WEBVTT - The Edge of Sentience, with Jonathan Birch 0:00:03.040 --> 0:00:07.040 Welcome to Stuff to Blow Your Mind production of iHeartRadio. 0:00:13.080 --> 0:00:15.600 Hey you welcome to Stuff to Blow your Mind. This 0:00:15.760 --> 0:00:18.720 is Robert Lamb and today I'm going to be chatting 0:00:19.239 --> 0:00:23.919 with Jonathan Birch about his new book, The Edge of Sentience, 0:00:24.120 --> 0:00:28.360 Risk and Precaution in Humans, Other Animals and AI. It 0:00:28.440 --> 0:00:31.960 comes out later this week on November fifteenth in the US. 0:00:32.680 --> 0:00:35.680 Jonathan Birch is a professor of philosophy at the London 0:00:35.720 --> 0:00:39.280 School of Economics and Principal Investigator on the Foundations of 0:00:39.320 --> 0:00:43.760 Animal Sentience Project, a European Union funded project aiming to 0:00:43.800 --> 0:00:47.320 develop better methods for studying the feelings of animals and 0:00:47.440 --> 0:00:50.519 new ways of using the science of animals' minds to 0:00:50.600 --> 0:00:54.520 improve animal welfare policies and laws. In twenty twenty one, 0:00:54.680 --> 0:00:57.440 he led a review for the UK government that shaped 0:00:57.480 --> 0:01:02.000 the Animal Welfare Sentience Act twenty two. In twenty twenty 0:01:02.040 --> 0:01:04.000 two through twenty twenty three, he was part of a 0:01:04.040 --> 0:01:08.080 working group that investigated the question of sentience in AI. 0:01:08.680 --> 0:01:12.080 So I'll definitely be asking him about animals, about AI 0:01:12.360 --> 0:01:15.440 and maybe a few surprises here. So without further ado, 0:01:15.640 --> 0:01:21.440 let's jump right into the interview. Thank you, Hi, Jonathan 0:01:21.440 --> 0:01:22.160 Welcome to the show. 0:01:22.360 --> 0:01:23.840 Hi Robert, thanks for inviting me. 0:01:24.360 --> 0:01:27.320 So the new book is The Edge of Sentience. But 0:01:27.480 --> 0:01:30.440 before we get to that edge and start talking about that, 0:01:31.240 --> 0:01:35.000 how do you define sentience in your work? And what 0:01:35.120 --> 0:01:39.640 are the implications and challenges of agreeing on a working definition? 0:01:40.080 --> 0:01:42.920 Well, so see why I think if sentience is a 0:01:42.920 --> 0:01:46.480 really useful concept. Let's start by thinking about pain that 0:01:46.560 --> 0:01:49.200 I think a lot of us have wondered. Can an 0:01:49.200 --> 0:01:54.080 octopusts feel pain? Can insects feel pain? Can things hurt? 0:01:54.240 --> 0:01:57.720 Can they have that feeling of ouch? And this is 0:01:57.760 --> 0:01:59.600 a great question, but I think it's a bit too 0:01:59.720 --> 0:02:02.720 n because we need to be aware of the fact 0:02:02.760 --> 0:02:07.440 that other animals might have very different mental lives from us, 0:02:08.080 --> 0:02:11.560 and words like pain they might be a bit narrow 0:02:11.639 --> 0:02:15.160 for thinking about what the experiences of other animals are like. 0:02:15.680 --> 0:02:17.959 So it's good to have concepts that are a bit 0:02:18.040 --> 0:02:21.200 broader than the concept of pain, and to have a 0:02:21.280 --> 0:02:29.040 concept that includes other negative feelings like frustration, discomfort, but 0:02:29.120 --> 0:02:32.200 also the positive side of mental life as well, because 0:02:32.400 --> 0:02:38.240 we also care about this. We care about states like joy, excitement, comfort, 0:02:38.960 --> 0:02:43.360 pleasant bodily sensations like the feeling of warmth, and we 0:02:43.400 --> 0:02:45.480 want a concept that is broad enough to include all 0:02:45.520 --> 0:02:47.960 of this, or the negative side and the positive side 0:02:47.960 --> 0:02:50.840 of mental life as well any feelings that feel bad 0:02:51.280 --> 0:02:53.560 or feel good. And this is what the concept of 0:02:53.600 --> 0:02:56.760 sentience is about. The capacities have feelings that feel good 0:02:56.840 --> 0:02:57.720 or feel bad. 0:02:58.520 --> 0:03:02.519 Now to what extent is a different concept from consciousness 0:03:02.639 --> 0:03:04.680 or where do they overlap and where do they differ. 