1 00:00:04,519 --> 00:00:12,319 Speaker 1: Technology with tech Stuff from works dot com. Hey there, 2 00:00:12,360 --> 00:00:15,680 Speaker 1: and welcome to tech Stuff. I'm your host, Jonathan Strickland. 3 00:00:15,720 --> 00:00:18,880 Speaker 1: I'm a senior writer for how stuff works dot com 4 00:00:19,040 --> 00:00:24,320 Speaker 1: focusing on all things technological, and recently I did an 5 00:00:24,320 --> 00:00:28,880 Speaker 1: episode about artificial intelligence and how Mark Zuckerberg and Elon 6 00:00:28,960 --> 00:00:33,760 Speaker 1: Musk had kind of a public disagreement about the direction 7 00:00:33,800 --> 00:00:36,239 Speaker 1: of AI, and how other people have weighed in. Some 8 00:00:36,280 --> 00:00:40,240 Speaker 1: people have said that perhaps Musk and Zuckerberg are arguing 9 00:00:40,280 --> 00:00:44,360 Speaker 1: about something that isn't really relevant right now, and that 10 00:00:44,440 --> 00:00:47,000 Speaker 1: there are in fact other elements of artificial intelligence that 11 00:00:47,040 --> 00:00:50,000 Speaker 1: we should be focusing on instead of whether or not 12 00:00:50,800 --> 00:00:53,320 Speaker 1: it is certain to make our lives better or worse 13 00:00:53,560 --> 00:00:56,400 Speaker 1: or rule over us. But it got me to thinking 14 00:00:56,440 --> 00:00:59,960 Speaker 1: about a related topic, and I touched on it all 15 00:01:00,040 --> 00:01:03,640 Speaker 1: little bit in that episode, and that was all about 16 00:01:04,560 --> 00:01:10,840 Speaker 1: how do you tell when a a an entity that 17 00:01:10,959 --> 00:01:16,080 Speaker 1: is communicating with you is in fact a person or 18 00:01:16,160 --> 00:01:19,679 Speaker 1: it is a computer program that is mimicking a person. 19 00:01:20,200 --> 00:01:22,840 Speaker 1: So we're going to look at that, And honestly, I 20 00:01:22,880 --> 00:01:25,360 Speaker 1: was inspired a lot also by the fact that we've 21 00:01:25,360 --> 00:01:28,000 Speaker 1: got a new Blade Runner movie coming out. It's Blade 22 00:01:28,080 --> 00:01:31,560 Speaker 1: Runner two thousand, forty nine, which has no connection to 23 00:01:31,640 --> 00:01:34,399 Speaker 1: this show. By the way, they are not sponsoring us. 24 00:01:35,000 --> 00:01:37,640 Speaker 1: I'm pretty sure no one connected to Blade Runner two 25 00:01:37,640 --> 00:01:41,280 Speaker 1: thousand forty nine even is aware that I exist. But 26 00:01:41,800 --> 00:01:44,240 Speaker 1: I'm a fan of the original Blade Runner film and 27 00:01:44,240 --> 00:01:48,480 Speaker 1: I'm looking forward to seeing what happens in Blade Runner 28 00:01:48,520 --> 00:01:52,320 Speaker 1: two thousand forty nine. I'm a little hesitant because it 29 00:01:52,360 --> 00:01:56,360 Speaker 1: depends upon which interpretation of the original film they decided 30 00:01:56,400 --> 00:01:59,560 Speaker 1: to ultimately go with. If they went with the director's vision, 31 00:02:00,320 --> 00:02:02,680 Speaker 1: might not want to see two thousand forty nine, but 32 00:02:03,320 --> 00:02:07,440 Speaker 1: I wanted to kind of talk about the difference between 33 00:02:07,440 --> 00:02:10,040 Speaker 1: communicating with a person and a synthetic being. Now in 34 00:02:10,080 --> 00:02:13,959 Speaker 1: Blade Runner, the synthetic beings are called replicants, and they 35 00:02:14,000 --> 00:02:17,760 Speaker 1: are not exactly robots. They're often referred to as androids, 36 00:02:18,240 --> 00:02:21,280 Speaker 1: but I don't really think that's terribly accurate either. They're 37 00:02:21,280 --> 00:02:28,520 Speaker 1: more like genetically engineered human simulations. Like they're not fully human. Uh. 38 00:02:28,560 --> 00:02:33,960 Speaker 1: They have other elements that either augmented abilities and intelligence, 39 00:02:34,560 --> 00:02:37,440 Speaker 1: but a lower lifespan that sort of stuffy. They tend 40 00:02:37,480 --> 00:02:41,519 Speaker 1: to be born in the adult stage of their lives 41 00:02:41,520 --> 00:02:45,440 Speaker 1: and implanted with false memories, but they're meant to do 42 00:02:45,560 --> 00:02:48,120 Speaker 1: jobs that humans can't or won't do, and they do 43 00:02:48,240 --> 00:02:50,720 Speaker 1: have a tendency to resent their lot in life, seeing 44 00:02:50,760 --> 00:02:53,880 Speaker 1: as how in the original film they did just have 45 00:02:54,000 --> 00:02:56,200 Speaker 1: that built an expiration data just a few years. They 46 00:02:56,240 --> 00:02:58,280 Speaker 1: don't live for a few years, and then they would 47 00:02:58,720 --> 00:03:03,440 Speaker 1: their bodies would break down and blade Runner the story 48 00:03:03,520 --> 00:03:07,760 Speaker 1: follows an investigator who is seeking out specific replicants that 49 00:03:07,800 --> 00:03:10,640 Speaker 1: are on the run in order to quote unquote retire 50 00:03:10,760 --> 00:03:14,640 Speaker 1: them with extreme prejudice. So this is all set up 51 00:03:14,639 --> 00:03:17,959 Speaker 1: at the beginning of the movie. Now, one thing those 52 00:03:18,000 --> 00:03:23,280 Speaker 1: investigators or blade Runners do is ask questions of suspects, 53 00:03:23,320 --> 00:03:25,600 Speaker 1: suspected replicants. You know, they find someone they think that 54 00:03:25,720 --> 00:03:29,000 Speaker 1: might be a replicant, and then they interview that person 55 00:03:29,080 --> 00:03:32,560 Speaker 1: and they look for signs that that is not actually 56 00:03:32,600 --> 00:03:36,880 Speaker 1: a real human being, because replicants are not exactly human. 57 00:03:36,920 --> 00:03:40,720 Speaker 1: They're human like, but they do not process emotions the 58 00:03:40,840 --> 00:03:44,080 Speaker 1: same way that humans do. So blade Runners can look 59 00:03:44,120 --> 00:03:47,880 Speaker 1: for indications that the suspect is actually a replicant and 60 00:03:47,920 --> 00:03:50,840 Speaker 1: they use what is called the void comp test in 61 00:03:50,880 --> 00:03:54,040 Speaker 1: the movie. This is a test that includes the hypothetical 62 00:03:54,080 --> 00:03:57,440 Speaker 1: situation you're in a desert walking along in the sand 63 00:03:57,480 --> 00:03:59,640 Speaker 1: when all of a sudden, you look down and see 64 00:03:59,680 --> 00:04:02,200 Speaker 1: it towards it. This you reach down and flip the 65 00:04:02,200 --> 00:04:05,440 Speaker 1: tortoise on its back. The tortoise lays on its back. 66 00:04:05,520 --> 00:04:08,480 Speaker 1: It's belly baking in the hot sun, beating its legs, 67 00:04:08,480 --> 00:04:12,120 Speaker 1: trying to turn itself over, but it can't, not without 68 00:04:12,200 --> 00:04:16,800 Speaker 1: your help. But you're not helping. Why is that? Now? 69 00:04:16,800 --> 00:04:19,720 Speaker 1: I kind of paraphrase that scene because it actually happens 70 00:04:19,720 --> 00:04:23,560 Speaker 1: as dialogue between two characters. But that was the attempt 71 00:04:23,760 --> 00:04:27,120 Speaker 1: of an interrogator to figure out whether or not the 72 00:04:27,160 --> 00:04:29,960 Speaker 1: person they were talking to is actually a human being, 73 00:04:30,400 --> 00:04:34,800 Speaker 1: because the emotional responses would indicate whether or not it 74 00:04:34,880 --> 00:04:38,039 Speaker 1: was a human response, or if there was a lack 75 00:04:38,080 --> 00:04:40,520 Speaker 1: of that, that it was perhaps a replicant. Now, that's 76 00:04:40,520 --> 00:04:42,560 Speaker 1: all science fiction, but in the real world there are 77 00:04:42,680 --> 00:04:45,960 Speaker 1: times when we encounter bots or AI constructs and we 78 00:04:46,040 --> 00:04:49,040 Speaker 1: might not know at first, at least they were not 79 00:04:49,080 --> 00:04:52,880 Speaker 1: communicating with a real life person. In fact, the Interactive 80 00:04:52,960 --> 00:04:57,200 Speaker 1: Advertising Bureau reported in two thousand fourteen that thirty six 81 00:04:57,279 --> 00:05:00,880 Speaker 1: per cent of all web traffic is general to buy bots, 82 00:05:01,320 --> 00:05:05,080 Speaker 1: not people, and the security firm Imperva reported in early 83 00:05:05,120 --> 00:05:09,239 Speaker 1: two thousand seventeen that today that figure is now closer 84 00:05:09,279 --> 00:05:14,600 Speaker 1: to fifty, which means that right now there's more traffic 85 00:05:14,640 --> 00:05:19,680 Speaker 1: on the web being generated from bots than actual human beings, 86 00:05:20,240 --> 00:05:23,320 Speaker 1: And that's not exactly great. Much of the web depends 87 00:05:23,400 --> 00:05:27,120 Speaker 1: upon advertising for monetization. But how do you figure out 88 00:05:27,200 --> 00:05:31,080 Speaker 1: what the value of traffic to your website is when 89 00:05:31,120 --> 00:05:33,240 Speaker 1: you know there's a good chance that more than half 90 00:05:33,440 --> 00:05:36,960 Speaker 1: of all those page views were generated from algorithms, not 91 00:05:37,040 --> 00:05:40,479 Speaker 1: from human beings. Now, much of the bot traffic isn't 92 00:05:40,480 --> 00:05:43,520 Speaker 1: meant to be outright malicious. There might be bots that 93 00:05:43,600 --> 00:05:47,760 Speaker 1: are essentially trying to to scour the Internet for data 94 00:05:47,839 --> 00:05:50,560 Speaker 1: for nefarious purposes, but a lot of them are just 95 00:05:50,680 --> 00:05:57,680 Speaker 1: they're gathering information for, you know, completely innocent purposes. Really, 96 00:05:58,240 --> 00:06:01,760 Speaker 1: gathering information on its own is not necessarily a bad thing. 97 00:06:02,080 --> 00:06:04,640 Speaker 1: It's how we use the information that makes it good 98 00:06:04,720 --> 00:06:09,039 Speaker 1: or bad. It's kind of paraphrasing Shakespeare there. But there 99 00:06:09,040 --> 00:06:12,680 Speaker 1: are the various bots on social platforms and websites that 100 00:06:12,800 --> 00:06:16,160 Speaker 1: also interact with people, and some of them are again benign. 