How consciousness evolved - and why AI can't have it | Sabine Hossenfelder, Nick Lane, Yoshua Bengio

There's a lot of discussion in physics about whether quantum processes play a role in consciousness. Some people say yes, other people say no, but either way, it's something physical, and consciousness is a property that we call emergent. So, it comes about from the interactions of other things, and I think that there is no particular reason why you can't replace the individual particles on some level. So, the individual constituents that interact. And this is where this idea comes from that you can upload your consciousness to computer.

So, the question that I would like each of our speakers to address is: is there a test for thought or consciousness in humans or even computers? Sabine?

Yeah, thank you. So, as you heard in the introduction, I'm a physicist, and I think it will surprise absolutely no one that I think consciousness is physical. It comes about from the interactions of particles in the human brain, and exactly what it is is difficult. So, to answer this question, is there a test for thought or consciousness in humans or computers? Not yet, but we're working on it. There are lots of theories for how you could measure consciousness in systems generally, not just in the human brain, but in any system that's made up of many components that interact with each other. Lots of mathematical theories. I recently wrote a review article about this by Robert Kuhn, which lists 200 theories of consciousness. I sent this to a friend who was very insulted that his theory wasn't on the list. So, I think this gives you an idea. I think they're all wrong. We're very far away from being able to quantify consciousness, but we'll get there eventually. I actually do think that one of the main ingredients to consciousness is the observer observing themselves. And I think that's something which can be in principle mathematically quantified at some point. I don't know how to do it, but in principle it should be possible. With regards to the Turing test, I never really thought of it as being a test of thought or consciousness, but it's more like a kind of a benchmark test for development. It's like when my first child had her first tooth, I was like, 'Look at this. We made a human being and it's got a tooth.' And so this isn't particularly meaningful to anyone except the parents, possibly, but it is a sign of development that is proceeding and this is how I've been thinking about the Turing test. It tells you we're making progress. We're on the path somewhere.

Fantastic. Thank you, Nick.

So you already know what I think. I'm a biologist, therefore I think it's biological, obviously. You're a physicist. I actually think it's physics as well. So I think we always conflate things and we're conflating here consciousness with intelligence. And consciousness is a word which conflates with itself as well. And I think for a long time the hard problem in consciousness has been why is there feeling that it's like to be alive, to do anything, to have any form of sensations when in principle with a philosopher zombie, you could behave the same, everything could be algorithmic, and yet we know that we have feelings. And so this is the hard problem of consciousness. And the reason it's hard is that I don't personally think that neural networks is the right way to ask that question about feelings. That's not to say neural networks can't become extremely intelligent because they're aware of the observer and so on. So intelligence I have no problem. It's the question of feelings that I am more interested in. And I think feelings are probably very grounded in how we think of ourselves as human beings and probably grounded deeply in morality and so on as well. So I think if we can't understand feelings biologically, we're getting nowhere. Now, in terms of that, there's a nice quote from Shakespeare, of course, 'Never the philosopher could endure the toothache patiently.' And we really can't. There's not much intelligence in toothache, but it's a searing pain that we all know what it is. And it's a very much a biological pain. There's no reason why quantum physics should come up with toothache. This is a biological problem that we're dealing with. And we don't know what generates it. Why would this neural network generate a feeling of pain and an apparently identical looking neural network be unconscious? We cannot answer that question in those terms at the moment. I would take it back. I think feelings evolved by natural selection. Therefore, they have utility. Where did they arise? Not in these neural networks. In my mind at least, anesthetics work, for example, on pretty much all animals. Some people even say plants, but certainly protists, which will effectively crawl around. You can stop them crawling around by anesthetizing them with the same anesthetic gases that we use. That says that there's something about consciousness which is a property of living things. And how do we get at it? Well, nobody really knows even now exactly how anesthetics work. This is a 100-year-old mystery. Some people will tell you, 'Oh, we do now.' But the arguments are not fully coherent. It seems that they affect mitochondria. This is my way into this question and why I have an interest. The mitochondria are the power packs in cells, which are providing you with energy. It's a little bit, you know, if you pull the plug out, obviously the system goes down. It's more interesting than that. Mitochondria were bacteria once, and bacteria, think of it as a little packet with a membrane around it interacting with the world. Inside that little packet, you've got a billion reactions every second. Sure, you've got genes, you've got enzymes, metabolic machinery, but what you're doing is a billion reactions a second, second after second after second. How the hell do you know what you're doing? How do you integrate that and say, 'Okay, I'm in trouble here. I should move over there.' That, I think, is the level where a feeling is beginning to come in. Is a biological system taking some kind of top-level appraisal of exactly how it's doing in relation to the environment and what it should choose to do next. And I think that's where feelings come in. It's linked with agency, with the ability to move, with the ability to make a decision, and it's got nothing at all to do with neural nets.