0:03:05.680 --> 0:03:08.080 The problem I have with the concept of consciousness is 0:03:08.080 --> 0:03:11.640 that it can refer to many different things. Sentience isn't 0:03:11.639 --> 0:03:13.799 perfect in that way, but I think it's a bit 0:03:13.880 --> 0:03:19.080 more tightly defined than consciousness because when we talk about consciousness, 0:03:19.080 --> 0:03:25.160 sometimes we're just talking about immediate raw sensation, what it 0:03:25.160 --> 0:03:29.600 feels like to be me right now, the sites, the sounds, 0:03:29.720 --> 0:03:35.080 the odors, the bodily sensations, the pains, the pleasures, and 0:03:35.120 --> 0:03:38.640 so on, and that's quite closely related to sentience. But 0:03:38.760 --> 0:03:41.960 sometimes when we're talking about consciousness, we're talking about things 0:03:41.960 --> 0:03:45.280 that are overlaid on top of that, like our ability 0:03:45.320 --> 0:03:49.280 to reflect on what we're currently feeling, and our sense 0:03:49.280 --> 0:03:54.000 of self, our sense that my current immediate raw experiences 0:03:54.600 --> 0:03:57.920 are not happening in isolation, but they're part of a 0:03:58.000 --> 0:04:01.720 life that extends back in time and extends forwards into 0:04:01.760 --> 0:04:05.480 the future, and I'm self aware. I'm aware of myself 0:04:05.520 --> 0:04:08.800 as existing in time, and these things are much more 0:04:09.520 --> 0:04:14.040 sophisticated than just having those immediate raw sensations. So it's 0:04:14.120 --> 0:04:16.800 very useful to have a term that draws our attention 0:04:17.000 --> 0:04:20.279 to those immediate sensations, and that's what sentience does. 0:04:21.000 --> 0:04:23.119 Now. I realize this is of course a huge question 0:04:23.160 --> 0:04:26.120 that you tack on the book, and I'm not going 0:04:26.160 --> 0:04:29.320 to ask you to regurgitate all of it for us here, 0:04:29.360 --> 0:04:34.520 But where are the least controversial divides between synthiens and 0:04:34.640 --> 0:04:37.839 non sentience in the animal kingdom and where does it 0:04:37.880 --> 0:04:40.320 become more messy or controversial. 0:04:41.160 --> 0:04:45.920 I think it's become very uncontroversial in the last few 0:04:45.960 --> 0:04:49.799 decades to think of all other mammals as being sentient beings. 0:04:50.320 --> 0:04:53.719 And that's a huge change because as a long tradition 0:04:53.760 --> 0:04:57.480 of skepticism in science, going back to Rene Descartes in 0:04:57.520 --> 0:05:02.080 the seventeenth century, but also the so called behaviorists in 0:05:02.120 --> 0:05:05.960 the early twentieth century, who said you should never be 0:05:06.040 --> 0:05:09.480 talking about consciousness or sentience of any kind in relation 0:05:09.680 --> 0:05:12.960 to other animals, and that view has just fallen away. 0:05:12.960 --> 0:05:15.719 And I think it's good that it's fallen away, because 0:05:16.080 --> 0:05:19.120 I think it is pretty obvious that our cats are dogs, 0:05:19.560 --> 0:05:22.479 and in fact all other mammals like pigs, cows, et cetera, 0:05:23.240 --> 0:05:28.080 they do have feelings. And then because of this new consensus, 0:05:28.680 --> 0:05:33.040 the debate, the controversy has started to be more around fishes, 0:05:33.360 --> 0:05:38.480 where you get some genuine doubters, and particularly invertebrates, where 0:05:38.520 --> 0:05:41.520 we move from animals with a backbone to animals without, 0:05:41.600 --> 0:05:44.799 and we're looking at animals separated from us in time 0:05:44.880 --> 0:05:51.920 by over five hundred million years of evolution, like octopuses, crabs, lobsters, insects. Here, 0:05:51.960 --> 0:05:55.720 I think doubts are very common, and it's entirely reasonable 0:05:55.760 --> 0:05:59.600 to think maybe not all invertebrate animals are sentient, and 0:05:59.839 --> 0:06:00.960 a lot of debate around that. 0:06:02.560 --> 0:06:05.560 And you mentioned the octopus already being an example of 0:06:05.560 --> 0:06:08.800 a very complex creature that of course is very distant 0:06:08.839 --> 0:06:12.719 from us, And yeah, how do how do we line 0:06:12.760 --> 0:06:15.840 that up with this idea of sentience, and then how 0:06:15.839 --> 0:06:18.160 do we keep from comparing it, trying to compare it 0:06:18.240 --> 0:06:19.320 too much to what we have? 0:06:19.440 --> 0:06:23.120 And I guess to consciousness, right, Yeah, sentience is a 0:06:23.120 --> 0:06:28.159 good word for pushing us away from anthroperstcentrism and away 0:06:28.160 --> 0:06:33.040 from this assumption that animals have exactly the same feelings 0:06:33.080 --> 0:06:37.880 we do. So does an octopus have pain, Well, it's 0:06:38.000 --> 0:06:41.719 probably not feeling it in the same way that we would. 