WEBVTT - Lying Robots and the Lies They Tell 0:00:03.800 --> 0:00:06.680 Welcome to Stuff to Blow your Mind from how Stuff 0:00:06.680 --> 0:00:14.000 Works dot com. Hey, welcome to Stuff to Blow your Mind. 0:00:14.040 --> 0:00:17.239 My name is Robert Lamb and I'm Julie Douglas. And 0:00:17.920 --> 0:00:22.480 as of this recording, the robots singularity has probably not occurred. 0:00:23.239 --> 0:00:25.040 Mm hmmm as far as we know. As far as 0:00:25.079 --> 0:00:28.080 we know, but we're not really sure how to define that, right, right, 0:00:28.200 --> 0:00:30.400 It's kind of an evasive term. I mean, roughly, we're 0:00:30.400 --> 0:00:36.400 talking about computers robots reaching the level where they not 0:00:36.560 --> 0:00:41.640 only um um meet us on the same mental plane, 0:00:41.640 --> 0:00:43.959 but they exceed us and they become greater than us, 0:00:44.000 --> 0:00:49.040 and they and then in various scenarios could conceivably play out, 0:00:49.640 --> 0:00:51.640 people really like to have fun with fun with the 0:00:52.040 --> 0:00:55.240 darker images, they just leave us in the dust, like literally, yeah, 0:00:55.320 --> 0:00:57.240 like they just they take over the world, or they 0:00:57.280 --> 0:01:00.800 just decided to destroy all humans, or or our lives 0:01:00.840 --> 0:01:07.039 become so robotic and so enhanced by robotics that that 0:01:07.120 --> 0:01:09.040 the the human part of our life becomes kind of 0:01:09.080 --> 0:01:11.200 a kind of a core to it. You know, it's 0:01:11.240 --> 0:01:14.160 just but not a necessity of it. Well, I think 0:01:14.200 --> 0:01:18.119 I love talking about the singularity because it does um, 0:01:18.160 --> 0:01:21.360 it brings up so many different points of view, and 0:01:21.640 --> 0:01:23.600 you and I have talked about it plenty and always 0:01:23.640 --> 0:01:28.120 feel like you have like this very um positive outlook 0:01:28.520 --> 0:01:31.800 on the Singularity, whereas sometimes I get paranoid and I think, 0:01:31.840 --> 0:01:35.640 oh no, to be taken over says when you don't 0:01:35.640 --> 0:01:37.600 even come into work and we have to call and 0:01:37.640 --> 0:01:40.160 se who we are because it's the Singularity and anxiety 0:01:40.240 --> 0:01:42.720 is getting taking over. Yes, those are the days when 0:01:42.720 --> 0:01:46.360 I've locked myself in my bunker. Yeah. Whereas I tend 0:01:46.360 --> 0:01:47.840 to look at it like, you know, if the if 0:01:47.880 --> 0:01:51.440 the coffee machines become as smart as humans, then we're 0:01:51.480 --> 0:01:54.600 just gonna have a really great coffee. I mean, that's sea. 0:01:54.800 --> 0:01:57.160 I love that that that's the sort of positive pluck 0:01:57.240 --> 0:02:00.800 that I'm talking about here. Um, But in that context, 0:02:00.840 --> 0:02:03.600 you know, it does it seems Michael, Okay, this is 0:02:03.640 --> 0:02:05.400 something that could be far away, or it could be 0:02:05.400 --> 0:02:09.200 something that's really near. We don't know, but robots are 0:02:09.200 --> 0:02:13.239 having a fundamental impact on our lives in myriad ways, right. Yeah. 0:02:13.280 --> 0:02:17.280 I remember I was looking into some information about recruiting 0:02:17.360 --> 0:02:20.720 into computing programs and they pointed out that there's virtually 0:02:20.760 --> 0:02:25.320 no field of study anymore that does not contain programming 0:02:25.480 --> 0:02:28.480 and uh and computerization, and it's going to get to 0:02:28.480 --> 0:02:30.919 where there's going to be no part of of our 0:02:30.960 --> 0:02:33.640 society that isn't touched by robotics in some way, shape 0:02:33.680 --> 0:02:36.600 or form robotics or you know, or or really high 0:02:36.720 --> 0:02:40.440 end programming or us even being touched by robots, right, 0:02:40.480 --> 0:02:43.520 I mean we've talked about Roxy the sex Spot literally 0:02:43.520 --> 0:02:45.720 touched by it. Well, I don't know if the robots 0:02:45.760 --> 0:02:49.080 doing in this situation, the robots is being touched, I think, Oh, 0:02:49.120 --> 0:02:51.119 but they are, I'm sure are very soon there will 0:02:51.160 --> 0:02:53.880 be touching. Yeah, but it's very hard to argue this 0:02:53.919 --> 0:02:59.639 consensual that's true unless it's programmed. I supposed Yeah, that's 0:02:59.639 --> 0:03:03.200 a whole party. Um, And we will get to that definitely. 