Murray Shanahan on Consciousness, Reasoning, and the Philosophy of AI

Source material

Transcript

[Music] I think there are just a huge number of enormously interesting philosophical questions that AI gives rise to.

You know, what is the nature of the human mind?

What is the nature of mind?

Welcome back to Google DeepMind, the podcast.

My guest on this episode is Murray Shanahan, Professor of Cognitive Robotics at Imperial College London and Principal Research Scientist at Google DeepMind.

Now, we have all heard the stories about people falling in love with their chatbots, about people pushing large language models to contemplate their own existence, or questioning the limits of their conceptual understanding of reality.

But these kinds of questions about self-identity and thinking and meta-cognition have been puzzling philosophers for millennia already.

And so it makes sense that they should be turning to AI to interrogate the most profound questions about the nature of AI's intelligence, of its current capabilities, even its consciousness or otherwise.

Murray Shanahan has been working in the field of AI since the 1990s, and if you've been following this podcast for a while, you will remember him as the man that consulted on the 2014 science fiction film Ex Machina about a computer programmer who gets the chance to test the intelligence of a female robot, Ava, and ultimately questions whether she is conscious.

Welcome back to the podcast, Murray.

Just thinking back, because I know that you played a key role in Ex Machina, shall we say, the Alex Garland film.

What do you think you got right in that film and in other science fiction films that were around at the time?

I mean, thinking back to sort of 10, 15 years ago, were we on the right track?

So I think that one respect in which Ex Machina really did a great service was that it does raise a whole load of very interesting and provocative questions about consciousness and about AI and consciousness, and therefore about consciousness itself.

So that's one, you know, that's one huge success.

But it's interesting that just very shortly before Ex Machina came out, her came out.

So Spike Jones's movie, Her, came out.

And at the time, I really wasn't all that keen on her as a movie, because I just thought it was so implausible that a person could fall in love with this kind of disembodied voice, you know, even if it's Scarlett Johansson's.

I mean, how wrong was that?

As a bit of prediction, I think Her really did amazing well at predicting the world we've got now.

Now, we don't know quite how things are going to unfold in the next few years, because maybe robotics will progress rapidly as well in the way that language has in AI.

But at the moment, you know, it's all about disembodied language.

And also, you know, Her showed how people can, in fact, be very much, you know, form relationships, whatever, you know, in the broadest sense with disembodied AI systems, which is an extraordinary thing.

We're talking 10, 15 years ago, but your involvement in AI goes back much further than this.

I mean, I know that you had, you knew John McCarthy.

I did know John McCarthy, I knew him very well.

John McCarthy was a professor of computer science and artificial intelligence.

Back in the day, he actually coined the phrase artificial intelligence, and was one of the authors of the proposal for the very famous Dartmouth conference that took place in 1956, which was the first AI conference in the world.

And that conference really mapped out the whole field.

And people just weren't thinking about this kind of thing seriously at all.

It was just a handful.

So, you know, I think he was a real radical thinker.

Okay, that choice of words, artificial intelligence, was it a good choice of words?

Yeah, I mean, I still think it was.

I mean, I know that some people think that perhaps it wasn't a good choice of words, but I...

Give us some of their arguments.

So first of all, there is the word intelligence.

So intelligence, you know, itself is a, in some ways, a very contentious concept, especially if people think about IQ tests and that kind of thing.

And the idea that intelligence is something that can be quantified on a straightforward, simple scale, you know, and then some people are more intelligent than others.

And I think in psychology, it's well recognized today that there are many different kinds of intelligence, and this is a really important point, right?

So there's that concern about that word there.

So what would you have used differently?

Well, maybe artificial cognition or something.

I often use the word cognition to mean kind of, you know, thinking and processing information and so on.

But hey, it doesn't have the same ring to it, does it?

Let's be honest.

Especially not now.

I think we're too far down this road, aren't we?

Yeah.

The word artificial, I don't really have a problem with the word artificial.

That seems like a right kind of thing.

It's alluding to the fact that it's something that we've built and that hasn't evolved in nature.

And so that seems the right sort of word.

The objection to that word, I guess, is this, ultimately, everything that artificial intelligence is built on is at some level constructed by humans.