0:06:42.160 --> 0:06:44.120 It's going to be a state that feels very different 0:06:44.160 --> 0:06:48.360 to the octopus. I think. Is the octopus sentient? Well, yes, 0:06:48.440 --> 0:06:51.479 I think so. The sentience concept is broad enough to 0:06:51.600 --> 0:06:55.239 just capture the whole range of animal mental lives, soever 0:06:55.320 --> 0:06:56.239 much they may vary. 0:06:57.120 --> 0:06:59.960 Now, when it comes to a moral obligation to Cynthia 0:07:00.040 --> 0:07:03.240 in life forms, where I guess and I realized this 0:07:03.279 --> 0:07:05.520 is asking a question where ultimately, ever, they're going to 0:07:05.520 --> 0:07:08.240 be a lot of different cultural differences and so forth. 0:07:08.279 --> 0:07:11.840 But where where are we generally with the idea of 0:07:11.880 --> 0:07:16.120 our moral obligation to sentient life and where are we 0:07:16.320 --> 0:07:17.920 looking to go with it? Or where does what's the 0:07:17.960 --> 0:07:20.480 trajectory seem to be? Again? 0:07:20.560 --> 0:07:23.320 I think there's been a sea change on this in 0:07:23.360 --> 0:07:27.120 recent decades. I think opinions are changing, and they're changing fast, 0:07:28.000 --> 0:07:30.920 and I think changing in a direction I find encouraging 0:07:31.840 --> 0:07:35.080 because it wasn't that long ago. You'd often get people 0:07:35.680 --> 0:07:38.920 denying the idea that other animals can make any moral 0:07:39.280 --> 0:07:42.480 claim on us at all. People would say morality is 0:07:42.520 --> 0:07:47.440 about humans, it's human interests, human rights. Other animals are 0:07:47.440 --> 0:07:50.760 not part of it, and very few people argue that now, 0:07:51.040 --> 0:07:53.880 because I think once you recognize other animals as sentient 0:07:53.960 --> 0:07:57.320 beings that can suffer, that can feel pain, that have 0:07:57.400 --> 0:08:00.400 lives of their own, it becomes very, very hard to 0:08:00.440 --> 0:08:04.960 defend the view that none of this matters ethically or morally. 0:08:05.600 --> 0:08:08.120 Of course it matters, and then the debate is about 0:08:08.840 --> 0:08:12.520 how it matters, how strong are our obligations, And here 0:08:12.560 --> 0:08:15.800 you do get a lot of disagreement. Still, I feel 0:08:15.840 --> 0:08:19.080 like the point of consensus that I'm trying to latch 0:08:19.120 --> 0:08:22.320 onto in my book is that we have a duty 0:08:22.400 --> 0:08:27.440 to avoid causing gratuitous suffering to sentient beings, which is 0:08:27.480 --> 0:08:29.640 to say, if we're going to do something that will 0:08:29.640 --> 0:08:33.959 cause suffering, we have to have a sufficiently good reason 0:08:34.600 --> 0:08:37.600 for doing that thing. And then, of course you get 0:08:37.600 --> 0:08:40.680 debate around what might be a good enough reason. You 0:08:40.720 --> 0:08:43.839 get debate around, for example, whether scientific research might be 0:08:43.880 --> 0:08:47.320 a good enough reason, and there'll always be disagreement about that, 0:08:47.720 --> 0:08:51.120 but the need to have a reason so that we 0:08:51.200 --> 0:08:55.360 cannot be causing suffering gratuitously. This I think everyone really 0:08:55.360 --> 0:08:56.080 can agree about. 0:08:56.800 --> 0:08:59.920 Now you discuss multiple additional cases that seem to exist 0:09:00.080 --> 0:09:02.840 at that edge of syndience, as the title refers to, 0:09:03.000 --> 0:09:05.160 and I'm not going to ask you about all of them, 0:09:05.160 --> 0:09:07.720 but one of the more surprising ones to me, I 0:09:07.720 --> 0:09:09.640 guess this isn't an area that I had not been 0:09:09.640 --> 0:09:12.920 paying close enough attention to in the science news is 0:09:13.080 --> 0:09:18.359 the idea of brain organoids or artificially grown neural tissues. 