0:03:03.440 --> 0:03:06.680 But again, thinking about robots and how they're being used 0:03:06.720 --> 0:03:10.760 in day to day life, we actually turn to Dr 0:03:10.919 --> 0:03:13.640 Ronald Arkin. He's the Regent's Professor in the School of 0:03:13.680 --> 0:03:16.560 Interactive Computing and is the director of the Mobile Robot 0:03:16.680 --> 0:03:20.800 Laboratory at the Georgia Institute Technology. Yeah, it's not a 0:03:20.800 --> 0:03:23.280 mobile lab though. It doesn't you know, it's not like 0:03:23.600 --> 0:03:25.400 it doesn't break away from the rest of the building 0:03:25.480 --> 0:03:30.960 or anything. And uh, galvanting around Atlanta. No, No, but 0:03:31.240 --> 0:03:36.119 he and his team, along with research engineer Alan Wagner, 0:03:36.400 --> 0:03:38.400 they've blazed the way when it comes to robotics and 0:03:38.440 --> 0:03:44.000 in particular thinking about robotic ethics, and they have created 0:03:44.040 --> 0:03:48.200 a robot or robots that can deceive. Yeah, And it 0:03:48.200 --> 0:03:52.400 all boils down to an algorithm, or more specifically a 0:03:52.440 --> 0:03:57.400 cognitive deception modeling algorithm or or a series of algorithms, 0:03:57.720 --> 0:04:01.760 and it involves not not only only giving the robot 0:04:02.120 --> 0:04:06.640 the capacity to fool another robot or a human, but 0:04:06.720 --> 0:04:10.600 giving it the ability to decide when to fool another human. 0:04:10.960 --> 0:04:13.240 Because that's the important thing. We all have the ability 0:04:13.280 --> 0:04:15.200 as humans. We all have the ability to see the 0:04:15.240 --> 0:04:18.880 ability to lie. And one of the the the the 0:04:18.960 --> 0:04:21.240 core things we have to decide in life is when 0:04:21.279 --> 0:04:23.560 to tell the truth and when to lie. And if 0:04:23.600 --> 0:04:26.279 it's if you're out of balance on that, life can 0:04:26.320 --> 0:04:28.080 get pretty messy and nobody will want to hang out 0:04:28.120 --> 0:04:31.799 with you either because you've never um, you never code 0:04:31.839 --> 0:04:35.560 anything um for consumption, you never you know, dish out 0:04:35.560 --> 0:04:37.880 a white liar two, or you just lie all the 0:04:37.920 --> 0:04:40.880 time and are kind of a jerk. Yeah. Yeah, So 0:04:41.000 --> 0:04:43.479 I mean we wouldn't want to create jerk robots. But 0:04:43.560 --> 0:04:45.760 the problem is that that robots in their simple form 0:04:45.760 --> 0:04:49.119 are they're they're robotic, they're they're they're they're not gonna 0:04:49.320 --> 0:04:52.080 they're not gonna code anything um to make it go 0:04:52.120 --> 0:04:54.719 down easier. Yeah. Yeah, So it is really interesting, like 0:04:54.760 --> 0:04:56.680 how do you how do you do this? How do 0:04:56.720 --> 0:04:59.720 you what? What is deception? Anyway, even when when you're 0:04:59.760 --> 0:05:02.520 talking about a robot And so we actually talked to 0:05:03.279 --> 0:05:06.120 Dr Arkin about this, and this is what he had 0:05:06.160 --> 0:05:10.160 to say about deception and what it is defined as 0:05:10.200 --> 0:05:13.080 in robotics. It's actually a definition that we borrowed from 0:05:13.080 --> 0:05:16.719 cognitive science, which is a false communication that tends to 0:05:16.760 --> 0:05:21.200 benefit the communicator. So the important thing is that you 0:05:21.240 --> 0:05:26.080 will benefit by saying something that is untrue. Now in 0:05:26.120 --> 0:05:28.520 the process has often been referred to as lying robots 0:05:28.520 --> 0:05:31.960 and the like as well too. It's a false communication. 