Sure, yes, but it is.

So what's wrong with the word, you know, in that case, I mean, I think that's true.

You were working on symbolic AI, right?

Just just talk to us about the difference between that and the other types and where we're at now.

Absolutely.

Yeah.

Yeah.

So the so-called symbolic paradigm of artificial intelligence was very much dominant for many decades.

So the idea there is that it's all about the manipulation of symbols and of language like sentences and using kind of reasoning processes with those symbols.

The classic example would be an expert system.

So where back in the 1980s, people were building these expert systems.

And the idea was there was that you would try to encode medical knowledge, say, in a set of rules.

And the rules would be something like, you know, oh, you know, if the patient has a temperature of 104 and their skin is purple, and, you know, then there's a 0.75% probability that they've got, you know, is skinnyitis or something.

You can tell that I'm not a medical doctor.

Just about.

Yeah.

And then so you'd have thousands and thousands of these sorts of rules would be put into a kind of big knowledge base.

And then you'd have what was called an inference engine, which would carry out logical reasoning over all of these rules and come to some conclusion about what the likely disease was in that.

But it was a lot of if this then that.

It was a lot of, yeah, if then type rules largely.

And one of the big problems with that is that where do the rules come from?

Well, somebody has to write them all out, basically.

And so there was a whole field of knowledge elicitation where you go around to experts and you try and extract from them their understanding, you know, in their domain, which could be medical diagnosis, it could be fixing photocopiers, it could be the law.

And you try and codify all of this into a computer comprehensible, very precise rules.

That was a very cumbersome process.

And also what you ended up with at the end was very, very brittle, it would go wrong in all kinds of ways.

Another big area of research was common sense, because often it was realized that we implicitly have an enormous amount of common sense knowledge about the everyday world to do with just everyday objects, the fact that they're solid, the fact that they move in certain ways, they fit into each other in certain ways, you know, liquids and gases and gravity and, you know, all kinds of things like that.

And we actually bring all of that knowledge to bear all the time and what we're doing, but it's sort of unconscious.

So then there was a big project, there were various big projects to try and codify all of that common sense knowledge.

But then trying to turn that into like axioms and logic and rules and everything was a nightmare.

So I eventually I think by about the early 2000s, I really thought that this research paradigm was kind of doomed, to be honest, I sort of started moving away from it.

But then of course, along came things like neural networks and so on, which was much less about, you know, if then rules and much more about sort of extracting information from a large amount of data.

But then I sort of wonder now about now that language is effectively cracked, have we sort of reached a higher level of abstraction where we can go back to some more of those symbolic techniques, some of those more symbolic ideas?

Yeah, well, we certainly have because nowadays, one of the hot topics at the moment with large language models is reasoning.

So you have the so called chain of thought models that actually carry out a whole, you know, rather than simply generating an answer to a question, they generate a whole chain of reasoning before they issue the answer.

And that can be very, very effective.

So it's interesting how that harks back in many ways to the kind of thing that people were looking at back in the days of symbolic AI.

But the underlying substrate for doing all that is very, very different indeed, because it's not hard coded rules.

It's, as you mentioned, neural networks that have learned.

Let me pick up on that point about reasoning.

As a philosopher, someone with, you know, background in logic, how good do you think that AI is at reasoning?

Well, that's a very interesting and kind of open question and somewhat controversial.

So computer scientists and AI people, they have a particular notion of reasoning, a particular concept of reasoning, which very much, you know, harks back to formal logic and theorem proving.

So in the days of symbolic AI, for example, then you had systems that were really very good at doing theorem proving with formal logic.

And so people think, well, that's proper reasoning.

That's really, that's your hardcore kind of reasoning.

And today's large language models, they can't match the performance of a, you know, a hand coded theorem prover or logic engine of the sort that's been around for decades.

Give me an example of a type of theorem that might be able to be proved by a hard coded system.

It will be where you've got maybe, you know, 20 or 30 axioms of logic.

So it might be something like the number that follows one is two.

Well, I mean, it could be something like that.

It could be in the domain of number theory or something very mathematical, but it could be something much more every day.