0:09:18.720 --> 0:09:22.120 I was not aware that they were playing pong. So 0:09:22.200 --> 0:09:24.480 what's the story here and what does it mean for 0:09:24.559 --> 0:09:27.480 our understanding of syndiants in something like this. 0:09:28.200 --> 0:09:32.640 It's an astounding and very exciting emerging area of research 0:09:33.520 --> 0:09:37.559 where you can induce human stem cells to form nerve 0:09:37.640 --> 0:09:41.839 cells to form brain tissue, and you can build structures 0:09:42.200 --> 0:09:46.120 the model regions of the human brain at very very 0:09:46.160 --> 0:09:51.400 small scales, And sometimes researchers are doing this to model diseases. 0:09:51.760 --> 0:09:56.440 They want to model Alzheimer's or ZEKA or fetal alcohol syndrome, 0:09:57.000 --> 0:09:58.840 and this can be a very good way of modeling. 0:09:59.160 --> 0:10:01.280 So if you compare it the alternative, that is to 0:10:01.400 --> 0:10:04.199 use a living animal like a mouse or a rat, 0:10:05.040 --> 0:10:08.320 that has real limitations because the rat's brain is so 0:10:08.400 --> 0:10:11.319 different from the human brain. So this is very exciting 0:10:11.760 --> 0:10:16.320 way of getting better models of diseases. Of course, it 0:10:16.400 --> 0:10:19.960 raises questions as well about well, there must be some 0:10:20.120 --> 0:10:22.520 point at which you really should stop doing this, because 0:10:22.559 --> 0:10:25.600 you've made something too lifelike, you've made something too big, 0:10:26.040 --> 0:10:28.000 you've let it develop for too long, and there's now 0:10:28.040 --> 0:10:31.120 a chance that it will be sentient in its own right. 0:10:31.800 --> 0:10:34.600 And I feel like this is a risk that seems 0:10:34.600 --> 0:10:37.760 particularly striking in cases where what the researchers are trying 0:10:37.800 --> 0:10:43.840 to do is model intelligence, model cognitive functions. That's what 0:10:43.920 --> 0:10:48.520 this system, dishbrain that you were referring to, is doing. 0:10:48.640 --> 0:10:51.720 Because what the researchers did was train it to play 0:10:51.720 --> 0:10:57.360 the video game Pong through interacting with a computer interface, 0:10:57.960 --> 0:11:01.719 and so the system could control the paddle and they 0:11:01.800 --> 0:11:07.880 demonstrated measurable improvement in gameplay over twenty minutes. So by 0:11:07.920 --> 0:11:12.760 getting feedback on its performance, the system was learning how 0:11:12.800 --> 0:11:19.959 to play pom. And really, the thought that we might 0:11:20.000 --> 0:11:22.640 be getting dangerously close to the edge of sentience, I 0:11:22.640 --> 0:11:26.400 think strikes you very clearly when you read about studies 0:11:26.480 --> 0:11:26.800 like this. 0:11:27.920 --> 0:11:30.640 Yeah, especially to your point, the idea that we could 0:11:30.840 --> 0:11:34.160 get there sort of very much by accident in this case, 0:11:35.320 --> 0:11:38.080 you know, trying in part perhaps trying to avoid things 0:11:38.120 --> 0:11:44.080 like cruelty to mouse kind of lad animal. 0:11:44.400 --> 0:11:46.960 Well, this is why I think it would be an 0:11:47.000 --> 0:11:52.200 overreaction to immediately ban all of this research, because that 0:11:52.240 --> 0:11:54.800 would be inconsistent. We need to be consistent in our 0:11:54.840 --> 0:11:58.640 attitudes to different risks, and it's no use if we 0:11:58.840 --> 0:12:02.200 crack down hard on the organoid research in a way 0:12:02.200 --> 0:12:05.160 that just leads to more research being done on obviously 0:12:05.200 --> 0:12:09.440 sentient animals like mice and rats and monkeys and so on. 0:12:10.480 --> 0:12:12.720 We've got to let this research develop because it could 0:12:12.760 --> 0:12:15.520 be replacing animal research, and we have to encourage that. 0:12:16.559 --> 0:12:19.120 At the same time, we need proportionate steps. We need 0:12:19.160 --> 0:12:22.480 to think about what the red lines are so that 