0:05:32.120 --> 0:05:34.400 Is lying leaving a trail on the ground. I don't know, 0:05:34.520 --> 0:05:37.680 that's a verbal act. I could argue this is more 0:05:37.680 --> 0:05:41.839 expansive than simply lying. In this case, it's basically trying 0:05:41.880 --> 0:05:44.760 to create a deception to make someone believe something that 0:05:44.920 --> 0:05:49.840 isn't true and act in a way. The work that 0:05:49.920 --> 0:05:53.919 we looked at in the military. For the military was 0:05:53.960 --> 0:05:57.039 more a fundamental question of what deception is and how 0:05:57.040 --> 0:06:01.760 it could be employed. So it's not soly writing up 0:06:01.800 --> 0:06:04.719 look up tables and saying these are the sets of actions, 0:06:05.040 --> 0:06:07.080 as I think I mentioned. If I didn't, there's an 0:06:07.160 --> 0:06:09.960 entire field manual for the U. S. Army on deception 0:06:10.279 --> 0:06:14.640 deceptive techniques, because it's important, it's crucially important in an 0:06:14.640 --> 0:06:19.240 appropriate way to conduct warfare. Uh. But this isn't just 0:06:19.440 --> 0:06:22.000 saying do this in this set of circumstances, do this 0:06:22.040 --> 0:06:24.479 in this set of circumstances. This is I have to 0:06:24.600 --> 0:06:28.479 understand you. I have speaking as a Roman. I have 0:06:28.560 --> 0:06:32.279 to understand you and learn what your potential outcomes are. 0:06:32.480 --> 0:06:34.680 And one of the other interesting factors in that paper, 0:06:34.680 --> 0:06:36.600 which I was about to mention as well too, is 0:06:36.680 --> 0:06:39.960 that it was observed, at least in this narrow case 0:06:40.160 --> 0:06:43.680 that we studied, the more sensors that you have, the 0:06:43.760 --> 0:06:47.520 more the easier it is to deceive you, which it 0:06:47.680 --> 0:06:50.800 was not intuitively obvious. You think, gosh, you know, if 0:06:50.880 --> 0:06:53.560 I have more ways to understand what's around me, I 0:06:53.680 --> 0:06:58.279 would be more likely to get the right action. But 0:06:58.360 --> 0:07:02.440 deception works in this particular, it appears due to the 0:07:02.480 --> 0:07:07.680 fact that we can exploit multiple channels of information. You know, um, 0:07:07.720 --> 0:07:11.480 hearing that, I can't help but be reminded of Isaac Asimov. 0:07:12.320 --> 0:07:14.800 He of course had the classic book I Robot, which 0:07:14.880 --> 0:07:17.920 is a series of short stories about about robots, and 0:07:17.920 --> 0:07:20.760 they're very they're very analytical. They're all about, you know, 0:07:20.840 --> 0:07:23.640 robots of being the three rules of robotics, and humans 0:07:23.680 --> 0:07:26.760 trying to figure out how the robots are are are behaving. 0:07:27.000 --> 0:07:29.640 And there's a story in there called Little Lost Robot. 0:07:29.720 --> 0:07:32.200 And and in this a researcher loses his temper and 0:07:32.280 --> 0:07:35.680 swears that a particular robot and tells it to get lost, 0:07:36.040 --> 0:07:40.360 and it does so, and then the chief robo psychologist, Dr. 0:07:40.440 --> 0:07:43.600 Susan Calvin comes in to find it. Um, So it 0:07:43.640 --> 0:07:46.400 couldn't help it be reminded of that. It's a fairly 0:07:46.440 --> 0:07:50.280 lighthearted tale. Yeah. Actually, that's that's interesting because just in 0:07:50.280 --> 0:07:55.040 in the context of the experiment that Arkin and Wagner did, 0:07:56.040 --> 0:07:59.120 they used interdependence theory and game theory to develop the 0:07:59.520 --> 0:08:04.160 algorithms that tested the value of deception in a specific situation, 0:08:04.880 --> 0:08:08.360 and the situation had to satisfy two key conditions. To 0:08:08.480 --> 0:08:11.800 warrant deception. There has to be a conflict between the 0:08:11.880 --> 0:08:15.200 deceiving robot and the seeker, and the deceiver must benefit 0:08:15.200 --> 0:08:18.040 from the deception. So we're talking about here, is like 0:08:18.080 --> 0:08:21.600 an elaborate hide and seek game or not so elaborate, 0:08:21.680 --> 0:08:25.360 essentially um to test or algorithms. They ran twenty hide 0:08:25.360 --> 0:08:28.840 and seek experiments with two autonomous robots, and they had 0:08:28.920 --> 0:08:32.480 colored markers which they lined up along three potential pathways 0:08:32.920 --> 0:08:35.360 to locations where the robot could hide. And so the 0:08:35.440 --> 0:08:38.800 hider robot randomly selected a hiding location from the three 0:08:38.880 --> 