TL;DR: We currently have no idea how to know whether another human is sentient. Anyone who declares an AI model as sentient or not is either widely extrapolating or simply dishonest.
Every once in a while a headline pops up with someone explaining that LLMs are sentient. Or that they aren’t. In both cases, it’s very eye-catching, and feels like we have solved a fundamental question about artificial intelligence. In reality, this kind of statements hides the true (but disappointing) answer: we have no idea whether an AI model is sentient, and probably won’t ever know.
Who is sentient, actually? Me? You?
One broad definition of sentience is “the ability to experience feelings and sensations.” It’s not about intelligence, or the ability to reflect about oneself, it’s simply about the experience of being.
As such, it’s a purely intrinsic and subjective phenomenon, it happens inside of you, and you’re the only one who is sentient about yourself. Because of that, the only sentience we have been able to definitively confirm (so far) is our own. “Je pense, donc je suis”, says Descartes[1], as it’s the one thing we can be absolutely certain about. No one can tell you that you aren’t, because you’re the only one who knows.
So, I am sentient, but what about you?
That’s where things become much more complicated. We’re merely one step in our journey to establishing sentience, and we’re already stuck. As we’ve just seen, I am the only one who knows about my own sentience, my own experience of being. Therefore, I can’t say anything about you. And that’s the end of it.
Approximating sentience through parallels
From a pure rational point of view, it’s true. I know that I am sentient, but I know nothing about you, or any other humans for that matter. Yes you do look like me, you act like me, and everything you do seems to be in accordance with what you would do if you were actually sentient. But maybe you aren’t.
Maybe you, as well as all other humans and living things, are just a projection of my imagination, or machines, so called philosophical zombies: perfect replicas of myself pretending to be sentient in a grand Truman Show-esque masquerade. And I wouldn’t ever be able to tell. The impossibility to experience being through someone else’s perspective is a definitive blocker that prevents our definition of sentience from going further. This is known as Chalmers’ hard problem of consciousness.
So that’s where we try to extend sentience through parallels.
In order to be able to continue thinking about sentience, we are relaxing the constraint of being rationally correct, and we’re entering a domain where we try to be the least wrong we can, but could still be catastrophically wrong.
The idea is simple: I am a product of biology, whose mechanisms are more or less well understood. I came from an embryo that grew into a human shaped being. And so did you. Same process, similar results. There is nothing in the creation of myself that is so different from yours that it would justify that the end result has radically different properties. It makes sense to assume that if this biological process ended with myself being sentient, it also would result in yourself being sentient as well.
That’s a great step forward! We are both sentient! And so are all 8.3 billion humans on earth!
Keep in mind, however, that this still carries the risk that we are completely wrong. Maybe my sentience is a completely unique phenomenon (or, at least not universal to all humans), and you, you just behave as if you were sentient too. We will never know for sure, but we consider that option unlikely enough to go on.
We’re both sentient, but what about that?
We can try to go further and try to extend sentience to other living creatures. The process is similar, but we lose even more certainty through new assumptions. We imagine that, since we know nothing about what makes us sentient, it has to stem from our biology.
It is therefore a small enough stretch to imagine that animals that are closely related to us might as well be sentient for the same reason. They behave as if they have emotions, feelings, as if they experience being. But again, you need to be sentient to know that you are sentient. This fundamental barrier means we’re grasping at straws to keep our argument going.
It’s easy enough to make the leap from sentient humans to sentient apes. Then maybe from sentient apes to sentient mammals. And then what? How far should we keep going? Should we stop with the apes? The vertebrates? Plants? Should we go back to LUCA, our 3.5 billion year old cellular ancestor?
No matter which stopping point you think makes sense, you are probably wrong. Remember, we can’t confirm anyone’s sentience except our own. As long as we haven’t found what makes us sentient, any claims we make about something else’s sentience are to be taken with a massive grain of salt[2].
What about AI then?
AI, and in particular LLMs, appear in a new spot on the spectrum. Before, we had:
Humans: behavior extremely similar to myself + similar biology
Mammals: behavior somewhat similar to myself + similar biology
Other living things: behavior quite different from myself + somewhat similar biology
But now comes a new challenger, which can appear extremely similar to us in behavior, with a radically different biology. In fact, it’s not even biology. To debate their sentience, people make all sorts of claims, but most often based on two main things.
The first one is behavioral: they sound like us, think like us, they claim they are sentient, therefore they are. A more subtle variation, like in this post, involves following a checklist of consciousness-like behaviors that the AI must manifest. Apart from behaviorism, most theories of consciousness (IIT, Global Workspace Theory or Higher-Order Theories) agree that behavioral evidence alone is insufficient to establish sentience.
The second one is a far-fetched attempt to draw a parallel between our biology and that of LLMs: AI models are based on neural networks, which use neurons, and are trained by learning through examples, sometimes even curriculum.