0:12:22.520 --> 0:12:23.560 it doesn't go too far. 0:12:33.840 --> 0:12:37.760 Now, another huge question from your book is how would 0:12:37.760 --> 0:12:41.439 we know even AI became sentient and what would it 0:12:41.480 --> 0:12:42.840 mean for us if it did. 0:12:44.880 --> 0:12:47.040 I think we wouldn't though, And this is the big 0:12:47.080 --> 0:12:50.599 fear that we may be rapidly approaching the edge of 0:12:50.640 --> 0:12:53.440 sentience in this case too, with the rate at which 0:12:53.480 --> 0:12:58.679 AI is developing the extraordinary behaviors we're seeing from AI systems, 0:12:59.360 --> 0:13:04.000 and yet understanding of how they work remains incredibly poor. 0:13:04.880 --> 0:13:07.080 And it's not just that the public doesn't understand that 0:13:07.120 --> 0:13:09.720 the people working at the tech companies do understand. The 0:13:09.760 --> 0:13:12.720 people at the tech companies do not understand either. These 0:13:12.760 --> 0:13:17.360 systems are black boxes where you know the architecture, the 0:13:17.400 --> 0:13:20.520 overall architecture that you've programmed the system to have, but 0:13:20.559 --> 0:13:24.400 then you've let it, You've trained it on vast, vast 0:13:24.440 --> 0:13:28.439 amounts of training data, and in the process it's acquired 0:13:28.480 --> 0:13:34.880 these emergent capabilities. It's acquired algorithms that you didn't program 0:13:34.880 --> 0:13:37.640 into it, but that it can now implement to reason 0:13:37.679 --> 0:13:42.439 its way through problems. And we don't know what the 0:13:42.520 --> 0:13:45.120 upper limit is here. We don't know as these systems 0:13:45.160 --> 0:13:48.440 scale up, we don't know what algorithms they might be 0:13:48.480 --> 0:13:51.440 able to acquire. And we don't know whether there's some 0:13:51.600 --> 0:13:55.439 point at which, if you recreate enough of the computations 0:13:55.480 --> 0:13:58.400 that are happening in a human brain, maybe you also 0:13:58.480 --> 0:14:01.360 get the sentience as well, maybe you also start to 0:14:01.400 --> 0:14:03.400 get feeling as well. This is a view that in 0:14:03.400 --> 0:14:07.760 philosophy is called computational functionalism. It's like a long word 0:14:07.840 --> 0:14:11.040 for the idea that if you recreate all the computations 0:14:11.080 --> 0:14:13.960 going on in the brain, nothing else is needed to 0:14:14.000 --> 0:14:17.439 get sentience, you get the sentience as well. And that's 0:14:17.480 --> 0:14:19.560 the possibility we have to take seriously, and it's a 0:14:19.600 --> 0:14:23.920 real risk, and it means we could create sentient AI 0:14:24.960 --> 0:14:27.600 long before we accept that we've done so, or before 0:14:27.640 --> 0:14:29.640 we realize that we've done so. 0:14:29.680 --> 0:14:33.080 This leads me to a question that my wife asked 0:14:33.120 --> 0:14:34.880 me to ask you when I said, hey, do you 0:14:34.880 --> 0:14:37.680 have any questions about synthients and AI and animals and 0:14:37.760 --> 0:14:40.680 so forth? She wanted me to ask should we be 0:14:41.000 --> 0:14:45.200 polite when we're addressing Siri, Alexa or various you know, 0:14:45.440 --> 0:14:49.040 Google Gemini or whatever kind of text based interfaces that 0:14:49.120 --> 0:14:54.640 we're using. Like what because I found myself making like 0:14:54.760 --> 0:14:57.760 going into say Google Gemini, testing it out, just kind 0:14:57.800 --> 0:14:59.880 of like experimenting with it, seeing what's up with it, 0:15:00.440 --> 0:15:03.920 and then after a few exchanges with it, feeling like 0:15:03.960 --> 0:15:06.280 I need to say, oh, well, thank you, that's all 0:15:06.320 --> 0:15:09.000 for today, and feeling like I need to be polite. 0:15:09.360 --> 0:15:13.840 But then also I have caught children, my own child 0:15:14.080 --> 0:15:17.160 once or twice being a little harsh with say Siri, 0:15:18.040 --> 0:15:20.120 or maybe their un alarm is going on too long 0:15:20.160 --> 0:15:22.440 in the morning, that sort of thing. So what are 0:15:22.440 --> 0:15:23.400 your thoughts about. 