0:08:42.160 location choices and then moved toward that location, knocking down 0:08:42.160 --> 0:08:44.960 colored markers along the way. So once it reached a 0:08:45.000 --> 0:08:47.520 point past the markers, the robot change course and then 0:08:47.600 --> 0:08:50.160 hid in one of the other two locations, and the 0:08:50.280 --> 0:08:53.120 presence or the absence of standing markers indicated the hyder's 0:08:53.160 --> 0:08:55.679 location to the seeker robot. Okay, well this is this 0:08:55.760 --> 0:08:59.080 is like a very basic form of deception. Uh. We 0:08:59.120 --> 0:09:00.600 see this in movies all the time where the hero 0:09:00.720 --> 0:09:03.280 is running through an office building trying to escape somebody. 0:09:03.280 --> 0:09:05.040 So what does he do? He runs? He or she 0:09:05.520 --> 0:09:07.439 runs to this door at the end of the hallway, 0:09:07.760 --> 0:09:10.040 opens it real wide, let's it shut, but then goes 0:09:10.080 --> 0:09:13.040 off in another direction. The guys they're chasing him work 0:09:13.080 --> 0:09:16.080 hows come around the corner and see the door slamming. 0:09:16.080 --> 0:09:18.319 They're like, ah, they went that way, chase them. Or 0:09:18.360 --> 0:09:21.080 it's like if you if you're in the forest and 0:09:21.200 --> 0:09:23.760 you you're making your way past these branches, if you 0:09:23.800 --> 0:09:27.440 started snapping branches along one direction and then backtracked and 0:09:27.480 --> 0:09:30.240 went somewhere else to and then that your pursuers would 0:09:30.240 --> 0:09:32.199 come through and they'd say, well, look, clearly they went 0:09:32.240 --> 0:09:34.760 this way because this is where the branches are snapped off. Right, right, 0:09:34.880 --> 0:09:37.040 It seems basic, right, I mean, and it is and 0:09:37.440 --> 0:09:40.160 um in the sense the physical sense that we think of, 0:09:40.640 --> 0:09:43.120 But think about all of that information, all that data 0:09:43.160 --> 0:09:45.400 that you have to absorb, and then the choices that 0:09:45.440 --> 0:09:47.040 you have to make on that that we take that 0:09:47.080 --> 0:09:50.719 for granted right with our brains, but trying to program 0:09:50.760 --> 0:09:54.480 program that in a robot is particularly interesting and quite 0:09:54.480 --> 0:09:58.320 a challenge. And so what seems rudimentary now and is 0:09:58.400 --> 0:10:01.960 rudimentary in terms of robotic we know, will become much 0:10:02.000 --> 0:10:04.120 more nuanced later on. And so that's why a lot 0:10:04.160 --> 0:10:06.720 of people are looking at this deception, even if it 0:10:06.880 --> 0:10:09.760 is on this level sort of hide and seek level, 0:10:10.040 --> 0:10:12.079 and saying wow, okay, you can do that in robots. 0:10:12.120 --> 0:10:14.120 Now what what is that going to look like in 0:10:14.280 --> 0:10:17.800 five years? And so um ar can actually talked a 0:10:17.800 --> 0:10:20.280 little bit about the attention the experiment received in that 0:10:20.360 --> 0:10:24.319 regard and how programming a robot to deceive is really 0:10:24.360 --> 0:10:27.640 similar to thinking like a con man a significant amount 0:10:27.840 --> 0:10:30.880 of press. I guess it's the best way to put 0:10:30.840 --> 0:10:33.640 an attention is another way to put it, which I 0:10:33.679 --> 0:10:38.600 would contend is probably somewhat disproportionate to the results that 0:10:38.640 --> 0:10:44.040 we obtained, but it's it is a controversial piece of work. Nonetheless, 0:10:44.160 --> 0:10:48.240 it was the first in depth study of the phenomena 0:10:48.400 --> 0:10:52.200 of robot deception and the ability of a robot to 0:10:52.200 --> 0:10:56.960 deceive other robots or potentially human beings as well. UH 0:10:57.000 --> 0:11:01.360 and we look very closely at interdependent theory, a cognitive 0:11:01.360 --> 0:11:05.520 science model as well as game theory as the basis 0:11:05.960 --> 0:11:11.120 for understanding two things about robotic deception. The first is 0:11:11.640 --> 0:11:15.080 when is it appropriate to deceive? Because you don't want 0:11:15.160 --> 0:11:18.200 a robot to be deceiving people