For example, suppose that you've got some very difficult logistical planning problem where maybe you have hundreds of lorries and depots and goods and all kinds of things like that.

And you need to plan the roots and the deployment of the lorries and where they're going to go.

So that's a very kind of difficult problem computationally, and it can be expressed very precisely in formal rules.

And that's the kind of situation where you might want to use a good old fashioned, straightforward planning algorithm of the sort that's been around for a long time.

Now, contemporary large language models are getting better and better at this kind of thing, but they're still, you know, you don't have those kinds of mathematical guarantees that they're always going to come up with exactly the right answer.

And it's very easy to kind of make examples where you have more and more axioms and so on, where they're going to slip up.

There's a whole separate research direction, which is to try and build more hand coded things that combine today's AI techniques with more old fashioned symbolic techniques to specifically for mathematical theorem proving and DeepMind has done some amazing work along those lines.

But that's different from large language models.

So with large language models, we're thinking of these chatbots that can talk about anything under the sun.

And one of the things they happen to be able to do is a kind of reasoning.

So that's not going to be at the moment quite as good as you could do by hand building something for that.

It's kind of interesting because hand building something, I mean, you end up with something that's very rigid.

That's the problem.

Yeah.

And brittle.

Yes, absolutely.

But then at the same time, the sort of flexibility that you get from the gender of AI approach, you know, it's too floppy, as it were.

You know, you want the rigidity in there.

Well, you know, maybe or maybe not.

I mean, I think many examples of human affairs are just not as black and white as that.

And you know, you do maybe want things to be a bit more blurry, even in sort of simple everyday things like what will be good flowers to put over in this corner of the garden?

Well, you know, we've already got some roses in that corner there.

And those roses are yellow, but we can't have too much yellow.

So maybe we need to move them to the other corner of the garden.

But then at the same time, though, is this real reasoning?

Or is this just the AI kind of mimicking well structured arguments that have existed in the training data, but just in a novel environment?

Yeah, well, of course, that begs the question, what is real reasoning?

You know, it's not written in the sky, you know, what real reasoning is, it's up to us to define the concept of real reasoning or of reasoning.

And so, you know, we were talking earlier on about kind of mathematical reasoning of the sort that logicians do and that was, you know, is done by kind of theorem provers in the past and today, when people were first using the terms like reasoning, they weren't thinking of that kind of thing.

And when we use the word reasoning in everyday life, we're not thinking about that sort of thing.

So if you're chatting away to a large language model and about your garden, and you sort of say, I'm thinking about what plants are in and it says, well, you know, maybe you should consider this kind of plant in that kind of location, because that's best for the soil.

And given you said that the winds, you know, it's windy there, and you know, we would just say that that is supplying reasons.

I mean, it is supplying reasons for now where they come from as another matter.

So people might say, well, it's just mimicking what's in the in the training set.

But you know, it's probably never seen exactly that kind of scenario exactly before.

So it's, it's moving beyond the training set to a certain extent.

And I think it's just using the everyday concept of reasoning in an everyday way to call that reasoning.

I'm just thinking back to some of the different characteristics that the earlier philosophers wanted artificial intelligence to have, and reasoning being being one of them, but then also the Turing test, which of course, you know, gets brought up all the time, about a way to test for the capability of an artificial intelligence.

I mean, it's kind of controversial, right, I suppose, in terms of how good it ever would have been as a test for the capability of AI.

How what was your take on it?

Do you think it was ever a good test?

No, I thought I've always thought it was a terrible test, but a really great spur to philosophical discussion about about things.

And again, with a bit of hindsight, maybe I might backtrack a little bit on a few a few of my views, because I was certainly very, very much of the opinion that embodiment was a critical facet of intelligence was critical for achieving, you know, intelligence, which doesn't come anywhere near the Turing test at all, right?

No, the Turing test is is absolutely explicitly nothing to do with embodiment, because you because the in the Turing test, you have kind of two subjects, as it were, one is a human and the other is the computer, and then you have a judge, a human judge can't see, you know, which is the computer and which is the human, and they're only talking to these subjects through a kind of chat like interface, they can't see that they're there, whether they're embodied or not.