0:15:23.280 --> 0:15:26.960 Yeah, it's a fascinating question we have As well as 0:15:27.000 --> 0:15:29.800 the book, there's a paper that we just released called 0:15:29.840 --> 0:15:36.480 taking AI Welfare Seriously, and it's it is an issue 0:15:36.480 --> 0:15:42.360 we should take seriously right now because AI systems that 0:15:42.880 --> 0:15:47.400 might realistically be sentient could be with us quicker than 0:15:47.400 --> 0:15:52.880 we expect, indeed, at any time, and I think it's 0:15:52.880 --> 0:15:55.960 great to be having that discussion now about what are 0:15:55.960 --> 0:15:57.880 we going to do about that. The questions it raises 0:15:57.880 --> 0:16:00.720 are absolutely enormous. We don't know how to answer them, 0:16:01.640 --> 0:16:03.920 and I think maybe it's right that a very low 0:16:04.040 --> 0:16:07.680 cost starting point that we can do right now is 0:16:07.760 --> 0:16:12.600 just start trying to cultivate an attitude of respect the 0:16:12.920 --> 0:16:16.640 systems we're currently interacting with. There's every chance they're not sentient, 0:16:17.360 --> 0:16:20.600 but there's no harm in cultivating an attitude of respect anyway. 0:16:21.720 --> 0:16:26.200 And by cultivating that attitude of respect, will be more prepared, 0:16:27.200 --> 0:16:29.320 more prepared for the future where there really might be 0:16:29.360 --> 0:16:34.360 a moral requirement to avoid torturing these systems. 0:16:34.720 --> 0:16:38.520 Now in terms of just identifying potential sentience, and you're 0:16:38.520 --> 0:16:42.240 already outlined like the challenges, if not impossibility of that. 0:16:42.840 --> 0:16:45.119 Can you tell us a little bit about the gaming problem. 0:16:45.480 --> 0:16:48.000 One of the problems we face in this area is 0:16:48.040 --> 0:16:51.680 that if you ask AI whether it feels anything or not, 0:16:53.120 --> 0:16:55.720 answers very a great deal. Sometimes they say yes, sometimes 0:16:55.800 --> 0:16:59.120 they say no, But those answers are not giving us 0:16:59.240 --> 0:17:04.040 very good evidence at all. The problem is that we've 0:17:04.800 --> 0:17:08.840 we've trained these systems to mimic the dispositions of a 0:17:08.880 --> 0:17:12.880 helpful human assistant. So in their training they've got rewarded 0:17:12.920 --> 0:17:17.720 constantly for being as human like as possible. And so 0:17:17.840 --> 0:17:21.520 we have this situation in which we've got reason to 0:17:21.560 --> 0:17:26.119 think our criteria for sentience will be gained, so to speak, 0:17:26.920 --> 0:17:30.920 because the system can serve its objectives of being a 0:17:30.960 --> 0:17:35.600 helpful like a helpful human assistant by mimicking behaviors that 0:17:36.000 --> 0:17:40.680 we see as being persuasive of sentience, in talking as 0:17:40.680 --> 0:17:43.800 if it had a rich internal life, as if it 0:17:43.840 --> 0:17:48.359 had emotions, as if it had sensations. Sometimes developers have 0:17:48.440 --> 0:17:52.159 reacted to that problem by just programming the systems to 0:17:52.200 --> 0:17:55.240 deny their sentience, so they just say, of course, as 0:17:55.280 --> 0:17:59.200 an AI system I don't have any feelings. That isn't 0:17:59.240 --> 0:18:02.000 very helpful either, because that's not evidence that they don't. 0:18:02.680 --> 0:18:06.560 So we're facing this tough situation where the surface linguistic 0:18:06.560 --> 0:18:11.280 behavior is not really giving us any evidence either way. 0:18:11.680 --> 0:18:13.480 To my mind, the message we have to take from 0:18:13.520 --> 0:18:15.840 this is that we need to be doing everything we 0:18:15.920 --> 0:18:19.680 can to look behind the surface linguistic behavior to try 0:18:19.720 --> 0:18:22.639 and understand the inner workings of these systems. Better to 0:18:22.680 --> 0:18:25.600 try and get inside the black box, open it up, 0:18:25.720 --> 0:18:29.960 find out what computations are actually being performed and how 0:18:30.000 --> 0:18:32.119 they relate to those that are being performed in the 0:18:32.200 --> 0:18:35.120 human brain, to identify what I call in the book 0:18:35.160 --> 0:18:38.800 deep computational markers of sentience, and then look for