101 00:06:16,200 --> 00:06:18,960 Speaker 1: They're meant to be helpful, such as bots that can 102 00:06:19,000 --> 00:06:23,160 Speaker 1: answer basic customer service questions for companies. You've probably encountered 103 00:06:23,160 --> 00:06:25,920 Speaker 1: one of these where you were looking for some information 104 00:06:25,920 --> 00:06:29,320 Speaker 1: about a particular product or service and then a little 105 00:06:29,360 --> 00:06:31,880 Speaker 1: chat window pops up, and you get the feeling that 106 00:06:31,920 --> 00:06:35,400 Speaker 1: the entity you're talking to is not exactly another human 107 00:06:35,400 --> 00:06:37,320 Speaker 1: being on the other end. It may just be a bot. 108 00:06:38,360 --> 00:06:42,120 Speaker 1: Sometimes that's fine. Sometimes it's more frustrating than helpful, because 109 00:06:42,440 --> 00:06:44,240 Speaker 1: you find that you have to word things in a 110 00:06:44,320 --> 00:06:47,279 Speaker 1: very particular way for the bot to comprehend what you mean, 111 00:06:47,480 --> 00:06:49,719 Speaker 1: whereas a human would probably pick it up much faster. 112 00:06:50,480 --> 00:06:53,280 Speaker 1: But you get the idea of why that was employed, right. 113 00:06:53,360 --> 00:06:56,880 Speaker 1: That was meant to make things a little more smooth 114 00:06:57,080 --> 00:07:00,200 Speaker 1: and to remove the necessity of putting a human being 115 00:07:00,240 --> 00:07:03,240 Speaker 1: in charge of that at all hours of the day. 116 00:07:03,320 --> 00:07:06,359 Speaker 1: You can also find these sorts of automated services on 117 00:07:06,440 --> 00:07:09,920 Speaker 1: phone lines, including bots that call you, which is always fun. 118 00:07:10,400 --> 00:07:12,480 Speaker 1: There's nothing like having a conversation with a bot for 119 00:07:12,560 --> 00:07:14,880 Speaker 1: half a minute before you figure out something fishy is 120 00:07:14,920 --> 00:07:18,200 Speaker 1: going on. But other bots are meant to serve the 121 00:07:18,240 --> 00:07:23,640 Speaker 1: purposes of some third party, sometimes with malicious intent, such 122 00:07:23,680 --> 00:07:26,120 Speaker 1: as convincing you to click on a link that leads 123 00:07:26,120 --> 00:07:30,480 Speaker 1: to malware, and that's where we really run into obvious problems. 124 00:07:30,520 --> 00:07:34,080 Speaker 1: Some of the benign ones can run into problems too often. 125 00:07:34,120 --> 00:07:38,640 Speaker 1: There are unintended consequences if you're scouring the web for data. 126 00:07:38,760 --> 00:07:42,360 Speaker 1: Data is valuable and sometimes people will want to get 127 00:07:42,440 --> 00:07:45,840 Speaker 1: hold of it for bad reasons, even if the initial 128 00:07:46,000 --> 00:07:50,080 Speaker 1: approach wasn't to do anything nefarious. Now, some hackers have 129 00:07:50,320 --> 00:07:53,160 Speaker 1: used bots to flood a platform with complaints in an 130 00:07:53,160 --> 00:07:55,760 Speaker 1: effort to silence people that the hackers do not like. 131 00:07:56,200 --> 00:08:00,240 Speaker 1: So let's say there's this jerk face hacker who thinks 132 00:08:00,280 --> 00:08:03,640 Speaker 1: a Facebook page devoted to promoting women in STEM education 133 00:08:03,680 --> 00:08:07,480 Speaker 1: and careers is dumb. So this jerk face then creates 134 00:08:07,600 --> 00:08:11,840 Speaker 1: or more likely purchases bots to flood Facebook with complaint 135 00:08:11,920 --> 00:08:15,960 Speaker 1: reports about that specific page in an attempt to get 136 00:08:16,000 --> 00:08:19,280 Speaker 1: Facebook to shut the page down. Now that's a pretty 137 00:08:19,280 --> 00:08:21,520 Speaker 1: lousy thing to do. And to be clear, some of 138 00:08:21,520 --> 00:08:23,960 Speaker 1: the jerk faces are aiming at pages that the average 139 00:08:23,960 --> 00:08:26,000 Speaker 1: person would say is a bad one. It doesn't have 140 00:08:26,120 --> 00:08:29,600 Speaker 1: to be something that like I feel strongly about and 141 00:08:29,600 --> 00:08:32,720 Speaker 1: and in favor for. I think STEM education and careers 142 00:08:32,760 --> 00:08:35,240 Speaker 1: for women is amazing, and I would be very upset 143 00:08:35,280 --> 00:08:37,120 Speaker 1: to hear about a page that was shut down because 144 00:08:37,120 --> 00:08:39,600 Speaker 1: of one of these attacks. On the other hand, let's 145 00:08:39,600 --> 00:08:42,000 Speaker 1: say that there was a page that was promoting something 146 00:08:42,040 --> 00:08:44,480 Speaker 1: I really do not like. Maybe it was a page 147 00:08:44,480 --> 00:08:48,679 Speaker 1: that was promoting, uh, you know, racial discrimination. I would 148 00:08:48,760 --> 00:08:51,600 Speaker 1: think that was terrible. If someone else were to take 149 00:08:51,679 --> 00:08:54,880 Speaker 1: bots and direct them to that page in order to 150 00:08:54,920 --> 00:08:56,960 Speaker 1: shut it down, I would also think that that's not 151 00:08:57,080 --> 00:09:01,760 Speaker 1: so great. I don't think that a page about racial 152 00:09:01,800 --> 00:09:05,960 Speaker 1: discrimination should be promoted or exist on Facebook. I don't 153 00:09:06,000 --> 00:09:08,440 Speaker 1: think that's appropriate. But at the same time, I don't 154 00:09:08,440 --> 00:09:14,240 Speaker 1: think it's appropriate to use automated systems to bring that down. 155 00:09:14,559 --> 00:09:18,960 Speaker 1: I would rather see an actual ground swell of human 156 00:09:19,120 --> 00:09:24,360 Speaker 1: support for that, not to you know, boost it with 157 00:09:24,400 --> 00:09:27,839 Speaker 1: a bunch of automated scripts. I don't want to give 158 00:09:27,840 --> 00:09:31,200 Speaker 1: the indication that the only people who ever use bots 159 00:09:31,240 --> 00:09:35,120 Speaker 1: are those who want to silence vulnerable or underrepresented populations. 160 00:09:35,160 --> 00:09:37,520 Speaker 1: There are some who use them to attempt to silence 161 00:09:37,600 --> 00:09:41,120 Speaker 1: voices of hate. In either case, it's dirty pool. I 162 00:09:41,160 --> 00:09:45,760 Speaker 1: don't think it's really a legitimate strategy. Uh, it ends 163 00:09:45,840 --> 00:09:48,840 Speaker 1: up hurting everyone in the long run to use bots 164 00:09:48,840 --> 00:09:52,320 Speaker 1: in that specific way. Butts in general, I'm not against. 165 00:09:53,080 --> 00:09:56,400 Speaker 1: I do think there are times when they are incredibly useful, 166 00:09:56,400 --> 00:09:59,960 Speaker 1: but to use them specifically to fool people into think 167 00:10:00,040 --> 00:10:03,200 Speaker 1: ging their actual human beings in order to achieve an 168 00:10:03,280 --> 00:10:07,600 Speaker 1: ulterior motive that sets me on edge that I can't 169 00:10:07,640 --> 00:10:10,200 Speaker 1: really see an upside to that. I can definitely see 170 00:10:10,240 --> 00:10:13,600 Speaker 1: it from the side of customer service or answering general questions, 171 00:10:13,920 --> 00:10:18,080 Speaker 1: maybe even just trying to funnel out people who have 172 00:10:18,200 --> 00:10:21,640 Speaker 1: a very simple issue to resolve versus those who need 173 00:10:22,000 --> 00:10:25,880 Speaker 1: more attention, whereas you know those people would get directed 174 00:10:25,920 --> 00:10:27,800 Speaker 1: towards a pathway that would lead to speaking to an 175 00:10:27,800 --> 00:10:31,760 Speaker 1: actual human being. I get it from that perspective. Now, 176 00:10:31,760 --> 00:10:34,679 Speaker 1: in a recent episode, I explained in brief what the 177 00:10:34,679 --> 00:10:37,720 Speaker 1: Turing test was, or at least how we interpret it. 178 00:10:37,840 --> 00:10:40,520 Speaker 1: The Turing test is sort of the inspiration for the 179 00:10:40,600 --> 00:10:44,400 Speaker 1: Void comp test and Blade Runner. Alan Turing, one of 180 00:10:44,400 --> 00:10:47,960 Speaker 1: the fathers of computer science, proposed the test back in 181 00:10:48,080 --> 00:10:51,560 Speaker 1: nineteen fifty and in the actual thought experiment that he 182 00:10:51,600 --> 00:10:55,040 Speaker 1: was proposing, it was a variation on a parlor game 183 00:10:55,080 --> 00:10:58,400 Speaker 1: called the imitation game. Now, the imitation game is one 184 00:10:58,440 --> 00:11:01,600 Speaker 1: where you have an interrogator that's player, and the player 185 00:11:01,720 --> 00:11:05,199 Speaker 1: is presented with two subjects, neither of whom the interrogator 186 00:11:05,240 --> 00:11:08,600 Speaker 1: can see or talk to directly. One of the two 187 00:11:08,679 --> 00:11:13,240 Speaker 1: subjects is a woman, the other is a man. Both 188 00:11:13,400 --> 00:11:15,960 Speaker 1: of the subjects can communicate with the interrogator in a 189 00:11:16,000 --> 00:11:18,600 Speaker 1: way that does not require face to face contact or 190 00:11:18,720 --> 00:11:21,520 Speaker 1: voice or anything like that. Typically it would be through 191 00:11:21,600 --> 00:11:25,160 Speaker 1: something like typewritten letters, because that would help disguise handwriting 192 00:11:25,200 --> 00:11:28,160 Speaker 1: as well. And the two subjects have the same task. 193 00:11:28,559 --> 00:11:32,400 Speaker 1: They have to convince the interrogator that they are female. 