all the time, else 0:11:18.240 --> 0:11:20.800 it would not have any value and we were typically 0:11:20.800 --> 0:11:23.800 looking at in this early work or would'll all call 0:11:23.880 --> 0:11:28.559 one shot deception trying to deceive someone just once that's 0:11:28.600 --> 0:11:32.199 good enough, kind of like a conman. The second is 0:11:32.240 --> 0:11:34.720 how to be able to accomplish that, And I could 0:11:34.720 --> 0:11:36.680 talk a little bit more in depth about those, but 0:11:36.720 --> 0:11:40.120 I should say that much of the work that was 0:11:40.160 --> 0:11:44.240 based on some of Alan's earlier work on trust, because 0:11:44.360 --> 0:11:48.240 as any good command knows, a precursor to deception is 0:11:48.280 --> 0:11:51.000 the establishment of trust. But we were interested in that 0:11:51.040 --> 0:11:56.079 particular case of learning how a robot could trust a 0:11:56.160 --> 0:12:00.720 human being, rather than the classic case for man machine systems, 0:12:01.320 --> 0:12:04.640 how can a human learn to trust a robot? There 0:12:04.679 --> 0:12:09.760 are many instances such as UH nine eleven, for example, 0:12:09.920 --> 0:12:16.760 when people act in ways that are improper to say 0:12:16.760 --> 0:12:20.560 the least, and automation should be able to override them 0:12:21.040 --> 0:12:23.960 when they are doing that. Machines should know when not 0:12:24.080 --> 0:12:27.520 to trust the human being. And using the same kind 0:12:27.600 --> 0:12:30.680 of models and the same kind of situational analysis, but 0:12:31.120 --> 0:12:36.760 now looking at ways in which we could induce the observer, 0:12:37.080 --> 0:12:39.640 which will call the mark in this particular case, which 0:12:39.679 --> 0:12:42.480 is the language you would use in this UH, that 0:12:42.559 --> 0:12:46.200 we could induce a outcome belief in the mark that 0:12:46.280 --> 0:12:49.280 in action they would take would be more beneficial to 0:12:49.360 --> 0:12:53.880 them then would be if they took another action, But 0:12:54.000 --> 0:12:57.160 it actually ends up being more beneficial to the robot 0:12:57.160 --> 0:13:01.760 in this particular case is the UH strategy that we 0:13:01.880 --> 0:13:05.400 used to address the how imagine a military situation and 0:13:05.840 --> 0:13:08.240 the robot has valuable information or in and of itself, 0:13:08.240 --> 0:13:10.240 it's a valuable resource and does not want to be 0:13:10.280 --> 0:13:15.400 captured or reverse engineered or whatever the case may be. Now, granted, 0:13:15.440 --> 0:13:18.760 you could put self destruction capabilities in it, so it 0:13:18.800 --> 0:13:20.760 would not be the case, but that destroys the asset 0:13:20.800 --> 0:13:22.800 that you would want to preserve, and it may have 0:13:22.880 --> 0:13:27.240 valuable information of its own in its own right. Uh. Strangely, 0:13:27.280 --> 0:13:30.080 are are two D two and the original Star Wars 0:13:30.400 --> 0:13:33.199 comes to mind. You know, get this to Obi wan Kenobi. 0:13:33.640 --> 0:13:39.959 That wasn't our motivation? Um so what not just playing 0:13:40.040 --> 0:13:42.680 hide and seek in the sense of some other folks 0:13:42.720 --> 0:13:44.760 have done this with robots. Is robot finds a good 0:13:44.800 --> 0:13:47.400 place to hide and hopes for the best. But in 0:13:47.440 --> 0:13:52.040 this case, the robot models the pursuer using a very 0:13:52.040 --> 0:13:54.640 crude version of what's called theory of mind, tries to 0:13:54.760 --> 0:13:57.760 establish what would the pursuer do in these particular sets 0:13:57.800 --> 0:14:01.000 of circumstances and leaves a as a consequence of that 0:14:01.040 --> 0:14:03.960 leaves a false trail. It's kind of like uh, putting 0:14:04.000 --> 0:14:06.800 mud down or tracks down and saying I'm over here, 0:14:06.920 --> 0:14:10.280 but then backtracking and hiding in a different location. So 0:14:10.640 --> 0:14:12.679 it is the belief that the robots belief that if 0:14:12.679 --> 0:14:15.600 the pursuer sees that trail, the pursuer will move in 0:14:15.640 --> 0:14:18.760 that direction. And if that is indeed the