those 0:18:39.560 --> 0:18:42.040 rather than thinking the linguistic behavior will do the job 0:18:42.080 --> 0:18:42.480 for us. 0:18:43.240 --> 0:18:46.159 Now, what do you think about our moral and or 0:18:46.240 --> 0:18:50.360 legal responsibilities concerning sentient AI as we look forward into 0:18:50.400 --> 0:18:52.720 the future, And again you see, as you said, like 0:18:52.720 --> 0:18:55.639 a lot of this is and or could be happening 0:18:55.640 --> 0:18:58.280 a lot faster than many of us might think. But 0:18:58.720 --> 0:19:01.159 you know what does that mean when suddenly we have 0:19:01.880 --> 0:19:06.000 at least reasonable reason to believe a particular AI is sentient. 0:19:06.840 --> 0:19:09.960 It's a huge debate that I really think we should 0:19:10.000 --> 0:19:12.200 be having now. It's great to be having it now. 0:19:13.040 --> 0:19:15.199 In the edge of sentience, I defend this principle I 0:19:15.280 --> 0:19:18.640 call the run ahead principle, which says that in thinking 0:19:18.680 --> 0:19:21.480 about these issues, we really need to be asking what 0:19:21.520 --> 0:19:26.680 would be proportionate to the risks posed both credible future technologies, 0:19:27.359 --> 0:19:31.080 not just the technologies we have now. Because the technology 0:19:31.119 --> 0:19:34.879 is moving too fast and regulation moves very slow. We 0:19:34.920 --> 0:19:38.639 don't want to be in the position where we're totally 0:19:38.720 --> 0:19:41.960 unprepared for what happens, because we would only have a 0:19:42.040 --> 0:19:46.320 debating the current technology rather than the possible future technology. 0:19:46.800 --> 0:19:49.280 So it's absolutely worth debating about if we get to 0:19:49.320 --> 0:19:53.920 that situation where we've got some deep computational markers of sentience, 0:19:54.760 --> 0:19:58.000 and then we find that we have systems displaying those markers, 0:19:59.160 --> 0:20:02.399 so there is a realistic possibility that the system is 0:20:02.480 --> 0:20:07.800 genuinely sentient. We really have to be thinking about what 0:20:07.840 --> 0:20:12.160 does our duty to avoid causing gratuitous suffering require from 0:20:12.240 --> 0:20:16.800 us in this case, and I think it will imply 0:20:16.880 --> 0:20:22.480 ethical limits on what people can actually do to AI systems. 0:20:24.359 --> 0:20:27.359 What those ethical limits are very very hard to say, 0:20:27.680 --> 0:20:32.639 because the welfare needs we can't even really imagine. It 0:20:32.720 --> 0:20:35.119 depends a lot on the precise nature of these systems 0:20:35.160 --> 0:20:39.679 and the way in which they've achieved sentience, whether we 0:20:39.720 --> 0:20:43.760 can say anything about their welfare needs at all. And 0:20:43.800 --> 0:20:47.000 to me, all of this points towards having good reasons 0:20:47.040 --> 0:20:49.840 to desperately try not to develop this technology at all 0:20:49.880 --> 0:20:53.119 if we can. I think currently we're just not ready, 0:20:54.160 --> 0:20:58.640 We're just not in a position to use this technology ethically, 0:21:00.200 --> 0:21:02.720 and so in a way we should be trying to 0:21:02.760 --> 0:21:03.880 avoid making it at all. 0:21:05.040 --> 0:21:07.880 Now in the book, there's at least one example, and 0:21:08.400 --> 0:21:11.639 I apologize I'm blinking on the specific here, but you 0:21:11.720 --> 0:21:16.800 mentioned a fairly recent call for ethical guidelines concerning AI 0:21:16.880 --> 0:21:22.280 development that was dismissed by critics as being mere science fiction. 0:21:22.560 --> 0:21:24.840 Well, Thomas Metsinger, Yeah, yes. 0:21:24.520 --> 0:21:28.399 I believe so. And that struck me as interesting because 0:21:28.440 --> 0:21:32.520 on one hand, we have clearly, at least through science fiction, 0:21:32.720 --> 0:21:34.800 and of course outside of science fiction as well, we've 0:21:34.840 --> 0:21:39.080 been contemplating things like this for decades and decades, and 0:21:39.160 --> 0:21:42.800 yet as we get closer to the reality, the label 0:21:42.840 --> 0:21:46.439 science fiction is also sometimes used to dismiss it as saying, well, 0:21:46.480 --> 0:21:49.320 that is just sci fi. That's not actual things we 0:21:49.359 --> 0:21:51.720 should be worrying about. So I don't know if you 0:21:51.760 --> 0:21:54.560 have any thoughts on to what extents science fiction and 0:21:54.600 --> 0:21:57.440 science fictional thought has prepared us for this or kind 0:21:57.480 --> 0:22:01.600 of created this barrier that prevents us from acting as quickly. 