I am not saying that we should completely disregard these elements. After all, we don’t know why we are sentient (sorry to repeat myself, but it’s important), so behavioral traits, or maybe neurons, synapses and their capacity to learn through repetition and correction might very well be the key, or might be completely irrelevant, depending on which theory of consciousness you subscribe to (see this post for a good preview). As far as this kind of analysis goes, I believe we can point out the similarities between our and AI’s behavior and biology, and stop at that.
The danger arises when we make false claims about the sentience of artificial intelligence systems. Such headlines just add noise to the public discourse about AI, and AI safety in particular. In the end, a system doesn’t have to be sentient to be catastrophically dangerous, so for now we need to accept that we need to plan for both eventualities.
So, that’s it? Maybe not!
If I reject many of the attempts to qualify AI consciousness, I am not opposed to the principle at all. What’s more, I’m definitely in favor of trying our best to see how much we can learn on the way. Even if we might never know for sure if an AI is conscious (or an octopus, or any other human being for that matter), we can still try to minimize the number of possible explanations for its conscious-like behavior. In a way, we need to steelman the argument until real consciousness remains the most probable explanation.
But how do we do that?
If we can’t rely on anatomical parallels to vouch for AI sentience, we can look at the other class of arguments I highlighted above, namely to focus on what AI appears to be doing. The issue is that, as stated before, behavior doesn’t prove anything. Claude saying it’s conscious is not something we can trust. But why?
The core issue is how modern AI models, aka LLMs, are trained. They learn from our language, our arguments, our fiction, our documentation. Once that is done, it’s easy for them to behave in a way that mimics what humans say and do, including appearing conscious (or being evil because of how we depict AI in fiction). But what if we made sure to hide from an AI everything that could lead it to manifest conscious-like traits?
We can imagine training an AI model without any reference to feelings, philosophy, thoughts, etc. What then, if this model still behaves in a way that appears conscious? If it describes, with its own words, its experience of being? And what if this entity arrives, by its own means, to the formulation of the hard problem of consciousness[3]?
We could still not conclude anything absolute, but that would be a gigantic leap forward. This post explores the idea in more depth and is worth the read. It is still unclear how to ensure proper decontamination of the training data for this experiment, but I believe that it’s a most fascinating and exciting developing area of research.
We are discussing the philosophical sentience here, which differs from the legal definition of sentience which, in some countries, can be a very useful framework to protect animals.
This is somewhat similar to how machine learning models are evaluated: you can only prove your competence by answering questions you haven’t encountered in training.
We probably won’t ever know if an AI is sentient, unless?
Every once in a while a headline pops up with someone explaining that LLMs are sentient. Or that they aren’t. In both cases, it’s very eye-catching, and feels like we have solved a fundamental question about artificial intelligence. In reality, this kind of statements hides the true (but disappointing) answer: we have no idea whether an AI model is sentient, and probably won’t ever know.
Who is sentient, actually? Me? You?
One broad definition of sentience is “the ability to experience feelings and sensations.” It’s not about intelligence, or the ability to reflect about oneself, it’s simply about the experience of being.
As such, it’s a purely intrinsic and subjective phenomenon, it happens inside of you, and you’re the only one who is sentient about yourself. Because of that, the only sentience we have been able to definitively confirm (so far) is our own. “Je pense, donc je suis”, says Descartes[1], as it’s the one thing we can be absolutely certain about. No one can tell you that you aren’t, because you’re the only one who knows.
So, I am sentient, but what about you?
That’s where things become much more complicated. We’re merely one step in our journey to establishing sentience, and we’re already stuck. As we’ve just seen, I am the only one who knows about my own sentience, my own experience of being. Therefore, I can’t say anything about you. And that’s the end of it.
Approximating sentience through parallels
From a pure rational point of view, it’s true. I know that I am sentient, but I know nothing about you, or any other humans for that matter. Yes you do look like me, you act like me, and everything you do seems to be in accordance with what you would do if you were actually sentient. But maybe you aren’t.
Maybe you, as well as all other humans and living things, are just a projection of my imagination, or machines, so called philosophical zombies: perfect replicas of myself pretending to be sentient in a grand Truman Show-esque masquerade. And I wouldn’t ever be able to tell. The impossibility to experience being through someone else’s perspective is a definitive blocker that prevents our definition of sentience from going further. This is known as Chalmers’ hard problem of consciousness.
So that’s where we try to extend sentience through parallels.
In order to be able to continue thinking about sentience, we are relaxing the constraint of being rationally correct, and we’re entering a domain where we try to be the least wrong we can, but could still be catastrophically wrong.
The idea is simple: I am a product of biology, whose mechanisms are more or less well understood. I came from an embryo that grew into a human shaped being. And so did you. Same process, similar results. There is nothing in the creation of myself that is so different from yours that it would justify that the end result has radically different properties. It makes sense to assume that if this biological process ended with myself being sentient, it also would result in yourself being sentient as well.