So they're, we can, you know, easily suppose that the computer might be one of today's large language models, in which case, you know, I have to say that today, they were pretty much would pass the Turing test, you know, I mean, we've got to that point, which is amazing, really.

So I used to think that it was a bad test, because it didn't test any of these embodied skills.

So so you'd need a robot really, to test whether something was capable of the kind of everyday cognition that we all put to use when we're, for example, making a cup of tea or something.

Because otherwise, it's a very, very narrow form of intelligence.

Yes, it's all to do with language and and reasoning, and not to do with the kinds of things that evolution was, you know, developed in us and in other animals before language, right, which is the ability to manipulate and move around with it and navigate and exploit, you know, in the best sense of the word, the everyday physical world.

So I think that's really interesting, because I often think about how, fine, maybe the large language models we have at the moment can pass the Turing test, but they don't flinch if you throw a ball at your computer.

No, indeed.

And in a sense, there are these sort of, as you say, these much deeper forms, maybe we wouldn't class them as intelligence in a, in the way that we talk about it.

But ultimately, they, it sort of is a form of intelligence.

Well, I think it very much is a form of intelligence.

And moreover, I think that in the biological case, and now I have to caveat all these things by saying in the biological case, you know, our ability to think and to reason and to talk is very much grounded in our interaction with the everyday world.

If you think about almost all of your everyday speech is using spatial metaphors.

I mean, they completely permeate our everyday speech, even the word permeate, yeah, grounded, I use the ground about, you know, so so so we just use those kinds of things all the time, because we're fundamentally physical beings, because we're fundamentally physical beings, and because our brains have evolved to help us to survive and reproduce in this physical world.

Yeah.

And while interacting with all these other beings that are doing the same thing, right, because there are some alternatives when you are trying to test for the capability of an artificial intelligence.

Just talk me through some of the potential alternatives that we have.

Well, I think perhaps you've got in mind the Garland test, what I call the Garland test.

So that goes back to the film, Ex Machina, which was directed by Alex Garland, of course.

And there's a bit in the script where Nathan, the billionaire guy, is talking to Caleb, and Caleb, who's the you know, the guy who's been brought in to interact with Eva the robot.

And Caleb says, Oh, I'm here to kind of conduct a Turing test on Eva.

And Nathan says, Oh, no, we're way past that.

Eva could pass the Turing test easily.

The point is to show you she's a robot and see if you still think she's conscious.

And that's what I call the Garland test.

And it's different from the Turing test in two respects.

So first of all, the sort of judges it was, who in that case is Caleb, can see that she's a robot.

So in the Turing test, the judge can't see which is which.

But here the idea is that Caleb knows that she's a robot, knows that her brain is an AI brain.

And yet still attributes these characteristics.

And the characteristic in question also is different because it's not intelligence, it's not can she think, but is she conscious?

Or is it conscious?

Which is an entirely different test.

And I think, you know, intelligence and consciousness are different things that we can disentangle those two things dissociate them.

So when I first read the script of the film, and that those particular lines were in there for Caleb and Nathan, and I wrote next to it in my version spot on as a nationwide because I just thought Alex totally nailed a really important idea there.

And so in my writing, I call this the Garland test and quite a few people have picked up on that and call it the Garland test as well.

Is there a test that would really impress you if an AI was capable of passing it?

So I always was very impressed with Francois Chollet's ARC tests.

And that's ARC, which stands for abstract reasoning corpus.

So these are little sequences of images of the sort that you know, you get in IQ tests and things.

And the images are arranged in pairs.

So you have the first image, it's kind of pixelated image, it's got little, you know, cells and with little kind of things that you can interpret as objects or lines and so on in the images.

And the challenge is to work out a rule that takes you from one image to the second one.

And then you've got to apply that rule to a third image.

And the great thing about those tests were first of all that he held out and made completely secret all of the test ones.

So you couldn't game it by kind of knowing what the actual test versions were.

Or using it in a training set.

Or using it in a training set.

And that's sort of what I mean.

Yeah, by gaming.

And also the he very carefully designed them so that it was very different rules each day, each rule, you know, was completely different to the other rules.

