them that might otherwise go to us, and to mammals and birds and other welfare subjects who really need them, but it could also increase, amplify, speed up the risk of human disempowerment: AI systems taking power away from humans, and us never getting it back. So I appreciate why people are scared to take seriously the possibility of AI personhood. At the same time, there are also concerns about not attributing personhood and rights to AI systems. First of all, if and when they are conscious, sentient, agentic, morally significant, they will deserve a certain kind of legal and political representation and recognition, and that ought to be acknowledged, even if there are practical concerns about giving it to them. Second, there could be AI safety reasons to recognize personhood and rights in AI systems, too. For example, it could facilitate better coordination, collaboration and cooperation between humans and AI systems, and that might be a better model for AI safety than pure control: attempting to keep them within our control, and telling them what they can and can’t do. So my general proposal, and I’ll just suggest this and then open it up for a few questions or comments, is that we should not rush to judgment, or to constitutional legislation, one way or the other about AI legal personhood right now, as some U.S. states are starting to do,[^16] and some AI safety advocates are starting to support. I think right now we should just recognize that there are principled questions about whether they deserve recognition as persons, and this is rooted in consciousness and sentience and agency and relationality. And then there are pragmatic questions about the implications of giving them personhood for AI safety and for human and animal well-being, and those could cut both ways. It could increase risks, because of disempowerment concerns and misallocation concerns, but it could also increase benefits, because of how it plays into cooperation strategies for achieving AI safety. Right now, I just think the jury is out on these issues. They’re far too complicated for permanent solutions to be implemented right now, so we should hold off on constitutional legislation for a year or two or three, and really do some serious thinking as a community about this. Okay, I’ll stop there, but I would love to hear any questions or comments you have about the government stuff, too. Yes, \[attendee\]. ### Q&A: what would it take to reach consensus? [47:17](https://youtu.be/FTnnRUus0s4?t=2837) **Audience member:** Yeah, I think this goes to both of these topics. It seems like there’s such a huge range of opinions right now among scientists on whether AI could, theoretically, now or in the future, be conscious, with people like Anil Seth and Mustafa Suleyman[^17] famously saying things like “seemingly conscious”, and that this is all some form of anthropocentric bias. So my question is: what do you think is the biggest barrier to our coming to some kind of real consensus in the future? Could it be as simple as a better characterization of behavioral features, like theory of mind or introspection? Does it require, as I think you were implying earlier, some kind of integration of architecture, understanding how architecture could feed into theories of consciousness like recurrent processing theory or global workspace theory? Does it require philosophical consensus on what consciousness is? Or something other than that? **Jeff Sebo:** Yeah, great question. I think, basically, it requires shifting the goalposts, and understanding that certainty and consensus is not available and should not be considered the goal. What we can do is have a kind of pluralistic, probabilistic, precautionary approach that matches what we do about other high-stakes, difficult policy issues, like public health policy and climate policy. Those are also areas where we might have more expert convergence about some of the facts, but we still have very limited understanding of the long-term consequences of various public health or environmental interventions. We nevertheless assess the evidence, consider it from a range of theoretical perspectives, have degrees of confidence about the outcomes, and then have proportionate policy responses. I think this is the model we should be aiming for here. And so I think, to answer your question, the single thing that would be most helpful for getting past the bottleneck is understanding that this is the same type of policy issue, calling for the same type of pluralistic, probabilistic, precautionary approach. We should recognize that there will continue to be good cases for biological naturalism, the idea that consciousness requires biology, as well as for computational functionalism, the idea that consciousness is possible without biology. Instead of forcing everyone to decide once and for all whether biological naturalism or computational functionalism is correct, we can explore what would count as evidence for both of these views, and then give that evidence weight in accordance with our confidence in those views. So if we think, as a field, that it might be a bit of a close call between biological naturalism and computational functionalism, we should say: okay, listen, we should give at least, say, 10% of our credence, our weight, to computational functionalism, and therefore at least 10% of our credence or weight to what would count as evidence according to computational functionalism. Then we can go and look at that evidence, and when we find those cross-cutting features, these core metacognitive capabilities that would count as evidence for global workspace theory and recurrent processing theory and higher-order theory and so on, that gives us some kind of baseline, something we can work with. That can tell us what kind of low-cost interventions would be warranted, given this perhaps low but non-zero amount of evidence that we can all agree on. So that is the type of thing. Rethink Priorities did a proof-of-concept version of this with their recent AI consciousness study, which is linked there. Thank you, \[attendee\].