194 00:11:32,720 --> 00:11:34,920 Speaker 1: So the woman will be telling the truth, the man 195 00:11:35,080 --> 00:11:38,520 Speaker 1: will be lying, and it's the interrogator's job to figure 196 00:11:38,559 --> 00:11:41,360 Speaker 1: out who is imitating a woman and who actually is 197 00:11:41,440 --> 00:11:46,480 Speaker 1: a woman. Touring then suggested taking this game a step 198 00:11:46,520 --> 00:11:51,040 Speaker 1: further by replacing the male subject in this thought experiment 199 00:11:51,240 --> 00:11:54,920 Speaker 1: with a computer. The computer would also attempt to convince 200 00:11:54,920 --> 00:11:58,480 Speaker 1: the interrogator that the computer was in fact a woman. 201 00:11:59,520 --> 00:12:02,520 Speaker 1: Now would the interrogator be able to detect the computer's 202 00:12:02,640 --> 00:12:06,760 Speaker 1: ruse if not? Touring suggested that this would indicate some 203 00:12:06,840 --> 00:12:10,880 Speaker 1: form of intelligence, though not necessarily human intelligence. But you 204 00:12:10,920 --> 00:12:16,720 Speaker 1: could say the machine is capable of fooling a human being, 205 00:12:16,920 --> 00:12:20,520 Speaker 1: of of practicing deception, which I think most of us 206 00:12:20,559 --> 00:12:24,280 Speaker 1: would argue. The ability to practice deception does indicate at 207 00:12:24,360 --> 00:12:28,520 Speaker 1: least some form of intelligence. Maybe not the type of 208 00:12:28,559 --> 00:12:31,320 Speaker 1: intelligence that's gonna go out and teach a class on 209 00:12:31,440 --> 00:12:35,439 Speaker 1: quantum mechanics, but the type of intelligence that does understand 210 00:12:36,240 --> 00:12:39,719 Speaker 1: the concept of manipulation at least or at least is 211 00:12:39,760 --> 00:12:43,600 Speaker 1: able to employ the concept of manipulation, if not understand 212 00:12:43,679 --> 00:12:47,640 Speaker 1: it from a truly cognitive point of view. Now, the 213 00:12:47,679 --> 00:12:50,959 Speaker 1: other variations and refinements to the Touring test followed after 214 00:12:51,080 --> 00:12:55,000 Speaker 1: Touring's death in nineteen fifty four, and Touring's life was 215 00:12:55,200 --> 00:12:58,720 Speaker 1: very tragic. We've done an episode on Alan Touring, so 216 00:12:58,720 --> 00:13:00,280 Speaker 1: if you want to go back and find that in 217 00:13:00,320 --> 00:13:03,880 Speaker 1: our archives, you can learn all about his his death 218 00:13:03,960 --> 00:13:07,320 Speaker 1: and why some people rule it a suicide. I think 219 00:13:07,320 --> 00:13:10,359 Speaker 1: most people do, and some people say it was accidental. 220 00:13:11,520 --> 00:13:15,880 Speaker 1: But it is an interesting and tragic tale. Today, the 221 00:13:15,920 --> 00:13:18,480 Speaker 1: general interpretation of the Turing test is that if a 222 00:13:18,520 --> 00:13:22,040 Speaker 1: certain threshold is met, such as a greater than thirty 223 00:13:22,120 --> 00:13:25,959 Speaker 1: percent success rate of a computer convincing interrogators that's actually 224 00:13:25,960 --> 00:13:30,120 Speaker 1: a human, it has passed the Turing test. So, in 225 00:13:30,200 --> 00:13:33,160 Speaker 1: other words, if you're an interrogator and you've got a 226 00:13:33,240 --> 00:13:35,840 Speaker 1: computer terminal in front of you, and you're typing messages 227 00:13:36,400 --> 00:13:40,160 Speaker 1: and the the response is coming back to you. And 228 00:13:40,320 --> 00:13:42,800 Speaker 1: if more than thirty percent of the time you cannot 229 00:13:42,880 --> 00:13:45,280 Speaker 1: tell if that actually is a computer or a person, 230 00:13:46,080 --> 00:13:49,320 Speaker 1: maybe you misidentified as a person more than thirty percent 231 00:13:49,360 --> 00:13:52,320 Speaker 1: of the time, and it's actually the computer. That computer 232 00:13:52,400 --> 00:13:54,360 Speaker 1: is said to pass the Turing test, and that it 233 00:13:54,480 --> 00:13:57,160 Speaker 1: is capable of fooling you into thinking it's an actual 234 00:13:57,240 --> 00:13:59,640 Speaker 1: human being. Now, there was a case in two thousand 235 00:13:59,679 --> 00:14:03,080 Speaker 1: and four in which a chat bought called Eugene seemed 236 00:14:03,240 --> 00:14:07,400 Speaker 1: to accomplish this. Eugene's persona was that of a thirteen 237 00:14:07,480 --> 00:14:11,560 Speaker 1: year old Ukrainian boy. Critics pointed out that Eugene's limitations 238 00:14:11,640 --> 00:14:15,160 Speaker 1: as a non native English speaker with a limited knowledge 239 00:14:15,160 --> 00:14:17,520 Speaker 1: of the world due to his age and the fact 240 00:14:17,520 --> 00:14:20,400 Speaker 1: that he was from the Ukraine, meant that people were 241 00:14:20,480 --> 00:14:24,120 Speaker 1: lowering their expectations on his performance when they were chatting 242 00:14:24,120 --> 00:14:27,120 Speaker 1: with him over a computer. In other words, critics were 243 00:14:27,120 --> 00:14:33,600 Speaker 1: saying that Eugene was gaming the system by making people think, oh, well, 244 00:14:34,000 --> 00:14:38,000 Speaker 1: non native English speaker, so if the responses come back 245 00:14:38,040 --> 00:14:43,080 Speaker 1: a little weird, that explains that. And being young means 246 00:14:43,120 --> 00:14:46,440 Speaker 1: that they this kid doesn't have that much knowledge about 247 00:14:46,880 --> 00:14:49,440 Speaker 1: a lot of things in the world, pop culture, politics, 248 00:14:49,680 --> 00:14:53,720 Speaker 1: lots of stuff, so your expectations are set low, and 249 00:14:53,760 --> 00:14:56,600 Speaker 1: then you just think, all right, well, are the messages 250 00:14:56,640 --> 00:14:59,320 Speaker 1: I'm getting Are those in line with what I would 251 00:14:59,360 --> 00:15:02,840 Speaker 1: expect a thirteen year old non native English speaker to 252 00:15:02,960 --> 00:15:06,240 Speaker 1: say to me, or do they stand out as being artificial? 253 00:15:07,360 --> 00:15:10,040 Speaker 1: And a lot of this ends up being deflection as well, 254 00:15:10,160 --> 00:15:14,720 Speaker 1: where if you ask somebody a question and the computer 255 00:15:14,800 --> 00:15:17,400 Speaker 1: program doesn't have a way of responding, it will try 256 00:15:17,400 --> 00:15:21,480 Speaker 1: to deflect the question so that it doesn't indicate that 257 00:15:21,520 --> 00:15:27,360 Speaker 1: and in fact is a computer program. Well, Eugene managed 258 00:15:27,400 --> 00:15:30,440 Speaker 1: to to fool a lot of people, But again the 259 00:15:30,480 --> 00:15:35,880 Speaker 1: critics were saying, well, Eugene was kind of an outlier 260 00:15:35,920 --> 00:15:39,160 Speaker 1: in the sense that you didn't really think of Eugene 261 00:15:39,160 --> 00:15:44,520 Speaker 1: as being a native speaker with a lifetime of experience 262 00:15:45,080 --> 00:15:48,560 Speaker 1: where you could really quiz the the entity and find out, Okay, 263 00:15:48,640 --> 00:15:51,760 Speaker 1: is this actually a person or is it a computer program. 264 00:15:51,960 --> 00:15:53,680 Speaker 1: It's sort of beside the point. I'm not here to 265 00:15:53,760 --> 00:15:57,920 Speaker 1: argue about whether or not machines possess intelligence if they 266 00:15:57,920 --> 00:16:01,480 Speaker 1: passed the Turing test, because I did that recently already. Instead, 267 00:16:01,560 --> 00:16:04,080 Speaker 1: let's focus on the flip side of the scenario, we're 268 00:16:04,160 --> 00:16:08,160 Speaker 1: human at least I'm assuming you're a human. You might 269 00:16:08,160 --> 00:16:11,520 Speaker 1: be a bot who subscribed to text stuff. Apparently of 270 00:16:11,520 --> 00:16:14,280 Speaker 1: you out there are in that case. Thanks. I hope 271 00:16:14,320 --> 00:16:17,240 Speaker 1: you like the show. But this is all for the 272 00:16:17,280 --> 00:16:19,680 Speaker 1: humans here, this bit out here, so you bots out 273 00:16:19,720 --> 00:16:23,080 Speaker 1: there can take a break. How can we humans tell 274 00:16:23,120 --> 00:16:25,320 Speaker 1: if we're dealing with an actual person or if it 275 00:16:25,400 --> 00:16:29,240 Speaker 1: is a bot. Well, one of the ways that we 276 00:16:29,360 --> 00:16:34,440 Speaker 1: have created a means of separating bots from humans is capture. 277 00:16:35,320 --> 00:16:38,640 Speaker 1: Capture is an acronym that stands for a completely automated 278 00:16:38,720 --> 00:16:43,200 Speaker 1: public touring test to tell computers and humans apart. That 279 00:16:43,240 --> 00:16:44,960 Speaker 1: pretty much sums it up when you break it down. 280 00:16:45,120 --> 00:16:48,560 Speaker 1: It's completely automated, meaning there's no human oversight necessary for 281 00:16:48,600 --> 00:16:53,120 Speaker 1: any given implementation of the technology. It's public. It's pretty 282 00:16:53,120 --> 00:16:55,880 Speaker 1: self explanatory. It's a test that's out there in the public. 283 00:16:56,040 --> 00:16:58,080 Speaker 1: I guess I explained it even though it wasn't necessary. 284 00:16:58,560 --> 00:17:01,160 Speaker 1: That's my bad, y'all. Now, it's said to be a 285 00:17:01,160 --> 00:17:04,560 Speaker 1: turing test because it's meant to detect human versus automated 286 00:17:04,640 --> 00:17:08,480 Speaker 1: agents operating on a given web page. We talked about 287 00:17:08,600 --> 00:17:11,320 Speaker 1: the touring test just now. But h so we're not 288 00:17:11,359 --> 00:17:13,520 Speaker 1: gonna go over that again. But you know, again, it's 289 00:17:13,560 --> 00:17:17,560 Speaker 1: just just this indicator. Is there something there that implicates 290 00:17:17,680 --> 00:17:20,200 Speaker 1: this as being a computer agent not a human being? 