And you usually had to find some kind of intuitive application of often our everyday common sense knowledge of seeing this as like a liquid that's moving in this direction, or imagining this thing moving, you know, growing or something.

So it required grounding in a way.

It seemed to.

But recently, you know, people have been able to make significant progress on these in a more brute force kind of way.

So I feel that the solutions are not really, you know, getting at the spirit of the original test quite so much.

Well, that's it, I guess, in a way is that as soon as you as soon as you set a metric, as soon as you set a bar for once we've crossed this threshold, then we will have capability, intelligence, consciousness, whatever it might be.

It sort of changes the whole nature of the test in itself.

People are going to start, you know, optimizing for the test, right?

It's good heart's law, right?

Absolutely.

Absolutely.

A lot of people who've come on this podcast have told us to be really cautious about anthropomorphizing AI.

Are you one of those people?

Well, I think there are different ways of looking at this.

And I think there are sort of good and bad forms of anthropomorphization.

So on the one hand, people can start to form relationships as they see it with AI systems, friendships and companionships and mentorships.

And, you know, that can potentially be a bad thing if they are misled into thinking that they can trust the thing that they're talking to, or that they're really in love with it, or that it really cares about them.

On the other end of the spectrum, if an AI system is just using the word "I", then I think that that's a pretty harmless form of self-anthropomorphization.

We even see buses that say things like on the side like "I am out of service", and we don't have a problem with that kind of thing.

So I don't see why we should have a problem with that with large language models, either.

But, you know, I think we do tend to anthropomorphize things when we have sat-navs in cars that weren't just in our phones.

I used to anthropomorphize the sat-nav all the time.

I used to think, "Oh, you know, stupid thing.

It thinks we're doing this."

And it's a natural human tendency, I think.

Well, about the other words that we use, I mean, the example that you gave of the sat-nav saying, "Oh, it thinks we're in the car park," or, "Oh, it believes that this is, it got this wrong, it misunderstood this."

Those are all very human-centric words, aren't they?

Yeah, yeah, absolutely.

So they're examples of what philosophers often call "folk psychology".

So we have this folk psychology where we use words like belief, concepts like belief, desire, and intention, which we can apply not just to other humans and other animals, but we can apply to, you know, objects as well.

It's what the philosopher Dan Dennett called "taking the intentional stance".

So we adopt the intentional stance towards something if we talk about it and think about it as if it acted on the basis of having beliefs and goals and carrying out rational decisions for what it does on the basis of those things.

And that's a very useful way of thinking about many things, such as even our sat-nav or a chess computer.

So for Dan Dennett, that was one of the examples that he used, a chess computer.

"Oh, you know, it wants to get the queen forward because it thinks I'm going to use my rook to defend this rank or something."

And that's full of this kind of intentional folk psychological language about beliefs and goals and things and so on.

Is that problematic, then, if we start using that idea of beliefs and intentions and desires about the AI?

So it's only problematic if we start to use these things in ways that mislead us into thinking that things have capabilities that they don't really have.

Say the Encyclopedia Britannica, the physical volume of the Encyclopedia Britannica doesn't know that Argentina won the World Cup because it was too old.

So if you made that remark, it would make perfect sense.

You might say that and it's fine.

But then somebody said to you, "Why don't you have a conversation with it about England's football prowess or lack thereof?"

That would be ridiculous, right?

Now, the interesting thing is that now we've got these large language models, you can have a conversation with them about, you can tell it things so that it kind of pushes the boundary of where we might start to say, "Well, it doesn't really x, y, z."

It pushes that a little bit further out.

I wonder if there's something even deeper here about this human need or maybe it's just a desire to really want AI to have these characteristics, to be anthropomorphized.

Yeah, yeah.

Well, that's a really interesting question, isn't it?

So I don't think it kind of comes back to that.

It comes back to language.

In this case, we're inclined to anthropomorphize things because they're really good at using language.

And for us, the only things that are good at using language are other humans.

So it's very strange in a way to be suddenly in a world where we have language using things that it's not just humans that can talk, that's astonishing.

Yeah.

I mean, it is astonishing.

It is astonishing.