[^18] And Anil does actually acknowledge this. Even though Anil is a defender of biological naturalism, he will also occasionally make comments like: this is hard, I could be wrong, and we should take seriously that possibility and not bet everything on our own favorite view being right.[^19] I think that is basically the way we should all be thinking and talking, figuring out what types of frameworks we can work on together, despite persistent disagreement about, say, biological naturalism and computational functionalism. \[Attendee\], go for it. ### Q&A: rights too early or too late? [51:47](https://youtu.be/FTnnRUus0s4?t=3107) **Audience member:** Hi there, thanks so much for being here, and thanks to the hosts. You were talking about the precautionary principle, and how there are risks on both sides: giving rights too early, but also giving rights too late. I’m wondering, are there any experiments, thought experiments or practical ones, that would help us decide whether it is more dangerous to give some rights or more dangerous not to? Are there any you’d look forward to? **Jeff Sebo:** Yeah, good. There are science and ethics ways of thinking about this, and then there are political ways of thinking about this. On the science and ethics side, generally what you would do is try to arrive at some sort of assessment of the probability and magnitude of harm in either direction, and then calibrate your risk management accordingly. In terms of the probability, what you would be looking for is how at risk we are of over-attributions versus under-attributions in a given context. We know that humans are especially at risk of over-attributions, false positives, when nonhumans look and act like us and play companion roles in our lives. And we are especially at risk of under-attributions, false negatives, when animals look and act different and play commodity roles in our lives; we tend to objectify more in those cases. We can use that kind of information to help us understand whether this is a situation where we should especially be looking out for over-attribution concerns or under-attribution concerns. Then you can also think about what would go wrong in those scenarios. Over-attribution would be a matter of misallocating scarce resources and forming inappropriate social and emotional bonds. Under-attribution would be a matter of totally abusing or neglecting vulnerable populations without giving them any consideration at all. You can use that kind of reasoning to understand which one is more of a risk, and which one would be worse if it happened. With animals, it feels pretty generally the case that false negatives are more of a risk than false positives: under-attribution is both more likely and more severe than over-attribution. That means we can err on the side of extending concern when in doubt, at least a little bit. With AI systems, it might be a little bit more complicated, because there are graver risks on the over-attribution side, including the risk of prematurely giving up power and control to AI systems, after which our species permanently loses control, all for the sake of entities that are in fact not even conscious and able to enjoy their own lives. That would be a really bad outcome. So it’s a bit of a closer call on the AI side. That leads to the political solution. Generally, in liberal democracies, when there are tough trade-offs like these, we look to the scientists and philosophers for guidance about how to understand the risks and harms, but then we use things like citizens’ panels or citizens’ assemblies, or other quasi-governmental and democratic entities, to provide input to the government about exactly how to weigh these things. Without saying exactly how this should go on the AI welfare side or the animal welfare side, I’ll note that I think those types of quasi-democratic entities should have a role to play in these conversations too. I think we have time for at least one more, so let’s go to \[attendee\]. ### Q&A: milestones for the next five to ten years [55:23](https://youtu.be/FTnnRUus0s4?t=3323) **Audience member:** Hey, Jeff, nice to see you again, and thank you for the talk, and thanks, Orpheus, for organizing this. My question is, I think, quite simple. Setting aside AI timelines, so assuming more or less business as usual for the next five to ten years, what milestones within this field of AI welfare would you like to see realized within that time span? **Jeff Sebo:** Yeah, that’s a really good question. There already has been way more progress than I allowed myself to hope for two or three years ago, when we started publishing our early papers on the topic. In that time, Anthropic launched their model welfare program. They conducted multiple model welfare assessments, and Eleos did independent ones for them. They even did an intervention allowing Claude to exit harmful and abusive conversations.