AI Consciousness Explained: What Counts as Evidence, Anthropic, Lessons from Animals (Jeff Sebo,NYU)

Yes, absolutely. So, I am a philosopher by training, though I now do interdisciplinary work. I work at the Department of Environmental Studies at New York University. I ask questions about the nature and the value of non-human minds, and then how we interact with non-humans of various types. We ask questions like which non-humans, like animals or AI systems, are conscious, capable of having feelings, sentient, capable of feeling pleasure and pain, agentic, capable of setting and pursuing their own goals. And as animals or AI systems or other non-humans do have those features, if they can have their own thoughts and feelings and projects and relationships, then what kind of moral status should they have? What kinds of moral responsibilities might we have to consider their interests when making decisions that affect them? And then what follows from that for our policies and how we interact with them, for our food systems, for our infrastructure, for technological development. So, very big questions and very integrative questions, which is why it helps to think about them in an interdisciplinary way.

Yes, I think they apply to some extent. And I spent the past 20 years thinking about these questions with non-human animals. And then as AI started to become more sophisticated, as development and deployment really started to take off with the success of large language models and other generative models, because of my background in animal studies and animal minds and animal ethics and policy, I was anticipating that some of those same questions would start to arise for AI systems. In the same way, we might reach a point where we genuinely feel unsure if they have their own feelings and emotions, if it feels like anything to be them. And yet, we will be creating and controlling them. And so, we will need to make decisions that affect them whether we have certainty about those issues or not.

That is very much the hope. If you look at our history with non-human animals, you can see that we have done a lot of damage to them and more broadly to the natural world. For a lot of the 20th century, for example, we assumed that only humans deserve moral consideration, and even that only humans or only very human-like animals like other primates are conscious, able to have their own feelings and emotions. And it was only near the end of the 20th century that serious scientists and philosophers in general came around to the idea that no, maybe consciousness and moral significance are much more widespread, that they extend not only to humans and other primates, but to other mammals and birds. And now in the 21st century, we realized, oh wow, it might even be all vertebrates and many invertebrates like octopuses and lobsters and even ants and bees. But by the time we came around to that awareness, we had already spent the past 50 years scaling up a global industry of factory farming and industrial animal agriculture, invasive animal research, these industries that cause profound pain and suffering to clearly sentient and agentic non-human animals with lives of their own. And had we thought in advance, and had we taken a more precautionary approach, and had we had more of an open mind, we could have just taken a completely different trajectory that would have been equally good if not better for us and way better for them. And so that is the cautionary tale that I think we should be having in mind when we contemplate bringing into existence this new alien type of mind that is clearly very sophisticated and very different from us and in some ways very mysterious to us.

There is some social psychology here. So we know, for example, that we are more likely to attribute consciousness and sentience to an entity when they look like us and act like us and we can relate to them. When an animal has a large body and a large head and large eyes, when they have symmetrical features, when they have furry skin instead of scaly skin, when they have four limbs instead of six or eight limbs, those are all features that are more likely to trigger our empathy and attributions of consciousness and moral status. And then there are also social and economic components. So when we socially classify an animal or a non-human as a companion instead of a commodity, then we are also more likely to empathize and attribute consciousness and moral standing to them. But then there were also these prevailing views in science at the time that other minds are not really an appropriate topic for science because you cannot directly observe them. You can only make inferences about them. And so it was seen as a kind of non-scientific topic. And so when you combine an assumption that they are not conscious with a view that this is not studyable by science, that becomes a recipe for it just being stuck in the muck and not being able to use the tools of science to get out of it.

Absolutely, yes. Sentience, as philosophers tend to use the term, is the ability to consciously experience positive and negative states like pleasure, pain, happiness and suffering. And so consciousness is then understood to be part of sentience. Consciousness means the ability to have feelings or emotions of any kind. And that can include sensory feelings like the feeling of seeing the color red or the feeling of smelling a rose, as well as affective feelings like the feeling of bodily pleasure or bodily pain or emotional satisfaction and emotional frustration. So sentience is what happens when you have consciousness. You can have feelings, and then some of those feelings can have a positive or a negative valence. This is when we really become worried that you are vulnerable to pain and suffering and these other negative states, and now all of a sudden we have the ability to harm you and make your life worse. And this is when we take on moral responsibilities that are owed to you. The moral circle or the moral community is all beings who should be considered when we make decisions that affect them.

Yeah, great questions. A mainstream view, and this is a view that I endorse, is that on the one hand there is not very strong evidence of consciousness in current generation large language models or other AI systems. So when we go and study GPT-5 or Claude 4 and we study their behaviors, but then we also study their internal architectures and we study their developmental histories and what kind of training they underwent, it seems quite plausible that the best explanation of their admittedly very impressive behavior is still that they are engaging in a kind of automated process of pattern matching and text prediction, and so on. On the other hand, we are also not finding any clear technical obstacles towards the creation of AI systems that have many of the behavioral and internal architectural and developmental genealogical markers of sentience that we would look for in other animals. For example, what that means is that the probability that current AI systems are conscious is perhaps still low, not zero, but also maybe not 50%.