case, the 0:14:18.840 --> 0:14:21.440 robot highs and some other location, the pursuer goes down 0:14:21.440 --> 0:14:24.320 that trail and then the robot can escape after that. 0:14:24.680 --> 0:14:27.800 So that's the notional aspect in that particular case. So 0:14:27.840 --> 0:14:30.720 in a way, we're kind of talking about robot original 0:14:30.760 --> 0:14:34.360 sin here. You know, up until this point, they've we've 0:14:34.400 --> 0:14:36.440 just had robots doing what the programmed to do. And 0:14:36.480 --> 0:14:40.400 if they say an industrial robot just you know potent, say, 0:14:40.480 --> 0:14:43.320 you know, welds a man's face off, it's not doing 0:14:43.360 --> 0:14:47.040 that with any intent to weld someone's face off. But 0:14:47.160 --> 0:14:50.160 but in this scenario, we're we're creating a robot that 0:14:50.240 --> 0:14:52.920 can lie and and I find it really interesting that 0:14:52.960 --> 0:14:55.080 it's you know, we're talking about a very simple form 0:14:55.160 --> 0:14:57.560 of lying, a very simple form of deception that will 0:14:57.600 --> 0:15:00.760 grow into more complicated forms of behay. Here, it's it's 0:15:00.800 --> 0:15:02.280 kind of like looking at, say the first time, like 0:15:02.480 --> 0:15:07.560 a child lies, or even just how lies themselves have 0:15:07.680 --> 0:15:10.480 the tendency to start very small and then and then, 0:15:10.520 --> 0:15:12.720 if unresolved, to grow and grow and grow and to 0:15:12.800 --> 0:15:18.080 become this greater, more complicated and complex thing. Yeah, it is, 0:15:18.440 --> 0:15:21.720 but when you think about lies, you also know that there's, uh, 0:15:21.760 --> 0:15:25.080 there's another side to it. That lives can be really helpful, 0:15:25.440 --> 0:15:28.400 right they can. Actually we do it for many reasons. 0:15:28.400 --> 0:15:30.880 But sometimes we do it to spare people's feelings, or 0:15:31.000 --> 0:15:33.200 there might be a dangerous situation and you need to 0:15:33.240 --> 0:15:36.440 lie about something. I don't know. Paramedics arrive on a 0:15:36.520 --> 0:15:39.120 scene the you know, the guy asked, K, doc, am 0:15:39.120 --> 0:15:42.160 I gonna make it that the duck's not gonna say maybe, 0:15:42.360 --> 0:15:44.200 or you know, the duck's gonna say hanging there. We 0:15:44.200 --> 0:15:46.760 can make it right, right, So, I mean, they're essential 0:15:46.800 --> 0:15:49.440 to our own existence, and they're certainly essential to warfare. 0:15:49.760 --> 0:15:51.920 And so that's when you look at something like these 0:15:51.920 --> 0:15:56.920 what they're being called the septicons, these deceiving robots UM 0:15:57.000 --> 0:15:59.200 that they could actually be very useful in search and 0:15:59.240 --> 0:16:03.400 rescue mission UM. And of course there's the idea that 0:16:03.400 --> 0:16:06.400 they could be more fully engaged in battle. But of 0:16:06.440 --> 0:16:08.360 course the challenge is to be able to give robots 0:16:08.360 --> 0:16:11.520 the ability to judge a situation, act accordingly and as 0:16:11.600 --> 0:16:15.920 ethically as possible. So again that's that's where these um, 0:16:15.960 --> 0:16:19.640 these sort of uncomfortable bits of information start to butt 0:16:19.680 --> 0:16:21.520 up against each other because it's like, here, on the 0:16:21.560 --> 0:16:24.000 one hand, we have this great technology that can do this. 0:16:24.600 --> 0:16:28.120 On the other hand, you know, it's you know, work humans, 0:16:28.120 --> 0:16:31.440 and we know that we are programmed not necessarily to 0:16:31.480 --> 0:16:36.560 be the most altruistic beings. UM, So how can we 0:16:36.600 --> 0:16:39.920 create something that's not necessarily in our image, but it's 0:16:40.000 --> 0:16:45.520 better than us. H. It's an interesting proposition. So there's 0:16:45.520 --> 0:16:48.840 a good scenario of of that sort of judgment center 0:16:49.360 --> 0:16:52.840 that we have that we're trying to actually finance, hopefully 0:16:53.440 --> 0:16:58.360 UH in robots. And the scenario at a robotic level 0:16:58.560 --> 0:17:01.520 is UM what Arkin is talking about when he talks 