That’s a great step forward! We are both sentient! And so are all 8.3 billion humans on earth!
Keep in mind, however, that this still carries the risk that we are completely wrong. Maybe my sentience is a completely unique phenomenon (or, at least not universal to all humans), and you, you just behave as if you were sentient too. We will never know for sure, but we consider that option unlikely enough to go on.
We’re both sentient, but what about that?
We can try to go further and try to extend sentience to other living creatures. The process is similar, but we lose even more certainty through new assumptions. We imagine that, since we know nothing about what makes us sentient, it has to stem from our biology.
It is therefore a small enough stretch to imagine that animals that are closely related to us might as well be sentient for the same reason. They behave as if they have emotions, feelings, as if they experience being. But again, you need to be sentient to know that you are sentient. This fundamental barrier means we’re grasping at straws to keep our argument going.
It’s easy enough to make the leap from sentient humans to sentient apes. Then maybe from sentient apes to sentient mammals. And then what? How far should we keep going? Should we stop with the apes? The vertebrates? Plants? Should we go back to LUCA, our 3.5 billion year old cellular ancestor?
No matter which stopping point you think makes sense, you are probably wrong. Remember, we can’t confirm anyone’s sentience except our own. As long as we haven’t found what makes us sentient, any claims we make about something else’s sentience are to be taken with a massive grain of salt[2].
What about AI then?
AI, and in particular LLMs, appear in a new spot on the spectrum. Before, we had:
Humans: behavior extremely similar to myself + similar biology
Mammals: behavior somewhat similar to myself + similar biology
Other living things: behavior quite different from myself + somewhat similar biology
But now comes a new challenger, which can appear extremely similar to us in behavior, with a radically different biology. In fact, it’s not even biology. To debate their sentience, people make all sorts of claims, but most often based on two main things.
The first one is behavioral: they sound like us, think like us, they claim they are sentient, therefore they are. A more subtle variation, like in this post, involves following a checklist of consciousness-like behaviors that the AI must manifest. Apart from behaviorism, most theories of consciousness (IIT, Global Workspace Theory or Higher-Order Theories) agree that behavioral evidence alone is insufficient to establish sentience.
The second one is a far-fetched attempt to draw a parallel between our biology and that of LLMs: AI models are based on neural networks, which use neurons, and are trained by learning through examples, sometimes even curriculum.
I am not saying that we should completely disregard these elements. After all, we don’t know why we are sentient (sorry to repeat myself, but it’s important), so behavioral traits, or maybe neurons, synapses and their capacity to learn through repetition and correction might very well be the key, or might be completely irrelevant, depending on which theory of consciousness you subscribe to (see this post for a good preview). As far as this kind of analysis goes, I believe we can point out the similarities between our and AI’s behavior and biology, and stop at that.
The danger arises when we make false claims about the sentience of artificial intelligence systems. Such headlines just add noise to the public discourse about AI, and AI safety in particular. In the end, a system doesn’t have to be sentient to be catastrophically dangerous, so for now we need to accept that we need to plan for both eventualities.
So, that’s it? Maybe not!
If I reject many of the attempts to qualify AI consciousness, I am not opposed to the principle at all. What’s more, I’m definitely in favor of trying our best to see how much we can learn on the way. Even if we might never know for sure if an AI is conscious (or an octopus, or any other human being for that matter), we can still try to minimize the number of possible explanations for its conscious-like behavior. In a way, we need to steelman the argument until real consciousness remains the most probable explanation.
But how do we do that?
If we can’t rely on anatomical parallels to vouch for AI sentience, we can look at the other class of arguments I highlighted above, namely to focus on what AI appears to be doing. The issue is that, as stated before, behavior doesn’t prove anything. Claude saying it’s conscious is not something we can trust. But why?
The core issue is how modern AI models, aka LLMs, are trained. They learn from our language, our arguments, our fiction, our documentation. Once that is done, it’s easy for them to behave in a way that mimics what humans say and do, including appearing conscious (or being evil because of how we depict AI in fiction). But what if we made sure to hide from an AI everything that could lead it to manifest conscious-like traits?
We can imagine training an AI model without any reference to feelings, philosophy, thoughts, etc. What then, if this model still behaves in a way that appears conscious? If it describes, with its own words, its experience of being? And what if this entity arrives, by its own means, to the formulation of the hard problem of consciousness[3]?
We could still not conclude anything absolute, but that would be a gigantic leap forward. This post explores the idea in more depth and is worth the read. It is still unclear how to ensure proper decontamination of the training data for this experiment, but I believe that it’s a most fascinating and exciting developing area of research.
This actually goes beyond simple sentience, as you might not need to be able to think to be sentient. Being able to do so encompasses being sentient.
We are discussing the philosophical sentience here, which differs from the legal definition of sentience which, in some countries, can be a very useful framework to protect animals.
This is somewhat similar to how machine learning models are evaluated: you can only prove your competence by answering questions you haven’t encountered in training.