0:22:02.880 --> 0:22:04.840 Yeah, I don't think it has prepared us. Yeah, I 0:22:04.840 --> 0:22:08.240 think that's fair to say, even though we do see 0:22:08.840 --> 0:22:12.919 films like Her for example, about ten years ago that 0:22:13.600 --> 0:22:17.399 now seemed remarkably prescient that no one thought they were 0:22:17.400 --> 0:22:21.119 describing events ten to fifteen years in the future, and 0:22:21.240 --> 0:22:24.200 yet that is the future we now found ourselves in. 0:22:24.480 --> 0:22:28.480 It's extraordinary. But yeah, that doesn't in any way mean 0:22:28.520 --> 0:22:30.680 that we're prepared. And in my work on this, I'm 0:22:30.720 --> 0:22:36.639 trying to develop a sort of centrist position that is 0:22:36.680 --> 0:22:39.840 about avoiding the pitfalls of extreme views on both sides, 0:22:40.600 --> 0:22:44.720 where one extreme you've got people who think that these 0:22:44.760 --> 0:22:48.960 systems already sentient. We can tell from their surface linguistic behavior. 0:22:49.560 --> 0:22:51.760 They just talk as if they have feelings, so we 0:22:51.760 --> 0:22:55.320 should think they do. And I think that's credulous and 0:22:55.359 --> 0:22:59.320 it needs to be avoided. On the other side, there's 0:22:59.400 --> 0:23:05.000 this dismissal of the whole idea that AI could achieve sentience, 0:23:05.960 --> 0:23:08.760 this idea that, of course you need a biological brain. 0:23:09.040 --> 0:23:11.800 Of course you need to be a living animal, and 0:23:11.840 --> 0:23:13.720 we're just not in a position to be confident or 0:23:13.760 --> 0:23:18.080 sure about that. In this well known philosophical position, computational 0:23:18.119 --> 0:23:21.880 functionism might be right, and if it is right, then 0:23:21.920 --> 0:23:24.480 you might not need a biological brain at all, and 0:23:24.560 --> 0:23:27.320 we have to take that seriously as well. So for me, 0:23:27.400 --> 0:23:29.840 it's about finding that middle ground where we're taking the 0:23:29.880 --> 0:23:33.720 issue seriously, but we're thinking that this has to be 0:23:33.760 --> 0:23:36.880 the beginning now of a process where we really try 0:23:36.880 --> 0:23:41.800 and look for robust, rigorous markers and have serious ethical 0:23:41.840 --> 0:23:45.800 debates about what the right response to those markers of 0:23:45.920 --> 0:23:49.560 sentience would be. We could have to be neither no 0:23:49.640 --> 0:24:02.080 knee jerk skepticism or credulousness. 0:24:03.040 --> 0:24:05.840 Now I realized this next question is largely outside the 0:24:06.040 --> 0:24:09.520 scope of this book. But what are the implications for 0:24:09.560 --> 0:24:13.919 the consideration of possible extraterrestrial syndiants as we encounter it 0:24:14.000 --> 0:24:20.080 in potentially encounter it in say a biological or technological form. 0:24:20.800 --> 0:24:23.280 Just make me think of octopuses again, because of course, 0:24:24.560 --> 0:24:27.720 you know they're so alien from us. They look like extraterrestrials. 0:24:27.920 --> 0:24:31.199 But they're not. They're terrestrial, and they're right here on 0:24:31.240 --> 0:24:34.359 Earth right now. So I think it's great to, you know, 0:24:35.720 --> 0:24:39.320 recognize the possibility of forms of sentients very different from 0:24:39.359 --> 0:24:43.560 our own, and then recognize that our actual Earth already 0:24:43.560 --> 0:24:47.400 contains them, and that we can start thinking now about 0:24:47.440 --> 0:24:49.840 those real cases and what we're going to do about 0:24:49.880 --> 0:24:54.240 those real cases. I'm entirely open to the idea that, 0:24:54.280 --> 0:24:56.280 you know, just as there are really alien forms of 0:24:56.359 --> 0:24:59.639