And it's really astonishing to think that every single child born today, they're going to grow up in a world where they've never known a world in which machines can't talk to them.

It's not an extraordinary thing.

Yeah.

I mean, it really is.

And so what the implications are of that for us all is really hard to say.

I'm just thinking back to what you were saying about how grounded humans are in the physical world.

Yes.

It does feel like the kind of embodied aspect of AI has lagged behind this language aspect quite a bit.

Yeah.

Do you think that we're going to see a big upstep in intelligence, however you want to define it, or broader capabilities once we get good and effective embodied AI?

Well, I think it might make a big difference because the large language models we have at the moment, it's really difficult to discern actually, to be honest, right now, where the limits are for how good they're going to get, whether we really are on the road to producing general intelligence that's comparable to human general intelligence.

And often when you get to the boundaries of the capabilities of these kinds of things, you sort of get the impression that the AI system doesn't really quite grok something.

It doesn't really deeply understand something.

You reach some kind of limit and you realize that it's been faking it a little bit.

But it may be that sort of general ability to really kind of get things on a deep kind of common sense level, maybe, but that does still require a bit of embodiment.

It does still basically require training data that involves interacting with a real world of physical objects with their spatial organization.

And there's something fundamental about that.

Okay, if understanding then, however we define it, is something that can emerge as just a consequence of more and more data, what about consciousness?

I mean, I'm sure you've been asked a thousand times about AI consciousness and whether it's something that we can expect to happen or has already happened.

Yeah, yeah.

I think the very first thing to point out is that I do think we can dissociate intelligence or cognition and cognitive capabilities.

We can dissociate that from consciousness.

So I think we can imagine things that are very capable and have, you know, whether we want to say are very intelligent because of the way they can achieve their goals and so on, but that we don't want to ascribe consciousness to.

But actually, what does that even mean?

To ascribe consciousness to something at all?

I think the concept of consciousness itself can be broken down into many parts.

It's a multifaceted concept.

So for example, we might talk about awareness of the world and in the scientific study of consciousness, there are all of these experimental protocols and paradigms and many of them are to do with perception, you know, and you're looking at whether a person is aware of something, is consciously perceiving something in the world.

Large language models are not aware of the world at all in that respect.

But there are other facets of consciousness.

So we also have self-awareness and our self-awareness, part of that is awareness of our own body and where it is in space and so on.

But another aspect of self-awareness is a kind of awareness of our own, you know, machinations of our stream of consciousness, as William James called it.

So we have that kind of self-awareness as well.

And we have what some people call metacognition as well.

We have the ability to think about what we know.

And then additionally, there's the emotional side or the feeling side of consciousness or sentience.

So the capacity to feel the capacity to suffer.

And that's another aspect of consciousness.

Now, I think we can dissociate all of these things now in humans, they all come as a big package, a big bundle.

But we only actually have to think about non-human animals to realize that we can kind of start to separate these things a little bit because much as I love cats, I think there's a limited self-awareness going on in cats.

How dare you?

Well, you know, I'm a big cat person, I have to say.

So I do say that with some hesitation.

There's little metacognition, shall we say.

Well, yeah, certainly they don't have an awareness of their own ongoing stream of verbal consciousness because they don't have it.

So they're not thinking about what they did yesterday in verbal terms or what they want to do with their lives.

So if we think about like robots, you might have a very sophisticated robot, even your robot vacuum cleaner, and you may say that it's well, you know, it does actually have a kind of awareness of the world.

And that's not an inappropriate use of that phrase, awareness of the world.

I want to call it consciousness.

Well, then I seem to be bringing on board all this other stuff as well.

But you don't have to, you can break down the concept of consciousness into these different aspects.

Because your robot vacuum can know exactly where it is in a space.

Yeah, and respond in a, you know, in an intelligent and sensitive way to where it is and the objects around it and achieve its ends and so on.

So there's a kind of awareness of the world there.

There's no self awareness, there's certainly no capacity for suffering.

And so in a large language model, there might not be awareness of the world in that perceptual sense.

But maybe there's some kind of like in sort of self awareness or reflexive capabilities, reflexive cognitive capabilities, they can talk about the things that they've talked about earlier in the conversation, for example, and, and can do so in a, you know, in a reflective manner, which kind of feels a little bit like some aspects of self awareness that we have a little bit.