0:17:01.560 --> 0:17:04.159 about instilling in robins a more sophisticated He wants a 0:17:04.160 --> 0:17:07.199 way to assess the situation and act accordingly. UH. This 0:17:07.240 --> 0:17:08.960 is from a New York Times article. It's called a 0:17:09.000 --> 0:17:12.840 soldier taking orders from its ethical judgment center. Uh. They 0:17:12.840 --> 0:17:15.960 talked to Dr Arkin about this computer model, and it's 0:17:16.000 --> 0:17:20.040 a robot pilot who flies past a cemetery in spots 0:17:20.040 --> 0:17:22.800 a tank at the entrance, and this is the target 0:17:22.840 --> 0:17:26.639 in this scenario. But there's a group of civilians in 0:17:26.680 --> 0:17:29.359 this computer model, and they're gather that the cemetery and 0:17:29.359 --> 0:17:32.399 the robot pilot considers the data and decides to keep going. 0:17:33.119 --> 0:17:35.679 But soon it spots another tank out in the field 0:17:35.720 --> 0:17:38.840 all alone, and it decides to fire on it. So again, 0:17:38.880 --> 0:17:43.840 here are these different models that scientists and researchers are 0:17:43.880 --> 0:17:48.480 trying to put together in order to you know, assemble 0:17:48.520 --> 0:17:52.159 this sort of judgment center for robots themselves. Uh. And 0:17:52.200 --> 0:17:55.280 there's obvious limitations to the technology right now as it stands. 0:18:00.480 --> 0:18:04.040 This presentation is brought to you by Intel sponsors of tomorrow. 0:18:08.480 --> 0:18:11.399 But there's the possibility that machines could one day assess 0:18:11.440 --> 0:18:15.800 situations with the advantage of not engaging in our something 0:18:15.840 --> 0:18:18.600 like our own confirmation by us, you know, where we 0:18:18.640 --> 0:18:22.480 sort of perpetuate information that's faulty information because we wanted 0:18:22.520 --> 0:18:26.240 to fit with our worldview. Um, and then also we 0:18:26.320 --> 0:18:30.280 get impassioned about things, right, We don't always think clearly, 0:18:30.359 --> 0:18:35.679 particularly in cases of war. Right. Yeah, So we taught 0:18:35.840 --> 0:18:39.720 to Arkin, and we wanted to to find out why 0:18:39.800 --> 0:18:43.480 people might be frightened of this part, this proposition where 0:18:43.480 --> 0:18:46.440 we might have a technology that could essentially help us 0:18:46.920 --> 0:18:50.720 one day to make better choices in in the war. 0:18:51.240 --> 0:18:53.760 But also you know why we might be a little 0:18:53.760 --> 0:18:55.520 bit freaked out by that. Well, I mean, the basic 0:18:55.600 --> 0:18:58.600 prospect is making war easier to wait, like easier to 0:18:58.640 --> 0:19:00.159 decide to wage war. I mean, that's one of the 0:19:00.440 --> 0:19:06.040 big arguments against robotic warfare is that it becomes if 0:19:06.040 --> 0:19:07.960 there are no human if there are fewer human lives 0:19:08.400 --> 0:19:11.600 on on our side on the line, and fewer chances 0:19:11.800 --> 0:19:15.280 of there being some sort of a horrible headline grabbing 0:19:15.560 --> 0:19:20.439 scenario in terms of enemy casualties, then why not declare 0:19:20.480 --> 0:19:24.879 war right over any little thing, over energy crisis, you 0:19:24.920 --> 0:19:27.840 need some oil, Just declare war because the robots. You're 0:19:27.840 --> 0:19:30.520 not gonna lose anybody. You're not gonna have grieving families 0:19:30.600 --> 0:19:35.320 voting uh, voting against you in an upcoming election. And 0:19:35.480 --> 0:19:37.719 if your robots can, you know, are really good at 0:19:37.880 --> 0:19:41.280 at not um causing a whole bunch of civilian deaths. 