Well, okay. None of those were two minutes apart from Sabine's. But that's fine. Let's dive into that first theme. So, we've actually all touched upon this. So, the first theme is kind of about this question of whether or not we can explain thought and experience in terms of physical material. Sabine, can I turn to you and ask you to sort of kick us off on that because you sort of acknowledge the importance of physical material and the sort of substrate. And so did Mick afterwards. But what do you think? Can we actually explain thought and experience in those terms?

The way that the question is phrased is a little bit weird because I think it's not so much the material that matters, but the connections of the system, which could be made of many different types of materials. And at the moment, we don't really know just exactly what is necessary about the substrate, as you put it. Like, what are the essential physical ingredients to consciousness? You just said it's something to do with my mitochondria. There's a lot of discussion in physics about whether quantum processes play a role in consciousness. Some people say yes, other people say no. Basically, the brief summary is we don't really know. But either way, it's something physical and consciousness is a property that we call emergent. So, it comes about from the interactions of other things. And I think that there is no particular reason why you can't replace the individual particles on some level. So, the individual constituents that interact. And this is where this idea comes from that you can upload your consciousness to computer because the computer has another substrate, so to speak. It's not biological, it's made of silicon or what have you. But it could reproduce exactly the same connections and pathways that are going on in your brain. And it seems to me that this is likely to be correct, but you know, we don't really know.

That's really interesting to me. So, you've basically said it could be biological substrates, it could be silicon chips, it doesn't make a difference because computation is effectively the basis of consciousness. Is that correct? And therefore, if it's computing, then that goes on. But you've also alluded to the sort of subjective experience thing that comes along with it, too, right? So, are you saying that the basis of consciousness is the capacity to compute in any form? And then secondarily, no. Okay.

I'd like to say something on the nature of substrates and computation. I completely agree with virtually all of that. In terms of computation, if you think of a neural net as a series of switches that are on or off and networking in one way or another, yes, all of that is computation. But, neurons are not just a switch. They are a lot more complex than that. And I alluded already to this question about why some neural nets give rise to a conscious experience and others don't. Now, it's interesting to me that the people from a scientific background who have explored consciousness, biologists are amazingly under represented. Biology is a really conservative discipline, I have to say, and biologists don't like making fools of themselves. I quite like making a fool of myself, so I'm here. But the people who do are mostly neuroscientists, and they're thinking about the human brain and about the network properties of the brain and emergence at the level of the brain and so on. What I was talking about before in a bacterial cell, a billion reactions a second, all of which can go wrong, all of which are happening in water, you've somehow got to come up with in a nanosecond, some kind of idea of how am I doing right now? Do I need to leave? It's a biological decision that you need to make with wetware. And I think that's very different to just silicon chips and dryware. So, we're probably less good at the computation. I think undoubtedly we already are less good, but we're doing something else, which is where feelings are coming in.

So, I think you can adopt a materialist view of the world, which is that everything is made of material bits, and try and account for everything in that framework. It's a very powerful model. It's one that I would want to use myself in lots of circumstances, but it's a model. It's not a description of what's going on out there. You can apply a biological model.

To continue watching this video, click the link in the top left or in the description below. With a free trial, you can enjoy the full talk and thousands more. Thank you for being part of the conversation.