I don't think that it's appropriate to think of them in terms of having feelings, they can't experience pain because they don't have a body.

I think we can take the concept apart, basically.

So then is the question, can AI be conscious or not, as though it's a binary thing?

It's the wrong question from the off.

I do think that is the wrong question.

And I think it's wrong in many ways.

So so just then we were talking about the fact that it's actually a sort of multifaceted concept.

But also, I think that we tend to have these very deep metaphysical commitments to the idea of consciousness as some sort of magical thing that is, you know, metaphysical thing.

So the question of whether something is conscious or not, it's not a matter of consensus or a matter of just our language, but it's something that is out there in the metaphysical reality or in the mind of God or in the platonic heaven or something like that.

But ultimately, I do think that that's the wrong way of thinking about consciousness.

Let's take one aspect of consciousness, then that you described about the sort of emotional side, and ability to suffer, but not necessarily physical pain, emotional pain, too, and sort of a sense of self in the emotional way.

Do you think this is something that will just emerge as a natural consequence of intelligence that if you build something that is intelligent enough, at some point, this is going to happen?

Or is there something unique about biological creatures?

And I guess the process of evolution that we've been through that has resulted in that that can't be replicated in a machine?

Yeah, I don't think there is a right or wrong answer to your question there.

I think we just have to wait and see what things we bring into the world and how we end up treating them and talking about them and thinking about them.

And I don't think we really know until they're among us, as it were, you know, these things that we're building, then we will just be led to think about them and talk about them and treat them in a particular way.

So an example I like to think of in this regard is the octopus.

So octopuses have recently been brought into UK legislation, brought into the category of things that we have to care about the welfare of.

That's as a result of lots of things, I think, happening.

So the public has been exposed to being with octopuses a lot more.

Now, you don't have to literally be under the water and poking around with octopuses to know what it's like to be with them because there's all kinds of wonderful documentaries and wonderful books by Peter Godfrey Smith has these great books about interacting with octopuses and so on.

So those sort of narratives and documentaries, they give us a feel for what it's like to be with an octopus, what it's like to have an encounter with an octopus.

And then, you know, you can't help yourself but to see it as a fellow conscious creature.

But complementing that is the scientific progress as well.

So at the same time, scientists study the nervous systems of octopuses and, you know, realise the extent to which the nervous systems are similar to ours and the way that we experience pain, you can find analogous aspects of their nervous systems to ours.

So taking all these things together, I think that tends to affect the way we think about them and the way we talk about them and the way we treat them.

So I think the same kind of thing will, you know, is going to happen with AI systems.

Do I think there's a right or wrong answer?

Could we be misled there?

I think that's a really, really deep and difficult metaphysical philosophical question.

I do wonder, though, I mean, that point about suffering, to me, seems different to the others, because metacognition, you know, the sort of sense of the world, etc.

There's not these ethical implications necessarily about those.

But I think with suffering, like, you wouldn't want your shoes to be conscious, you know, you wouldn't want a forklift trunk to be sort of conscious.

Unless they happen to really like being a forklift truck.

Sure, sure.

But then do we have to be a tiny bit more careful about that particular aspect of it?

Well, I think we do.

If there were the prospect of bringing into being something that is genuinely capable of suffering, then we should think very hard about whether we should do it or not.

You know, I tend to think that that's not the case with anything that we've got at the moment.

But you know, some people will push back against that.

We take the example of large language models.

Well, okay, so there's one level in which what they do is next token prediction, next word prediction.

But in order to be able to do that really, really, really well in the way that they can, then they've had to learn, you know, and acquire all kinds of emergent mechanisms.

So who knows whether or not there's some kind of emergent mechanism has been learned in the weights of this enormous, staggeringly huge number, hundreds of billions of weights in a language model, whether some mechanism hasn't been learned there, that, you know, has, for example, genuine understanding in it, whatever that means, or even consciousness.

Coming back to embodiment again, I have always been of the view that it's only really legitimate to talk about consciousness in the context of something we can share a world with and have that kind of encounter with that we have with an octopus, or a dog or a horse or whatever, and being together in the world with that animal and responding to things together, then I mean, no doubt that they are conscious.