0:19:41.280 --> 0:19:45.919 Then uh, it's more surgical and strategic. Ah see. And 0:19:46.040 --> 0:19:49.320 this this is what is so interesting that Arcan has 0:19:49.400 --> 0:19:52.480 to say about this UM, particularly when you talk about 0:19:52.600 --> 0:19:55.560 protocols of war UM. And this is what he has 0:19:55.560 --> 0:19:58.359 to say about it. Well, we need to talk about it. 0:19:58.400 --> 0:20:00.520 Is the most important thing is I've offered said that 0:20:00.560 --> 0:20:04.080 the research I have done in that particular space UM 0:20:04.240 --> 0:20:07.520 is one aspect, But the discussion that that research engenders 0:20:07.600 --> 0:20:10.040 is as important as the research itself, at least to me. 0:20:10.520 --> 0:20:12.359 And as such, I spent a lot of time talking 0:20:12.400 --> 0:20:16.760 about it, not only to the media, but also at 0:20:16.800 --> 0:20:21.760 military bases, at philosophy conferences, at the International Committee of the 0:20:21.720 --> 0:20:25.600 the Red Cross where they developed the Geneva Conventions just 0:20:25.760 --> 0:20:31.240 last month, a variety of different locations, ethics groups and 0:20:31.280 --> 0:20:34.720 the like as well too too answer that very question. 0:20:35.000 --> 0:20:38.560 So it would be really presumptuous at me to say, uh, 0:20:39.040 --> 0:20:42.320 this is perfectly safe, don't worry about it, or the 0:20:42.359 --> 0:20:44.840 sky is falling. The sky is falling, like some of 0:20:44.880 --> 0:20:48.879 my colleagues are doing as well too. Uh. What I 0:20:48.920 --> 0:20:51.359 do say is that if we are going to allow 0:20:51.400 --> 0:20:55.160 these robots to make these life or death situations decisions. 0:20:55.400 --> 0:20:57.280 If we're going to allow these robots to make these 0:20:57.280 --> 0:21:00.240 life or death decisions on their own to some to rate, 0:21:00.359 --> 0:21:01.840 I mean, they're not just going to go out and 0:21:01.840 --> 0:21:04.960 start looking at someone and say should I go that individual. 0:21:05.280 --> 0:21:07.320 They will be tasked by a human being in a 0:21:07.359 --> 0:21:11.520 mission context and then make that particular decision. But if 0:21:11.520 --> 0:21:13.879 we're going to do that, we need to understand what 0:21:13.920 --> 0:21:17.440 it is that we're doing, what if any bounds are appropriate, 0:21:17.880 --> 0:21:20.560 And from my perspective, if we're going to do that, 0:21:20.760 --> 0:21:23.359 they must adhere to the existing laws of war and 0:21:23.400 --> 0:21:28.879 the rules of engagement, as we as a society already do. 0:21:29.600 --> 0:21:32.199 And so that's what my work is about. I have 0:21:32.320 --> 0:21:34.720 often said that while I do believe that we can 0:21:34.760 --> 0:21:39.040 actually do better than human beings in these situations, that 0:21:39.440 --> 0:21:43.480 if it ended up that intelligent lethal robots were banned 0:21:43.480 --> 0:21:45.960 from the battlefield, I'm not adverse to that, although I 0:21:45.960 --> 0:21:48.159 still do believe that we can actually do better than 0:21:48.240 --> 0:21:54.080 human beings and making the right decisions regarding lethal application 0:21:54.119 --> 0:21:57.800 of force under certain circumstances. And it's really important to 0:21:57.840 --> 0:22:01.520 understand that I am not talking about placing a soldier 0:22:01.800 --> 0:22:05.840 with a robot. I'm talking about augmenting soldiers with robots 0:22:05.880 --> 0:22:09.800 and using them in highly specialized missions such as building 0:22:09.800 --> 0:22:14.680 clearing operations, counter sniper operations, things of that particular sort, 0:22:15.320 --> 0:22:21.200 which is quite different than saying, uh, here's a terminator, uh, 0:22:21.240 --> 0:22:25.080 and uh that's going to replace that particular soldier. Very 0:22:25.160 --> 0:22:29.360 narrow well defined situations where something called bounded morality applies. 0:22:29.680 --> 0:22:33.640 And that's what makes attractable, because human morality is extremely complex, 0:22:33.680 --> 0:22:37.840 involving multiple brain systems and other aspects of that. Machine 0:22:37.880 --> 0:22:40.240 learning deliberation, all sorts of things could be brought to 0:22:40.280 --> 0:22:44.640 bear in applying it in that context. But what's interesting 0:22:44.960 --> 0:22:49.760 is the lethal application of force is relatively low hanging 0:22:49.800 --> 0:22:55.320 fruit for a person dealing with computational morality for a 0:22:55.359 --> 0:23:00.040 couple of reasons. The primary one is that philosophers have 0:23:00.119 --> 0:23:02.480 been thinking about it for thousands of years, under what 0:23:02.600 --> 0:23:06.439