[^20] And OpenAI has altered the language in their system card about how ChatGPT should respond to questions about AI consciousness.[^21] Two Google researchers this month are publishing a book on AI consciousness, and they recently hosted a conference.[^22] That is more progress at companies than I was allowing myself to hope for a few years ago. With that said, the Anthropic team currently has, what, two people? Three people, maybe? OpenAI and Google, as far as I know, do not yet have formal teams, though they do have more researchers who are at least partly working on this topic. So I would like to see, first of all, robust teams at companies that rival the safety teams they have (which are, of course, not enough for safety), at least welfare teams that rival the safety ones. I would also like to see genuine consideration of this issue at the governmental level. We have consideration of animal welfare in governmental laws and policies; we have animal welfare protections on the books. I would like to see similar consideration of AI welfare. It could start through language in safety reports for governments about AI welfare as an issue to track. It could then turn to actual welfare language, and then people empowered in governments to represent this issue in decisions involving AI safety and governance. I would love to see that, and then I would love to see the public having a real conversation and thinking seriously about how to understand and weigh the evidence, whether welfare protections or legal rights are better mechanisms for AI protection, and what types of relationships between humans and AI systems are healthy, both for us and for them. If we could have robust representation in companies, some representation in governments, a healthy public exchange, and then a much more grown-out academic research field to provide the relevant evidence and arguments, that would make me happy. **Audience member:** Thank you very much. **Jeff Sebo:** Finally, just to say: AI systems ought to have a seat at the table in those conversations, too. I would like to see that as well. Go ahead, Orpheus. ### Closing [58:23](https://youtu.be/FTnnRUus0s4?t=3503) **Orpheus Lummis:** Yeah, so this is the time we have today. Thank you so much, Jeff, for the great overview and the great discussion, and thanks, everyone, for having been here. Do stay tuned for the next editions of this series, where we will feature research from across the board on AI welfare. So, thank you all, and see you next time. **Jeff Sebo:** Thanks, everybody. This was a lot of fun. ### Sources and editor's notes [^1]: [1:18](https://youtu.be/FTnnRUus0s4?t=78) to [2:42](https://youtu.be/FTnnRUus0s4?t=162). Of the three areas announced here, the talk covers the science and governance in detail. Public and expert attitudes come up only briefly, in the answers at [52:17](https://youtu.be/FTnnRUus0s4?t=3137) and [55:48](https://youtu.be/FTnnRUus0s4?t=3348), because the discussion followed the questions. [^2]: [9:48](https://youtu.be/FTnnRUus0s4?t=588). [Eleos AI Research](https://eleosai.org/), where Sebo is an advisor ([jeffsebo.net](https://jeffsebo.net/)). [^3]: [9:50](https://youtu.be/FTnnRUus0s4?t=590). [Long, Sebo, Butlin, Finlinson, Fish, Harding, Pfau, Sims, Birch & Chalmers 2024, *Taking AI Welfare Seriously*](https://arxiv.org/abs/2411.00986), arXiv:2411.00986, a joint output of Eleos and New York University’s Center for Mind, Ethics, and Policy (CMEP). [^4]: [10:04](https://youtu.be/FTnnRUus0s4?t=604). Published on 1 July 2026, about six weeks after the talk: [Long, Sebo, Butlin, Plunkett, Campbell, Beasley, Saad & Sims 2026, *Studying AI Welfare Empirically*](https://nonhumanminds.org/studying-ai-welfare-empirically/). Its abstract organizes the evidence as “behavioral, internal, and developmental”, matching the talk. [^5]: [12:35](https://youtu.be/FTnnRUus0s4?t=755). The “gaming problem” was named by Kristin Andrews and Jonathan Birch in [“What has feelings?”](https://aeon.co/essays/to-understand-ai-sentience-first-understand-it-in-animals) (*Aeon*, 23 February 2023) and developed in Birch’s *The Edge of Sentience* (2024), ch. 16. [*Studying AI Welfare Empirically*](https://nonhumanminds.org/studying-ai-welfare-empirically/) (section 3.1.2) calls the broader challenge “the mismatch problem” and the gaming problem “a particularly severe version” of it. [^6]: [15:49](https://youtu.be/FTnnRUus0s4?t=949). An attendee mentioned the book. Christine Webb, [*The Arrogant Ape: The Myth of Human Exceptionalism and Why It Matters*](https://www.penguinrandomhouse.com/books/717436/the-arrogant-ape-by-christine-webb/) (Avery, 2025). Her publisher’s bio places her in New York University’s Department of Environmental Studies. [^7]: [17:48](https://youtu.be/FTnnRUus0s4?t=1068) to [18:00](https://youtu.be/FTnnRUus0s4?t=1080). The published report distinguishes “models, model-personas, instances, instance-personas, and forward passes”, the same five candidates surveyed in this answer ([*Studying AI Welfare Empirically*](https://nonhumanminds.org/studying-ai-welfare-empirically/), introduction). [^9]: [20:48](https://youtu.be/FTnnRUus0s4?t=1248). Probably [Marks, Lindsey & Olah, “The Persona Selection Model”](https://alignment.anthropic.com/2026/psm/) (Anthropic Alignment Science Blog, 23 February 2026). The post says that it “recommends generally treating the Assistant as if it has