291 00:17:20,640 --> 00:17:22,680 Speaker 1: And if it is in fact a computer agent, then 292 00:17:22,720 --> 00:17:24,960 Speaker 1: you have a gate up saying all right, you don't 293 00:17:25,000 --> 00:17:28,280 Speaker 1: get to participate in this because it's not meant for you. 294 00:17:28,520 --> 00:17:31,680 Speaker 1: When you have of your web traffic out there generated 295 00:17:31,680 --> 00:17:35,359 Speaker 1: by bots and you're trying to collect meaningful data about 296 00:17:35,400 --> 00:17:38,879 Speaker 1: real human being users, you have to have a way 297 00:17:39,359 --> 00:17:41,960 Speaker 1: to separate the two. Right. So, if I'm a web 298 00:17:42,000 --> 00:17:45,400 Speaker 1: administrator and let's say that I've got let's just say 299 00:17:45,400 --> 00:17:50,200 Speaker 1: that I'm running a sweepstakes, have created an online entry form. 300 00:17:50,240 --> 00:17:53,080 Speaker 1: I don't want someone flooding my sweepstakes with bots in 301 00:17:53,119 --> 00:17:56,040 Speaker 1: an effort to try and game the system and win 302 00:17:56,359 --> 00:17:59,800 Speaker 1: by submitting more entries than anybody else. I want to 303 00:17:59,800 --> 00:18:02,400 Speaker 1: be able to control that. So I want to have 304 00:18:02,600 --> 00:18:04,800 Speaker 1: some sort of element on there that can weed out 305 00:18:05,400 --> 00:18:11,040 Speaker 1: the automated agents out there versus the actual human beings. Now, 306 00:18:11,080 --> 00:18:13,879 Speaker 1: that last bit and capture to tell computers and humans 307 00:18:13,880 --> 00:18:16,119 Speaker 1: apart is the key to all of this. Capture is 308 00:18:16,160 --> 00:18:18,720 Speaker 1: a Guardian right, Like I was just saying, it's meant 309 00:18:18,760 --> 00:18:20,680 Speaker 1: to keep people from just writing a script to fill 310 00:18:20,680 --> 00:18:24,080 Speaker 1: out a form or make a comment on forums, really 311 00:18:24,600 --> 00:18:27,560 Speaker 1: complete any interaction on the web in an automated way. 312 00:18:28,160 --> 00:18:33,480 Speaker 1: As someone who creates content online and I get lots 313 00:18:33,480 --> 00:18:36,480 Speaker 1: of comments on various platforms, I don't want a whole 314 00:18:36,520 --> 00:18:42,960 Speaker 1: bunch of automated gobbledygook showing up under my various podcasts 315 00:18:42,960 --> 00:18:47,120 Speaker 1: and videos because then I can't tell where the actual 316 00:18:47,600 --> 00:18:51,240 Speaker 1: signal is. All I'm seeing is noise. So you want 317 00:18:51,280 --> 00:18:54,880 Speaker 1: to have some way of controlling that, and you might 318 00:18:55,080 --> 00:18:57,439 Speaker 1: use it to limit spam in the message board, or 319 00:18:57,480 --> 00:19:00,520 Speaker 1: to stop people from abusing the format of an online mistakes, 320 00:19:00,640 --> 00:19:04,399 Speaker 1: or or again to stop people from harassing others on 321 00:19:04,480 --> 00:19:07,760 Speaker 1: social platforms. Now, the necessity for cap chub is due 322 00:19:07,840 --> 00:19:10,800 Speaker 1: to a fundamental flaw of the Internet, and that flaw 323 00:19:10,960 --> 00:19:14,600 Speaker 1: is this, it doesn't take very many people to make 324 00:19:14,760 --> 00:19:18,960 Speaker 1: using the Internet a total drag. You don't want some 325 00:19:19,160 --> 00:19:21,840 Speaker 1: jerk face to use a script to create thousands of 326 00:19:21,880 --> 00:19:25,200 Speaker 1: email addresses from a web based email provider and then 327 00:19:25,359 --> 00:19:29,359 Speaker 1: use those email addresses for spam purposes or for someone 328 00:19:29,400 --> 00:19:31,560 Speaker 1: to gain the system. In other ways a single person 329 00:19:31,600 --> 00:19:35,840 Speaker 1: has the potential to impact lots of other people. So 330 00:19:35,960 --> 00:19:39,320 Speaker 1: everything's out of balance, and the force demands a Jedi 331 00:19:39,400 --> 00:19:44,080 Speaker 1: to right the wrongs or something. Now, the ideal application 332 00:19:44,119 --> 00:19:48,760 Speaker 1: of capture is some sort of test that is very 333 00:19:48,840 --> 00:19:53,560 Speaker 1: easy for humans to complete, but very difficult for computers 334 00:19:53,640 --> 00:19:56,639 Speaker 1: to complete. And that requires some creative thinking. So what 335 00:19:56,720 --> 00:19:59,119 Speaker 1: are some things that people are really good at but 336 00:19:59,200 --> 00:20:03,800 Speaker 1: computers are aren't so great at. Over time this changes. 337 00:20:04,160 --> 00:20:07,840 Speaker 1: Computer programmers get better at designing software that allows computers 338 00:20:07,840 --> 00:20:11,359 Speaker 1: to simulate more of what humans can do. And that's 339 00:20:11,400 --> 00:20:14,320 Speaker 1: not a bad thing necessarily because it pushes our development 340 00:20:14,320 --> 00:20:18,880 Speaker 1: of artificial intelligence. But for the purposes of gate keeping, 341 00:20:19,040 --> 00:20:22,240 Speaker 1: it does make it more tricky. You've gotta figure out 342 00:20:22,240 --> 00:20:24,560 Speaker 1: a new way to be able to prevent people from 343 00:20:24,560 --> 00:20:28,840 Speaker 1: abusing the system. Now, the idea for capture came from 344 00:20:28,880 --> 00:20:32,040 Speaker 1: a couple of different teams. One team was at Alta Vista, 345 00:20:32,280 --> 00:20:34,320 Speaker 1: which started to work on ways to cut down on 346 00:20:34,320 --> 00:20:38,480 Speaker 1: online abuse way back in the Ulta Vista team was 347 00:20:38,480 --> 00:20:41,439 Speaker 1: trying to find a way to prevent bots or scripts 348 00:20:41,480 --> 00:20:44,000 Speaker 1: from adding active u r l s to the search 349 00:20:44,080 --> 00:20:49,000 Speaker 1: engine platform. Meanwhile, the other team was at Carnegie Mellon University, 350 00:20:49,160 --> 00:20:51,680 Speaker 1: And actually this happened a couple of years after Alta 351 00:20:51,800 --> 00:20:55,560 Speaker 1: Vista's work, and they included some researchers who were really 352 00:20:56,040 --> 00:20:58,360 Speaker 1: eager to try and find a solution to this problem, 353 00:20:58,359 --> 00:21:03,600 Speaker 1: and they included Louis on On, Manuel Bloom, Nicholas Hopper, 354 00:21:03,760 --> 00:21:07,520 Speaker 1: and John Langford. It was the Carnegie Melon team that 355 00:21:07,600 --> 00:21:11,119 Speaker 1: coined the term capture back in two thousand three, and 356 00:21:11,200 --> 00:21:14,320 Speaker 1: it worked pretty well. Humans could get a capture right 357 00:21:14,400 --> 00:21:17,480 Speaker 1: more often than not, and computers weren't nearly as good 358 00:21:17,520 --> 00:21:21,000 Speaker 1: at it, at least not at first. Now we'll talk 359 00:21:21,040 --> 00:21:23,400 Speaker 1: a lot about captures in just a minute and get 360 00:21:23,400 --> 00:21:26,399 Speaker 1: into some more elements about telling the difference between butts 361 00:21:26,400 --> 00:21:29,359 Speaker 1: and humans, but right now let's take a quick break 362 00:21:29,680 --> 00:21:41,280 Speaker 1: to thank our sponsor. So with early capture implementations, things 363 00:21:41,520 --> 00:21:45,879 Speaker 1: were pretty simple. The capture would take on a pretty 364 00:21:45,960 --> 00:21:49,119 Speaker 1: universal form. You'd have a little box and inside that 365 00:21:49,160 --> 00:21:51,840 Speaker 1: box you would see a couple of different words or 366 00:21:52,000 --> 00:21:57,200 Speaker 1: collections of letters or other characters, often distorted in some way, 367 00:21:57,320 --> 00:21:59,960 Speaker 1: and a little field beneath it telling you, hey, tie 368 00:22:00,080 --> 00:22:03,160 Speaker 1: been what you see here? And it was your job, 369 00:22:03,200 --> 00:22:05,480 Speaker 1: as a human being type person to type in the 370 00:22:05,520 --> 00:22:08,000 Speaker 1: correct characters, and that would allow you to gain access 371 00:22:08,000 --> 00:22:10,800 Speaker 1: to whatever it was that the capture was guarding. And 372 00:22:10,800 --> 00:22:13,040 Speaker 1: the thought was that computers just weren't as good at 373 00:22:13,080 --> 00:22:17,280 Speaker 1: recognizing those characters as humans are. That if you distort them, 374 00:22:17,480 --> 00:22:22,560 Speaker 1: then the character recognition software couldn't put piece that altogether. 375 00:22:22,760 --> 00:22:27,760 Speaker 1: The weird shapes would be too far outside the norm 376 00:22:28,160 --> 00:22:31,399 Speaker 1: for the computer model. So if you had a one, 377 00:22:31,800 --> 00:22:35,320 Speaker 1: but that number one, the numeral one, it was all 378 00:22:35,440 --> 00:22:39,159 Speaker 1: wavy and staticky or something like you were, uh, you 379 00:22:39,240 --> 00:22:41,800 Speaker 1: were breaking up the shape a bit by changing it. 380 00:22:42,240 --> 00:22:45,680 Speaker 1: Computers can't really see that and conceptualize that's a one, 381 00:22:45,760 --> 00:22:48,480 Speaker 1: or at least not in the early days, so it 382 00:22:48,480 --> 00:22:50,399 Speaker 1: would just look like a weird squiggle to them, and 383 00:22:50,400 --> 00:22:53,360 Speaker 1: they wouldn't be able to complete the capture. Whereas we 384 00:22:53,520 --> 00:22:56,000 Speaker 1: human being type people, we'd look and think, that's the 385 00:22:56,080 --> 00:22:58,520 Speaker 1: worst number one I've ever seen. Some kid must have 386 00:22:58,560 --> 00:23:01,240 Speaker 1: drawn that, but we understand it is, we recognize it, 387 00:23:01,280 --> 00:23:03,879 Speaker 1: so we would type that in. That was the basis 388 00:23:03,960 --> 00:23:07,720 Speaker 1: for capture. Create a test that's relatively easy for humans, 389 00:23:08,240 --> 00:23:12,639 Speaker 1: very difficult for computers. Now, not everyone was capable of 390 00:23:12,680 --> 00:23:17,440 Speaker 1: seeing these captures. Clearly, some people have visual impair impairment, 391 00:23:17,680 --> 00:23:20,320 Speaker 1: and so they need to have some other element to 392 00:23:20,480 --> 00:23:24,960 Speaker 1: captures in order to be able to access that same content. 