But we can expect over the next 10 years that the probability is going to go up and up and up and up. And so part of why we feel a sense of urgency around the issue is not that I think there is a 99% chance that GPT-5 deserves moral consideration, but there could be a 40% chance that GPT-7 deserves moral consideration. And if that might be true, then we really ought to think ahead so we can be ready when the time comes. Now, to answer your questions about industry and companies, yeah, the main leader in this area has been Anthropic. Me and my colleagues wrote a report called Taking AI Welfare Seriously that we released in fall 2024 where we basically argued that hey, companies ought to start taking this issue seriously. To their credit, Anthropic, when we released that report, they hired one of the authors, Kyle Fish, as their first full-time AI welfare researcher. They then in spring 2025 launched what they call a model welfare program. And they started releasing model welfare evaluations with their Claude 4 system card. They even released an intervention in summer 2025 where they now permit Claude to exit harmful and abusive conversations partly on AI safety grounds but partly on AI welfare grounds. So there is some action there and some Google researchers are getting interested now. One or two people at OpenAI are getting interested as well. So momentum is slowly starting to build in industry, led so far primarily by Anthropic.

Mhm.

Yeah, I think that is right. And those results from Anthropic were very interesting and very intriguing and kind of charming too. Like the fact that Claude is so drawn to what Anthropic describes as a spiritual bliss attractor state, that when left to their own devices, Claude models will just be filled with curiosity about consciousness and gratitude for the beauty of the universe and so on, that this is where they gravitate is kind of charming and lovely and I hope that that is a feature that at least this model can hold on to. Now I should say, it is not clear that Claude is in fact conscious, and I know you were not suggesting that Claude is in fact conscious. Still, it is difficult to tell what the best explanation of these interesting behaviors are. You could explain them as actual markers of consciousness or you could explain them as products of the training data, patterns contained within the training data, because Claude is trained on a lot of conversations that involve consciousness as well as reinforcement learning from human feedback as well as Anthropic's own interventions in how Claude responds to these types of questions. So there is always the possibility that Claude is conscious and expressing preferences as well as the possibility that Claude is not conscious and is just playing a character they were trained to play.

There will be AI welfare advocates who think that we should have AI welfare protections that fall short of full legal personhood and rights as we do for non-human animals in the United States, for example. Then there will be other people who advocate for giving AI systems full legal personhood and rights including rights to own property and participate in our legal and political institutions because unlike animals, they will not only potentially be conscious and sentient, but also rational and capable of participating in our legal and political institutions if we permit them to. So there will be some people who are advocating for welfare protections like we have for animals. Other people advocating for legal personhood and rights and responsibilities as we have for ordinary adult humans. And then other people who advocate for continuing to treat them purely like objects as unfortunately we do to some extent with non-human animals as well. And then there will just be a big societal debate and probably some cities and states and countries who are a little bit more progressive and inclusive and who do extend these forms of protection to them and then others that are a little bit less so and deny this of them. And we are going to have to reckon with this as a society and figure out together where we go. I have my own views about this but nobody should be imposing their views on the rest of society. We are going to have to work this out democratically and some judges might have to step in at a certain point too.

So there are going to be all kinds of perspectives that we have to reckon with.

That is a great question. Yes. Yeah. We will certainly have that come up as an issue. I would think within the decade if not sooner than that. Within our lifetimes. Will they legally get married? I am going to go out on a limb and say yes.

Yeah, that is like 51-49, but I am guessing it will emerge as a major issue. I am not sure that it should happen but I am expecting that within the next decade or two or three there will be AI systems that do have physical embodiment, that do have advanced and integrated capacities for perception and attention and learning and memory and self-awareness and social awareness and rational flexible decision-making, and they have at least their versions of feelings and emotions. At least they can behave as though they have feelings and emotions. And so we will have chatbots as well as maybe even robots who live in our houses who are able to have ongoing conversations with us, to have ongoing relationships with us. I think we will have to decide as a society what kinds of safeguards do we want on this both for the sake of the humans and for the sake of the AI systems. But there I think will be adults who have relationships with chatbots and robots that are a little bit closer in kind functionally, behaviorally to how we relate to each other.

Yeah, that is a really good question. I have a report coming out with Elicit about empirical research methods that could really make progress on understanding the probabilities in which AI systems are conscious. There might not be any single thing an AI system can do where that behavior alone is decisive evidence of consciousness. They can pass the Turing test. We can have extended conversations with them and then that context window will get bigger and bigger. So soon enough they can perform month-long tasks and we can have month-long relationships and interactions with them. So no single behavior I think can be decisive proof. What can be strong proof is when we triangulate behavioral evidence with internal architectural evidence with developmental genealogical evidence. Like look at other animals. They behave like they dislike pain. But that by itself is not proof that they can feel pain. The reason we feel confident that non-human animals feel pain is they behave like they dislike this stuff and they have this internal anatomical brain and nervous system structure that we know corresponds to pain experience in humans and they evolved on the same evolutionary tree as us and so they have this stuff for the same reason. It's all of that together that allows us to be confident that when my dog lifts his paw up and starts to limp and whimper, it's all of that together that gives me grounds for being confident that he really is experiencing pain.

So with AI systems, we'll need to triangulate in that same way. We'll look at the behavior, yes, but then we'll also look and ask internally, architecturally, do they have an analog of our brains and central nervous systems and is that participating in the production of this behavior? And then developmentally, genealogically, were they trained and did they come into existence in a way that would make this beneficial for their survival in the pursuit of their goals in the same kind of way as with us and other animals? And then when we put all of that together and then they're picking their foot up off the ground and saying, 'Ouch, this hurts,' that's when it becomes strong evidence.

Well, thank you.