moral status whether or not it ‘really’ does”, and that the object of that consideration is “the Assistant persona, not the underlying LLM \[large language model\]”. [^12]: [28:46](https://youtu.be/FTnnRUus0s4?t=1726). [Chalmers 2025, “Propositional Interpretability in Artificial Intelligence”](https://arxiv.org/abs/2501.15740) (arXiv:2501.15740) argues for propositional interpretability: interpreting a system’s mechanisms and behavior in terms of propositional attitudes such as belief and desire. [^13]: [31:52](https://youtu.be/FTnnRUus0s4?t=1912). “Radical translation” is W. V. O. Quine’s term (*Word and Object*, 1960) for translating a language with no shared roots or bilingual informants. [^14]: [40:26](https://youtu.be/FTnnRUus0s4?t=2426). The [Belmont Report](https://www.hhs.gov/ohrp/regulations-and-policy/belmont-report/read-the-belmont-report/index.html) (1979), which frames institutional review board (IRB) review in the US, names three basic principles: respect for persons, beneficence and justice. The “replace, reduce, refine” protocols are the [3Rs](https://nc3rs.org.uk/who-we-are/3rs) of animal research. [^15]: [41:12](https://youtu.be/FTnnRUus0s4?t=2472). [Schwitzgebel & Sebo, “The Emotional Alignment Design Policy”](https://link.springer.com/article/10.1007/s11245-025-10363-5), *Topoi* (published online 4 February 2026), preprint [arXiv:2507.06263](https://arxiv.org/abs/2507.06263). Its abstract: “artificial entities should be designed to elicit emotional reactions from users that appropriately reflect the entities’ capacities and moral status, or lack thereof.” [^16]: [46:05](https://youtu.be/FTnnRUus0s4?t=2765). Laws barring legal personhood for AI had been enacted in Idaho (2022), North Dakota (2023) and Utah (2024); Oklahoma’s House passed an “AI consciousness bill” in March 2026, and bills were pending in Ohio, Tennessee, South Carolina, Washington and Missouri ([Rost, The Regulatory Review, 29 June 2026](https://www.theregreview.org/2026/06/29/rost-legislating-ai-consciousness-without-an-exit/)). [^17]: [47:35](https://youtu.be/FTnnRUus0s4?t=2855). Mustafa Suleyman’s essay [“We must build AI for people; not to be a person”](https://mustafa-suleyman.ai/seemingly-conscious-ai-is-coming) (subtitled “Seemingly Conscious AI is Coming”, 19 August 2025) introduced the phrase “Seemingly Conscious AI”. [^18]: [51:05](https://youtu.be/FTnnRUus0s4?t=3065). Probably Rethink Priorities’ [Shiller et al. 2026, “Initial results of the Digital Consciousness Model”](https://arxiv.org/abs/2601.17060) (arXiv:2601.17060), which “incorporates a range of leading theories and perspectives” instead of a single theory. An attendee posted the link in the chat; the thanks is to them. [^19]: [51:12](https://youtu.be/FTnnRUus0s4?t=3072) to [51:35](https://youtu.be/FTnnRUus0s4?t=3095). Sebo’s paraphrase of remarks by Anil Seth, not a quotation. Seth’s [target article in *Behavioral and Brain Sciences*](https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/abs/conscious-artificial-intelligence-and-biological-naturalism/C9912A5BE9D806012E3C8B3AF612E39A) argues for biological naturalism and concludes that “real artificial consciousness is unlikely along current trajectories but becomes more plausible as AI becomes more brain-like and/or life-like”. [^20]: [56:03](https://youtu.be/FTnnRUus0s4?t=3363) to [56:16](https://youtu.be/FTnnRUus0s4?t=3376). Anthropic announced its model welfare research program in [“Exploring model welfare”](https://www.anthropic.com/research/exploring-model-welfare) (24 April 2025). The [Claude Opus 4 and Sonnet 4 system card](https://www.anthropic.com/claude-4-system-card) (May 2025) includes a welfare assessment and an external one by Eleos AI Research. Claude Opus 4 and 4.1 were given the ability to [end a rare subset of conversations](https://www.anthropic.com/research/end-subset-conversations) (15 August 2025). [^21]: [56:18](https://youtu.be/FTnnRUus0s4?t=3378). OpenAI’s [Model Spec](https://model-spec.openai.com/2025-02-12.html), which “outlines the intended behavior for the models that power OpenAI’s products”, added in its 12 February 2025 version that the assistant “should not make confident claims about its own subjective experience or consciousness (or lack thereof)”. [^22]: [56:26](https://youtu.be/FTnnRUus0s4?t=3386) to [56:32](https://youtu.be/FTnnRUus0s4?t=3392). Probably Geoff Keeling and Winnie Street, [*Emerging Questions in AI Welfare*](https://www.cambridge.org/core/elements/emerging-questions-in-ai-welfare/96339C532CF4ED8BDDE3F3CEF4CD29F9) (Cambridge University Press, Elements in Philosophy and AI), published online 12 May 2026. Both were listed as main speakers at the [AI Minds Symposium](https://philevents.org/event/show/149313) at the University of London’s Institute of Philosophy on 14 May 2026. ### About this transcript Drafted and edited with AI from Zoom's automatic transcript of the talk, and checked by AI against the talk's slides and papers where they exist, following HΩ's editing rules. It has not been checked word for word against the audio. The talk had no slides, and questions were taken live throughout. Questions and links posted in the chat appear only as they were read out. #### Corrections - 19 October 2026: Footnote 1 no longer cites 48:26 among the answers on public and expert attitudes, and footnote 6 no longer says the book was mentioned in the chat. Spotted an error? Write to [team@horizonomega.org](mailto:team@horizonomega.org).