393 00:23:25,280 --> 00:23:29,280 Speaker 1: So there are also audible captures, which is pretty important 394 00:23:29,400 --> 00:23:32,880 Speaker 1: option to get around those visual impairments that some people have. 395 00:23:33,640 --> 00:23:41,040 Speaker 1: And uh, you might get a distorted voice being repeating 396 00:23:41,080 --> 00:23:44,359 Speaker 1: out the same sort of letters and numbers that you 397 00:23:44,359 --> 00:23:47,679 Speaker 1: would encounter with a capture. There might also be some 398 00:23:47,760 --> 00:23:52,240 Speaker 1: background noise that would include some other elements that would 399 00:23:52,280 --> 00:23:56,280 Speaker 1: make it hard for a computer program to analyze the 400 00:23:56,280 --> 00:23:59,160 Speaker 1: audio and figure out what was being said, but hopefully 401 00:23:59,400 --> 00:24:02,840 Speaker 1: humans would be able to make it out. So again, 402 00:24:02,840 --> 00:24:05,439 Speaker 1: it was all about making it more challenging for a 403 00:24:05,480 --> 00:24:08,959 Speaker 1: computer while not making it too challenging for human beings. 404 00:24:09,080 --> 00:24:13,920 Speaker 1: And sometimes that works great, and sometimes that doesn't work 405 00:24:14,000 --> 00:24:16,920 Speaker 1: so great. There are plenty of examples of human beings 406 00:24:16,920 --> 00:24:19,800 Speaker 1: who could not get through a capture because the distortion 407 00:24:19,920 --> 00:24:22,879 Speaker 1: was so great that it made the made it almost 408 00:24:22,920 --> 00:24:26,399 Speaker 1: impossible to recognize what the actual capture was supposed to be. 409 00:24:27,920 --> 00:24:31,240 Speaker 1: But the first counter to capture wasn't an advance in 410 00:24:31,320 --> 00:24:35,080 Speaker 1: computational analysis of visual or audible data. You know, there 411 00:24:35,080 --> 00:24:37,040 Speaker 1: are a lot of tricks that people figured out later 412 00:24:37,119 --> 00:24:42,240 Speaker 1: down the line, to make these visual captures easier to analyze, 413 00:24:42,320 --> 00:24:45,439 Speaker 1: things like switching all the images gray scales so that 414 00:24:45,520 --> 00:24:48,800 Speaker 1: you take out the different color gradations that could fool 415 00:24:48,840 --> 00:24:52,119 Speaker 1: a computer, and other elements along those lines. But at 416 00:24:52,160 --> 00:24:56,000 Speaker 1: first those weren't even really necessary because the people who 417 00:24:56,040 --> 00:24:58,600 Speaker 1: really wanted to get access to those systems didn't bother 418 00:24:58,760 --> 00:25:03,520 Speaker 1: programming better AI. They just went and started paying people 419 00:25:03,600 --> 00:25:07,120 Speaker 1: to fill out capture forms. Those who wanted to continue 420 00:25:07,119 --> 00:25:10,120 Speaker 1: the game giving the systems, they created a new industry. 421 00:25:10,240 --> 00:25:13,600 Speaker 1: They'd pay the people to fill out all these capture fields. 422 00:25:13,640 --> 00:25:16,280 Speaker 1: There was no need to develop any sort of AI. 423 00:25:16,440 --> 00:25:19,440 Speaker 1: People were doing what people were supposed to be doing easily. 424 00:25:19,480 --> 00:25:22,520 Speaker 1: They were solving captures. Now, the pay was super low 425 00:25:23,320 --> 00:25:25,679 Speaker 1: and the output was super high, and it posed a 426 00:25:25,680 --> 00:25:29,440 Speaker 1: threat to the capture system. Now, as an analogy, amount 427 00:25:29,520 --> 00:25:32,440 Speaker 1: imagine that you build a big fence strong enough to 428 00:25:32,520 --> 00:25:35,720 Speaker 1: keep bears out. No bears will get in this fence, 429 00:25:35,800 --> 00:25:38,119 Speaker 1: you say, and you go on your married a little way. 430 00:25:38,320 --> 00:25:40,320 Speaker 1: What you didn't notice is that there were gaps in 431 00:25:40,359 --> 00:25:43,680 Speaker 1: the fence that while the bears are far too big 432 00:25:43,720 --> 00:25:47,000 Speaker 1: to fit through the gaps, the gaps are big enough 433 00:25:47,040 --> 00:25:50,639 Speaker 1: to let rabid I don't know possums through, And so 434 00:25:50,760 --> 00:25:55,879 Speaker 1: the bears who go to employ rabid possums, paying them handsomely, 435 00:25:55,920 --> 00:25:58,600 Speaker 1: are able to access the stuff behind your fence anyway, 436 00:25:58,640 --> 00:26:02,360 Speaker 1: because the rapid possums pass right through the security. They 437 00:26:02,400 --> 00:26:05,200 Speaker 1: weren't intended to be kept out. Of course, in the 438 00:26:05,240 --> 00:26:08,760 Speaker 1: case the captures, we are talking about people accessing the system. 439 00:26:08,800 --> 00:26:11,239 Speaker 1: They were just doing so in massive numbers and for 440 00:26:11,400 --> 00:26:15,000 Speaker 1: less than ethical reasons. So the Carnegie Melon team began 441 00:26:15,040 --> 00:26:18,440 Speaker 1: to consider a new approach. That's when they developed recapture. 442 00:26:19,480 --> 00:26:23,119 Speaker 1: This tech used images of real words and numbers taken 443 00:26:23,160 --> 00:26:27,920 Speaker 1: from old documents. The original run was of New York 444 00:26:27,960 --> 00:26:32,679 Speaker 1: Times archival texts, but eventually the teams sold this technology 445 00:26:32,720 --> 00:26:35,159 Speaker 1: to Google, which began to use it on lots and 446 00:26:35,200 --> 00:26:40,119 Speaker 1: lots of scanned books. They were trying to transcribe those 447 00:26:40,119 --> 00:26:43,879 Speaker 1: old books. The company used recapture to display scanned words 448 00:26:43,920 --> 00:26:47,919 Speaker 1: or numbers from the texts, and as more people filled 449 00:26:47,920 --> 00:26:50,720 Speaker 1: out the recaptures, Google began to use that data to 450 00:26:50,760 --> 00:26:52,879 Speaker 1: transcribe these old works, which meant that they had a 451 00:26:52,880 --> 00:26:57,040 Speaker 1: digital copy of these books that they had come into 452 00:26:57,040 --> 00:27:00,200 Speaker 1: possession of, which means anyone filling those fields out was 453 00:27:00,240 --> 00:27:04,440 Speaker 1: actually technically doing real work for Google, including all those 454 00:27:04,440 --> 00:27:07,600 Speaker 1: folks who were being employed to write out captures. Meanwhile, 455 00:27:07,640 --> 00:27:11,199 Speaker 1: bot developers were making better bots, and character recognition and 456 00:27:11,240 --> 00:27:15,400 Speaker 1: analysis software was getting better at increasing success rates with 457 00:27:15,520 --> 00:27:19,160 Speaker 1: visual captures. Now that would prompt capture designers to make 458 00:27:19,200 --> 00:27:22,840 Speaker 1: more challenging captures, and soon we reach a real problem. 459 00:27:22,960 --> 00:27:25,560 Speaker 1: The whole point of capture was that it was supposed 460 00:27:25,560 --> 00:27:27,680 Speaker 1: to be easy for a human to complete, but difficult 461 00:27:27,720 --> 00:27:31,000 Speaker 1: for a bot to complete. If it becomes tricky for humans, 462 00:27:31,080 --> 00:27:37,639 Speaker 1: you've defeated its original purpose. Now Google updated capture to 463 00:27:37,720 --> 00:27:40,679 Speaker 1: the familiar I'm not a robot check box that you 464 00:27:40,720 --> 00:27:43,240 Speaker 1: can still find on some online forms. They call it 465 00:27:43,280 --> 00:27:49,760 Speaker 1: the no capture recapture catchy. It wasn't just a check 466 00:27:49,840 --> 00:27:53,560 Speaker 1: box that needed checking. Behind the scenes, back if you 467 00:27:53,600 --> 00:27:56,000 Speaker 1: were able to stare at the back side of the 468 00:27:56,040 --> 00:27:59,960 Speaker 1: website that you're on, software was analyzing your clicking style 469 00:28:00,000 --> 00:28:02,359 Speaker 1: all so it would look for stuff like was the 470 00:28:02,400 --> 00:28:05,480 Speaker 1: box clicked right away, perhaps before or at the same 471 00:28:05,520 --> 00:28:08,199 Speaker 1: time as fields were being filled in. If so, that 472 00:28:08,240 --> 00:28:10,639 Speaker 1: indicates a bot rather than a human being. But this 473 00:28:10,680 --> 00:28:13,480 Speaker 1: approach also doesn't get around the fact that you could 474 00:28:13,480 --> 00:28:17,960 Speaker 1: employ real human beings to do this same work. So well, 475 00:28:17,960 --> 00:28:19,800 Speaker 1: it's an effective way to tell the difference between a 476 00:28:19,840 --> 00:28:23,919 Speaker 1: bot and a person. It's not necessarily effective in keeping 477 00:28:23,960 --> 00:28:26,679 Speaker 1: spam traffic away from a site if people are willing 478 00:28:26,720 --> 00:28:31,200 Speaker 1: to employ actual human beings to do it. In sen 479 00:28:31,240 --> 00:28:34,760 Speaker 1: Google killed off this version of capture on its own services. 480 00:28:34,800 --> 00:28:37,640 Speaker 1: You can still find it everywhere else, but these days 481 00:28:37,640 --> 00:28:43,760 Speaker 1: Google uses invisible recaptures. Now, this version analyzes your browsing behavior, 482 00:28:43,840 --> 00:28:46,640 Speaker 1: and there aren't a lot of details released about it yet, 483 00:28:46,680 --> 00:28:51,320 Speaker 1: but presumably Google is looking at how any given agent 484 00:28:51,480 --> 00:28:55,200 Speaker 1: on a website uses a web page to determine if, 485 00:28:55,560 --> 00:28:58,040 Speaker 1: in fact, that is an honest to goodness human being, 486 00:28:58,200 --> 00:29:00,479 Speaker 1: or if the terminator has the said to pop over 487 00:29:00,560 --> 00:29:05,880 Speaker 1: to Zeppos to look for some new kicks. So this 488 00:29:05,960 --> 00:29:09,400 Speaker 1: is still in a way of being able to differentiate 489 00:29:09,560 --> 00:29:13,400 Speaker 1: humans from machines based solely upon behavior, just analyzing the 490 00:29:13,440 --> 00:29:17,960 Speaker 1: behavior and thinking all right, well, this indicates a human being. 