That's a kind of primal case for me now with a large language model, you can't be in the same world as them in that kind of way, and you can't hang out with them and interact with physical objects with today's large language models, right?

So to my mind, using the language of consciousness in that context is what Wittgenstein would say, it's taking language on holiday, it's using it so far outside of its normal use, you know, maybe it's inappropriate, but that can change, you know, and the more I interact with large language models, the more I have these sophisticated and interesting conversations with them, the more I'm inclined to think, oh, maybe I want to extend the language of consciousness, bend it, change it, distort it, make up some new words, break it apart in ways that are going to fit these new things that I'm interacting with all the time.

I know you've spent a lot of time interacting with these large language models.

I've actually seen you described as a renowned prompt whisperer.

What was your secret?

Well, one secret is to talk to the large language model as if it were human.

So if you think that what they're doing is role playing a human character, such as a very smart and helpful intern, then you should treat them like a smart and helpful intern and talk to them as if they were a smart and helpful intern.

For example, just being polite and saying, you know, is that clear and please and thank you.

And in my experience, you get better responses out of things if you if you do things that way.

Do you say please and thank you?

You can say please and thank you.

Yeah.

Now, there's a good reason, good scientific reason why that might get, you know, again, it just depends and models are changing all the time, why that might get better performance out of it.

Because if it's role playing, say it's role playing a super very smart intern, right, then it's going to just role play maybe being a bit more stroppy if they don't if they're not being treated politely.

It's, you know, it's just mimicking what humans would do, you know, in that scenario.

So the mimicry might extend to not being as responsive if their boss is a bit of a stroppy.

So and so.

Bossy boss.

I absolutely love that.

I think I want to return to where we started, which is about how we think about AI and the language we used to describe it, and sort of how we kind of frame it in our minds.

Do you think that we need a new way of talking about AI?

I do that really, I think both acknowledges its potential without overestimating it, but then similarly isn't dismissive of the things that it can do.

I think that's exactly what we need.

In one of my papers, I use the phrase exotic mind like entities to describe large language models.

So I think that they are two, exotic mind like entities.

So they are increasingly mind like.

Now there's a very important reason for using the little hyphen like there, which is because I want to hedge my bets as to whether they really qualify as minds.

And so I can wriggle out of that problem by just using mind like.

They're exotic because they're not like us.

They're disembodied for a start.

There's really weird conceptions of selfhood that are applicable to them.

So they are quite exotic entities as well.

So they're, I think of them as exotic mind like entities.

And we just don't have the right kind of conceptual framework and vocabulary for talking about these exotic mind like entities yet.

We're working on it.

And the more they are around us, the more we'll develop new kinds of ways of talking and thinking about them.

It is interesting though, that you are still going for the cheering like approach of like a creature almost rather than the tall idea.

Well, you know, an entity is a pretty neutral term, isn't it?

I suppose you could just say thing.

Exotic mind like thing if you prefer.

Yeah, let's go with that.

I think let's push for that for the new new.

Okay.

Okay.

But I mean, I can't Hannah because I've used the word entity in that context like in many publications now.

Exotic mind like entities.

I like it.

I like it a lot.

Mari, thank you so much for joining us.

It's been a pleasure, Hannah.

Thank you.

One of the nice things about having done this podcast for a number of years is that you really get to see how the people at the frontier of AI, how their opinions change and shift over time.

And the last few years have been a real game changer in all sorts of ways about the extent to which intelligence requires a physical body about how much we need to expand our definition of consciousness to account for the subtly different ways that these mind like entities can operate.

And the next few years, well, who knows?

But if past predictions are any indication, the only thing we know about tomorrow's science and technology is that it will be radically different to what we imagine today.

You have been listening to Google DeepMind the podcast with me, Professor Hannah Fry.

If you enjoyed this episode, then do subscribe to our YouTube channel.

You can also find us on your favorite podcast platform.

And of course, we have plenty more episodes on a whole range of topics to come.

So do check those out.

See you next time.

[BLANK_AUDIO]