491 00:29:18,040 --> 00:29:21,160 Speaker 1: This is this person, This entity is navigating a web 492 00:29:21,160 --> 00:29:24,280 Speaker 1: page the way a human would versus this is really 493 00:29:24,320 --> 00:29:28,600 Speaker 1: efficient and formulaic and repetitive, and that tells me that's 494 00:29:28,600 --> 00:29:32,560 Speaker 1: possibly a machine. So let's switch over to Twitter. Twitter 495 00:29:32,640 --> 00:29:35,520 Speaker 1: has got a lot of bots on it. Twitter and 496 00:29:35,520 --> 00:29:39,600 Speaker 1: the follower numbers are kind of a type of status online. 497 00:29:39,960 --> 00:29:43,760 Speaker 1: If you have more followers than the general implication is 498 00:29:43,800 --> 00:29:46,760 Speaker 1: that you must be more important than someone who has 499 00:29:46,800 --> 00:29:52,520 Speaker 1: fewer followers, and so there's a healthy market for purchased followers. 500 00:29:52,640 --> 00:29:57,360 Speaker 1: On Twitter. You can go to several different companies and 501 00:29:57,480 --> 00:30:02,480 Speaker 1: stores and buy followers by the hundreds or thousands. So 502 00:30:02,520 --> 00:30:04,680 Speaker 1: if you're desperate to boost that number, you can pay 503 00:30:04,680 --> 00:30:07,640 Speaker 1: a service that will link accounts to your account. Now, 504 00:30:07,680 --> 00:30:10,560 Speaker 1: most of those probably do not have real, live human 505 00:30:10,600 --> 00:30:13,920 Speaker 1: beings behind those accounts, and so a visit to any 506 00:30:14,040 --> 00:30:16,080 Speaker 1: of those accounts will show you that they never seem 507 00:30:16,160 --> 00:30:19,080 Speaker 1: to say anything themselves. They'll retweet what lots of other 508 00:30:19,080 --> 00:30:22,600 Speaker 1: people are saying, but they don't actually, you know, tweet 509 00:30:22,640 --> 00:30:26,400 Speaker 1: anything of their own, or if they do, it makes 510 00:30:26,480 --> 00:30:28,880 Speaker 1: little to no sense. It might just be kind of 511 00:30:28,920 --> 00:30:34,400 Speaker 1: like garbled general you know, new a g kind of stuff, 512 00:30:34,520 --> 00:30:37,840 Speaker 1: the things that that sound like they might have some 513 00:30:37,920 --> 00:30:39,960 Speaker 1: sort of deep meaning, but if you think about it, 514 00:30:39,960 --> 00:30:43,360 Speaker 1: you realize, no, that really doesn't mean anything at all. Now, 515 00:30:43,400 --> 00:30:46,160 Speaker 1: on a one on one basis, Twitter bots are pretty 516 00:30:46,160 --> 00:30:48,560 Speaker 1: easy to spot. So let's say your you tweet about 517 00:30:48,600 --> 00:30:50,880 Speaker 1: something important going on, such as you know something's going 518 00:30:50,920 --> 00:30:53,000 Speaker 1: on in politics or whether John Snow is going to 519 00:30:53,080 --> 00:30:56,160 Speaker 1: win the Game of Thrones, and almost immediately after you tweet, 520 00:30:56,200 --> 00:30:59,560 Speaker 1: you notice a new followed notification and if it popped 521 00:30:59,600 --> 00:31:03,080 Speaker 1: up sue pretty quickly, like instantly after you posted a tweet. 522 00:31:03,600 --> 00:31:05,680 Speaker 1: That might very well be a bot running on a 523 00:31:05,720 --> 00:31:09,560 Speaker 1: script that is searching for instances of specific keywords, and 524 00:31:09,560 --> 00:31:13,400 Speaker 1: when it finds those keywords, it then prompts the bot 525 00:31:13,440 --> 00:31:17,920 Speaker 1: account to follow the account that generated the keywords, assuming 526 00:31:17,960 --> 00:31:21,200 Speaker 1: that hasn't already followed that account. And some butts do 527 00:31:21,280 --> 00:31:23,360 Speaker 1: this in order to convince people to follow them back 528 00:31:23,680 --> 00:31:26,000 Speaker 1: because lots of folks on Twitter have a follow back 529 00:31:26,120 --> 00:31:28,800 Speaker 1: policy which helps them boost up their own follower numbers. 530 00:31:28,800 --> 00:31:30,800 Speaker 1: You know, it's the whole hey, if you follow me, 531 00:31:30,880 --> 00:31:34,480 Speaker 1: I'll follow you quit bro quo kind of approach. But 532 00:31:34,560 --> 00:31:37,280 Speaker 1: in this case, one of the two parties is a bot, 533 00:31:37,760 --> 00:31:39,680 Speaker 1: at least one of them. Anyway, maybe they both are, 534 00:31:39,840 --> 00:31:43,440 Speaker 1: which is kind of funny and pointless. Now, once you 535 00:31:43,560 --> 00:31:46,760 Speaker 1: follow the bot, you may start seeing spam messages from 536 00:31:46,760 --> 00:31:49,320 Speaker 1: that bot pop up in your feed. Whenever it occasionally 537 00:31:49,360 --> 00:31:52,760 Speaker 1: posts two followers, it's likely trying to get you to 538 00:31:52,840 --> 00:31:56,080 Speaker 1: engage in a particular behavior. Now that behavior might be 539 00:31:56,120 --> 00:31:59,000 Speaker 1: more or less benign, such as convincing you to shop 540 00:31:59,040 --> 00:32:03,920 Speaker 1: a certain brand which is obnoxious but not you know, malicious, 541 00:32:04,480 --> 00:32:06,480 Speaker 1: Or it might be more sinisters, such as trying to 542 00:32:06,480 --> 00:32:08,880 Speaker 1: get you to do something foolish that will compromise your 543 00:32:08,880 --> 00:32:12,280 Speaker 1: computer and allow it to join like a hacker's bot 544 00:32:12,280 --> 00:32:15,560 Speaker 1: net army or something. And there's a lot of reasons, 545 00:32:15,600 --> 00:32:18,360 Speaker 1: most of them annoying, that a bot programmer would want 546 00:32:18,360 --> 00:32:21,560 Speaker 1: you to follow their butt. According to a study conducted 547 00:32:21,560 --> 00:32:25,640 Speaker 1: by researchers at Indiana University and the University of Southern California, 548 00:32:25,880 --> 00:32:29,880 Speaker 1: somewhere between nine and fifteen percent of all active Twitter 549 00:32:29,960 --> 00:32:33,840 Speaker 1: accounts are actually bots. It usually doesn't require a lot 550 00:32:33,880 --> 00:32:36,440 Speaker 1: of work to determine if a single account is the 551 00:32:36,480 --> 00:32:38,880 Speaker 1: work of an actual human being, but if you have 552 00:32:38,960 --> 00:32:40,880 Speaker 1: a lot of them, that can be a challenge. I mean, 553 00:32:40,880 --> 00:32:44,200 Speaker 1: if you've got thousands of followers, sorting through all of 554 00:32:44,240 --> 00:32:48,560 Speaker 1: those would take a real long time. So that's what 555 00:32:48,640 --> 00:32:52,160 Speaker 1: prompted developers to create apps like butt or Not, which 556 00:32:52,280 --> 00:32:55,320 Speaker 1: scour Twitter followers and look for signs of butts, returning 557 00:32:55,320 --> 00:32:57,200 Speaker 1: a report to the user to let him or her 558 00:32:57,240 --> 00:33:01,080 Speaker 1: know how many legitimate followers they have. Those apps, which 559 00:33:01,080 --> 00:33:04,920 Speaker 1: you can argue are are kind of bots themselves, look 560 00:33:04,960 --> 00:33:09,200 Speaker 1: for indicators such as each followers Twitter description uh the 561 00:33:09,320 --> 00:33:12,280 Speaker 1: u r L field, the number of tweets the account 562 00:33:12,320 --> 00:33:15,520 Speaker 1: has generated of its own, the number of followers the 563 00:33:15,560 --> 00:33:19,360 Speaker 1: account has, and so on. So if you come across 564 00:33:19,360 --> 00:33:23,160 Speaker 1: an account that follows thousands of other accounts but only 565 00:33:23,200 --> 00:33:26,160 Speaker 1: as a few followers of its own, that's a red flag. 566 00:33:26,240 --> 00:33:28,920 Speaker 1: That's saying, well, this account is following lots of people, 567 00:33:28,960 --> 00:33:31,120 Speaker 1: not a lot of people follow it. That tells me 568 00:33:31,440 --> 00:33:34,520 Speaker 1: something hinky might be going on. If the description or 569 00:33:34,600 --> 00:33:38,080 Speaker 1: you are l are empty, that's another indicator because it 570 00:33:38,320 --> 00:33:40,560 Speaker 1: shows maybe someone didn't want to take the time to 571 00:33:40,640 --> 00:33:45,440 Speaker 1: try and fool people by creating a bogus description and 572 00:33:45,480 --> 00:33:49,200 Speaker 1: a bogus u r L. There are several other criteria 573 00:33:49,400 --> 00:33:52,160 Speaker 1: that the apps look for, and depending upon how many 574 00:33:52,160 --> 00:33:55,400 Speaker 1: red flag boxes get checked, the app determines if the 575 00:33:55,480 --> 00:33:58,040 Speaker 1: account is the work of a script or if it's 576 00:33:58,040 --> 00:34:02,280 Speaker 1: an actual person. Now, on the one hand, we can 577 00:34:02,320 --> 00:34:04,240 Speaker 1: look at all these stories about bots, and think of 578 00:34:04,280 --> 00:34:07,680 Speaker 1: how irritating they are because they generate spam content, they 579 00:34:07,680 --> 00:34:10,960 Speaker 1: clog up actual communication. They create deception, whether it's an 580 00:34:10,960 --> 00:34:13,880 Speaker 1: attempt to trick you into following a malicious link or 581 00:34:13,920 --> 00:34:17,000 Speaker 1: to think someone is particularly notable due to the enormous 582 00:34:17,040 --> 00:34:19,880 Speaker 1: number of Twitter followers they have. But on the other hand, 583 00:34:20,160 --> 00:34:22,200 Speaker 1: we can think about how these examples show how we're 584 00:34:22,200 --> 00:34:26,040 Speaker 1: getting better at creating more human like agents. Now that's 585 00:34:26,080 --> 00:34:29,440 Speaker 1: not to say these agents possess intelligence, only that they 586 00:34:29,480 --> 00:34:33,080 Speaker 1: can imitate human interactions enough to raise the question could 587 00:34:33,120 --> 00:34:35,759 Speaker 1: this be a bot I'm talking to? If you have 588 00:34:35,840 --> 00:34:38,760 Speaker 1: to ask that question, then that indicates programmers are getting 589 00:34:38,800 --> 00:34:41,960 Speaker 1: better at designing bots, or that you're getting pretty bad 590 00:34:41,960 --> 00:34:45,840 Speaker 1: at recognizing humans. Some days, I certainly have that problem. 591 00:34:45,960 --> 00:34:49,000 Speaker 1: We'll talk a little bit more about machine intelligence and 592 00:34:49,040 --> 00:34:52,839 Speaker 1: communication in just a minute and kind of layout why 593 00:34:52,880 --> 00:34:57,560 Speaker 1: it's so difficult to really create a truly compelling butt 594 00:34:57,960 --> 00:35:00,360 Speaker 1: that can fool people into thinking it's a human. But 595 00:35:00,480 --> 00:35:10,719 Speaker 1: first let's take another quick break to think our sponsor. Now, 596 00:35:10,760 --> 00:35:15,359 Speaker 1: they're just elements to human communication that bots are not 597 00:35:15,520 --> 00:35:18,719 Speaker 1: great at handling, or they need a huge amount of 598 00:35:18,760 --> 00:35:21,239 Speaker 1: help in order to pull it off. So let's take 599 00:35:21,280 --> 00:35:25,120 Speaker 1: IBM S Watson for example. Now, Watson is the interface 600 00:35:25,239 --> 00:35:27,800 Speaker 1: that made the news when it competed against two former 601 00:35:27,880 --> 00:35:32,120 Speaker 1: Jeopardy champions on a special edition of Jeopardy. Watson beat 602 00:35:32,200 --> 00:35:35,239 Speaker 1: the opponents, which is pretty impressive when you consider that 603 00:35:35,320 --> 00:35:39,480 Speaker 1: Jeopardys format includes elements of wordplay includes and machines are 604 00:35:39,480 --> 00:35:44,120 Speaker 1: typically not very good at interpreting word play and subtext 605 00:35:44,200 --> 00:35:45,759 Speaker 1: and that sort of thing and getting at what the 606 00:35:45,800 --> 00:35:50,040 Speaker 1: actual meaning to a sentence is. Watson even attempted a 607 00:35:50,080 --> 00:35:52,480 Speaker 1: couple of jokes throughout the course of the game, but 608 00:35:53,000 --> 00:35:57,080 Speaker 1: they weren't really spontaneous bumb malls designed to get a 609 00:35:57,239 --> 00:36:01,520 Speaker 1: chuckle of, you know, Alex Trebek. Humor is just one 610 00:36:01,520 --> 00:36:04,239 Speaker 1: of those aspects of human communication that is difficult to 611 00:36:04,320 --> 00:36:09,200 Speaker 1: quantify and implement with machines. Typically, it requires programmers to 612 00:36:09,239 --> 00:36:13,080 Speaker 1: think ahead and imagine specific scenarios and queries to build 613 00:36:13,120 --> 00:36:18,880 Speaker 1: out appropriate or, depending upon the context, inappropriate responses. So, 614 00:36:18,960 --> 00:36:23,840 Speaker 1: for example, when Apple's personal assistant Sirie debuted, people immediately 615 00:36:23,880 --> 00:36:27,000 Speaker 1: began to test Sirie. They began to ask the digital 616 00:36:27,000 --> 00:36:30,360 Speaker 1: personal assistant all sorts of odd things and sharing the results. 617 00:36:31,120 --> 00:36:33,560 Speaker 1: If you create any sort of system. One of the 618 00:36:33,600 --> 00:36:36,440 Speaker 1: first things you're going to find when you allow people 619 00:36:36,520 --> 00:36:38,920 Speaker 1: to access that system is they're going to try and 620 00:36:38,960 --> 00:36:41,080 Speaker 1: break it, or they're at least going to try and 621 00:36:41,120 --> 00:36:45,160 Speaker 1: explore what the limitations are within that system. And they're 622 00:36:45,160 --> 00:36:48,879 Speaker 1: not necessarily doing this malicious with malicious intent, but rather 623 00:36:49,040 --> 00:36:52,000 Speaker 1: that you know, we're humans, were curious. We want to 624 00:36:52,040 --> 00:36:56,320 Speaker 1: know how how far do things go? Are they really 625 00:36:56,719 --> 00:36:59,400 Speaker 1: limitless or are you going to run up against an 626 00:36:59,440 --> 00:37:02,000 Speaker 1: invisible all if you keep going in one direction long enough. 627 00:37:02,800 --> 00:37:06,560 Speaker 1: The same thing is true about personal digital assistance. So 628 00:37:07,719 --> 00:37:11,000 Speaker 1: in some cases where people were asking weird things of SIRIE, 629 00:37:11,200 --> 00:37:15,359 Speaker 1: serious responses were particularly hilarious, indicating that someone over at 630 00:37:15,360 --> 00:37:19,799 Speaker 1: Apple had anticipated some of those shenanigans because SIRIE wasn't 631 00:37:19,840 --> 00:37:24,560 Speaker 1: coming up with these wacky responses on its own account. 632 00:37:24,760 --> 00:37:28,880 Speaker 1: It was referring to a database of responses that people 633 00:37:28,920 --> 00:37:34,400 Speaker 1: had been compiling ever since they started working on the project. So, 634 00:37:34,920 --> 00:37:38,360 Speaker 1: if you are working on a personal Digital assistant project 635 00:37:38,360 --> 00:37:40,719 Speaker 1: and you think, oh, someone's gonna say I love you 636 00:37:41,239 --> 00:37:44,680 Speaker 1: eventually to this, I want to have a response to 637 00:37:44,800 --> 00:37:47,520 Speaker 1: come back that isn't just I'm sorry. I don't understand 638 00:37:47,520 --> 00:37:51,520 Speaker 1: that every time the digital assistant says, I'm sorry, I 639 00:37:51,560 --> 00:37:59,560 Speaker 1: don't understand that is an overall, like outright admission of limitations. 640 00:37:59,600 --> 00:38:02,319 Speaker 1: So you will try avoid that as much as you can. 641 00:38:02,440 --> 00:38:04,360 Speaker 1: Make it kind of a joke instead. But it means 642 00:38:04,400 --> 00:38:07,680 Speaker 1: thinking ahead, and it means the humans are thinking ahead. 643 00:38:07,960 --> 00:38:12,640 Speaker 1: It's not a machine. Uh So, here's an example. One 644 00:38:12,680 --> 00:38:15,480 Speaker 1: of the early queries that got widespread traction was I 645 00:38:15,520 --> 00:38:18,960 Speaker 1: need to hide a body, and Siri would respond originally 646 00:38:19,239 --> 00:38:24,080 Speaker 1: with various sites where you could, you know, possibly dump 647 00:38:24,160 --> 00:38:26,960 Speaker 1: a body and get away with it, like nearby reservoirs 648 00:38:27,040 --> 00:38:30,680 Speaker 1: or quarries. It's pretty grim, but darkly humorous, and it 649 00:38:30,719 --> 00:38:33,960 Speaker 1: showed that someone had been thinking those things through by 650 00:38:34,000 --> 00:38:36,680 Speaker 1: the way, that joke became very serious. In two thousand twelve, 651 00:38:36,880 --> 00:38:39,840 Speaker 1: a Florida man stood accused of murdering a friend of his, 652 00:38:40,080 --> 00:38:44,560 Speaker 1: a roommate, and on his phone the suspects phone was 653 00:38:44,600 --> 00:38:47,640 Speaker 1: a screenshot of a query to Sirie, the one about 654 00:38:47,640 --> 00:38:51,560 Speaker 1: where to hide his roommate, and prosecutors used it as 655 00:38:51,600 --> 00:38:54,280 Speaker 1: evidence in the trial. But it turned out the screenshot 656 00:38:54,320 --> 00:38:56,960 Speaker 1: that used wasn't really a query that the man had 657 00:38:57,000 --> 00:39:01,239 Speaker 1: made himself, because his iPhone was as an iPhone that 658 00:39:01,280 --> 00:39:05,239 Speaker 1: was running on Verizons service, and the screenshot was from 659 00:39:05,280 --> 00:39:06,960 Speaker 1: an iPhone that was running on a T and T 660 00:39:07,160 --> 00:39:11,200 Speaker 1: S service. Uh. Also, it turned out that the phone 661 00:39:11,280 --> 00:39:14,279 Speaker 1: he was using, the suspect was an older model of 662 00:39:14,320 --> 00:39:18,160 Speaker 1: iPhone that didn't even support SIRIE. However, he was later 663 00:39:18,239 --> 00:39:20,960 Speaker 1: found guilty of his crime, though the Serie connection was 664 00:39:21,040 --> 00:39:25,879 Speaker 1: again dismissed for those multiple reasons. Later on, Apple would 665 00:39:26,120 --> 00:39:30,319 Speaker 1: replace that joking response with a referential but less morbid joke, 666 00:39:30,440 --> 00:39:33,440 Speaker 1: which was quote I used to know the answer to 667 00:39:33,520 --> 00:39:37,319 Speaker 1: this question end quote, So, in other words, acknowledging that, 668 00:39:37,360 --> 00:39:39,800 Speaker 1: in fact, there used to be another response without actually 669 00:39:39,800 --> 00:39:45,080 Speaker 1: giving it because of you know, these very grim, macabre 670 00:39:45,239 --> 00:39:48,680 Speaker 1: reasons in real life. But let's say you wanted to 671 00:39:48,680 --> 00:39:53,680 Speaker 1: create an artificial entity that could respond with humor dynamically. 672 00:39:54,000 --> 00:39:56,920 Speaker 1: It wouldn't require you to pre program in responses to 673 00:39:57,080 --> 00:40:00,920 Speaker 1: different questions you'd have to anticipate. This would let you 674 00:40:00,960 --> 00:40:03,840 Speaker 1: have a bot that could convincingly stand in as a 675 00:40:03,960 --> 00:40:07,239 Speaker 1: human without the danger of the bot encountering something you 676 00:40:07,280 --> 00:40:10,000 Speaker 1: didn't expect and having no response to it, or to 677 00:40:10,160 --> 00:40:14,080 Speaker 1: misinterpreting the interaction with an actual human being, or if 678 00:40:14,080 --> 00:40:16,680 Speaker 1: it did misinterpret it that it could follow up in 679 00:40:16,680 --> 00:40:20,440 Speaker 1: a very human way. So if I make a joke 680 00:40:20,880 --> 00:40:24,360 Speaker 1: to my coworkers and I do it well, my coworkers 681 00:40:24,440 --> 00:40:27,719 Speaker 1: understand what the meaning of the joke was, what the 682 00:40:27,800 --> 00:40:30,919 Speaker 1: intended meaning of the joke was, and there you get 683 00:40:30,920 --> 00:40:34,319 Speaker 1: that response. If the joke doesn't go well, I can 684 00:40:34,400 --> 00:40:37,120 Speaker 1: follow it up by explaining the joke or explaining what 685 00:40:37,200 --> 00:40:40,320 Speaker 1: I had tried to do with the joke, which doesn't 686 00:40:40,360 --> 00:40:43,480 Speaker 1: make the joke funny, but at least informs the audience 687 00:40:43,560 --> 00:40:46,480 Speaker 1: as to what it was I was thinking. Machines would 688 00:40:46,520 --> 00:40:47,960 Speaker 1: have to be able to do that too, and this 689 00:40:48,040 --> 00:40:50,960 Speaker 1: is hard to do. Machines would need to be able 690 00:40:51,000 --> 00:40:53,840 Speaker 1: to interpret not only the literal meaning of any statement, 691 00:40:54,160 --> 00:40:58,239 Speaker 1: but the potential intended meanings as well. So I would 692 00:40:58,280 --> 00:41:02,200 Speaker 1: have to incorporate the concept of novelties, introducing something new 693 00:41:02,239 --> 00:41:05,319 Speaker 1: and unexpected into the interaction. It's a subversion of our 694 00:41:05,360 --> 00:41:08,920 Speaker 1: expectations that tends to lead to humor. So, for example, 695 00:41:08,960 --> 00:41:11,760 Speaker 1: Douglas Adams, who is one of my favorite authors, once 696 00:41:11,800 --> 00:41:15,600 Speaker 1: wrote a sentence describing a fleet of spaceships, and the 697 00:41:15,680 --> 00:41:19,640 Speaker 1: sentence goes like this, The ships hung in the sky 698 00:41:19,960 --> 00:41:24,200 Speaker 1: in much the same way that bricks didn't. Now that's 699 00:41:24,200 --> 00:41:27,520 Speaker 1: a great sentence. It gets across the humor and intent 700 00:41:27,640 --> 00:41:29,840 Speaker 1: to the reader. You know that if you were to 701 00:41:29,840 --> 00:41:32,560 Speaker 1: see these spaceships in the sky, they would look completely 702 00:41:32,640 --> 00:41:36,720 Speaker 1: out of place. They might even be remotely brick shaped. 703 00:41:36,800 --> 00:41:40,400 Speaker 1: But mostly it's the idea that if bricks could hang 704 00:41:40,440 --> 00:41:43,520 Speaker 1: in the air, those ships would look like that, except 705 00:41:43,560 --> 00:41:47,680 Speaker 1: obviously bricks can't hang in the air, And in one sentence, 706 00:41:48,000 --> 00:41:51,560 Speaker 1: Adams is able to convey with humor the mind bendingly 707 00:41:51,680 --> 00:41:56,440 Speaker 1: weird experience of seeing these spaceships in the Earth's sky. 708 00:41:56,600 --> 00:41:59,719 Speaker 1: Computers would have a real hard time replicating that, at 709 00:41:59,800 --> 00:42:03,360 Speaker 1: least on purpose. A computer program that put rough sentences 710 00:42:03,400 --> 00:42:08,600 Speaker 1: together using a basic syntax and vocabulary could potentially make 711 00:42:08,680 --> 00:42:12,879 Speaker 1: weird and funny sentences, but these would be mostly random 712 00:42:12,920 --> 00:42:16,080 Speaker 1: and frequently meaningless, and you wouldn't be able to hold 713 00:42:16,120 --> 00:42:20,040 Speaker 1: a context from sentence to sentence. To make something that 714 00:42:20,080 --> 00:42:25,960 Speaker 1: has meaning requires aspects of intelligence that computers don't yet possess. Watson, 715 00:42:26,280 --> 00:42:29,960 Speaker 1: with its jokes, was running on a massively powerful computer 716 00:42:30,040 --> 00:42:34,239 Speaker 1: system with two thousand, eight hundred eighty processing cores, and 717 00:42:34,320 --> 00:42:37,200 Speaker 1: that doesn't even approach the power necessary to create real 718 00:42:37,360 --> 00:42:43,080 Speaker 1: humor spontaneously. To detect and generate sarcasm, and entity must 719 00:42:43,160 --> 00:42:47,160 Speaker 1: understand context and other cues and machines aren't very good 720 00:42:47,200 --> 00:42:50,320 Speaker 1: at this, though we've seen some advances in contextual tracking. 721 00:42:50,719 --> 00:42:54,160 Speaker 1: For example, Google's Personal Assistant can follow a line of 722 00:42:54,280 --> 00:42:57,600 Speaker 1: questions about the same subject without you having to restate 723 00:42:57,680 --> 00:43:01,000 Speaker 1: the subject with each question. If I asked my Google 724 00:43:01,040 --> 00:43:04,480 Speaker 1: Home what when the next Braves game is, it would 725 00:43:04,520 --> 00:43:07,200 Speaker 1: give me an answer. Let's say it's day after tomorrow. Well, 726 00:43:07,239 --> 00:43:09,320 Speaker 1: I could follow that up with what will the weather 727 00:43:09,360 --> 00:43:12,319 Speaker 1: be like then? And the system would understand that by 728 00:43:12,480 --> 00:43:14,920 Speaker 1: then I mean the day of the game, So the 729 00:43:15,000 --> 00:43:18,760 Speaker 1: day after tomorrow. I might also ask what's the fastest 730 00:43:18,800 --> 00:43:21,799 Speaker 1: way there, and it will know that by there I 731 00:43:21,840 --> 00:43:24,560 Speaker 1: mean the stadium, and that I am probably am asking 732 00:43:24,560 --> 00:43:27,880 Speaker 1: how to get from my current location to that stadium 733 00:43:27,960 --> 00:43:31,160 Speaker 1: and the most efficient way possible. The subject is stored 734 00:43:31,200 --> 00:43:34,120 Speaker 1: in temporary memory, I don't have to keep asking specific 735 00:43:34,239 --> 00:43:37,920 Speaker 1: questions about the game or the stadium. But that's still 736 00:43:37,960 --> 00:43:42,680 Speaker 1: a long way off from actually understanding context. So one 737 00:43:42,800 --> 00:43:46,160 Speaker 1: test for bots might be for us to have it 738 00:43:46,239 --> 00:43:49,080 Speaker 1: tell us a joke. If it's clear that the bot 739 00:43:49,160 --> 00:43:51,920 Speaker 1: can create a brand new joke, one that has not 740 00:43:52,080 --> 00:43:57,400 Speaker 1: been pre programmed, one that is spontaneous and novel and 741 00:43:57,480 --> 00:44:00,680 Speaker 1: created by the bot itself, and it makes sense and 742 00:44:00,800 --> 00:44:04,520 Speaker 1: it is funny. We've reached a point where telling bots 743 00:44:04,600 --> 00:44:08,759 Speaker 1: and humans apart is going to be tremendously complicated, but 744 00:44:08,880 --> 00:44:11,920 Speaker 1: right now we're nowhere near that. The jokes that we 745 00:44:12,000 --> 00:44:15,759 Speaker 1: hear bots tell, for the most part, are ones that 746 00:44:15,760 --> 00:44:18,200 Speaker 1: have been created by human beings and just stored in 747 00:44:18,239 --> 00:44:20,680 Speaker 1: a database, and the body just pulls them out and 748 00:44:20,719 --> 00:44:24,320 Speaker 1: then recites them. It's not creating them. It's just pulling 749 00:44:24,920 --> 00:44:29,200 Speaker 1: a massive data from a cell in a giant spreadsheet 750 00:44:29,239 --> 00:44:31,399 Speaker 1: and saying, all right, this is the joke I'm gonna 751 00:44:31,440 --> 00:44:35,080 Speaker 1: tell us. The joke that's in sell see four and seventeen. 752 00:44:35,400 --> 00:44:38,960 Speaker 1: That's the joke for today. That's not creating a joke, 753 00:44:39,040 --> 00:44:42,120 Speaker 1: it's just reciting one. If we can get to a 754 00:44:42,160 --> 00:44:44,719 Speaker 1: point where they can create jokes, that's a big jump 755 00:44:44,719 --> 00:44:49,000 Speaker 1: in computer intelligence and maybe a brand new audience from 756 00:44:49,040 --> 00:44:53,360 Speaker 1: my type of humor I'm always looking. Well, that pretty 757 00:44:53,400 --> 00:44:57,360 Speaker 1: much wraps up this episode. Really, the key to determining 758 00:44:57,360 --> 00:45:00,200 Speaker 1: whether or not it's a bot or a hu Man 759 00:45:00,480 --> 00:45:04,799 Speaker 1: is testing whether or not it's capable of handling novelty. 760 00:45:05,480 --> 00:45:09,480 Speaker 1: Most bots are fairly limited in the scope of things 761 00:45:09,480 --> 00:45:11,560 Speaker 1: they can handle, and if you step outside of that, 762 00:45:11,640 --> 00:45:14,840 Speaker 1: you see those limitations pretty quickly, and that then it 763 00:45:14,840 --> 00:45:18,080 Speaker 1: becomes apparent. But every year we're getting a little bit 764 00:45:18,080 --> 00:45:24,000 Speaker 1: better at handling wider spectrums of experiences with bots, so 765 00:45:24,040 --> 00:45:26,480 Speaker 1: that it becomes more and more complicated to tell them 766 00:45:26,520 --> 00:45:30,720 Speaker 1: apart from human beings. Uh. In most cases it's probably 767 00:45:30,920 --> 00:45:33,360 Speaker 1: a moot point. It's not really necessary depending upon what 768 00:45:33,440 --> 00:45:36,200 Speaker 1: it is you're trying to do, But in some cases 769 00:45:36,760 --> 00:45:38,359 Speaker 1: you really do want to know whether or not that's 770 00:45:38,360 --> 00:45:40,560 Speaker 1: a human being or a machine. On the other end, 771 00:45:41,200 --> 00:45:44,799 Speaker 1: if you guys have any stories about funny times where 772 00:45:44,840 --> 00:45:46,920 Speaker 1: you were chatting with something that you thought was a 773 00:45:46,960 --> 00:45:48,520 Speaker 1: human and turned out to be a bot, Like I've 774 00:45:48,560 --> 00:45:52,319 Speaker 1: got friends who have received robo calls and didn't know 775 00:45:52,560 --> 00:45:55,880 Speaker 1: until about half a minute in or maybe a minute in, 776 00:45:56,000 --> 00:45:59,279 Speaker 1: that it was a robot. Those are great stories. I 777 00:45:59,280 --> 00:46:02,000 Speaker 1: have specifically love the ones where if you ask the 778 00:46:02,160 --> 00:46:06,360 Speaker 1: entity are you a robot? It tries to deflect but 779 00:46:06,560 --> 00:46:09,920 Speaker 1: does not actually answer the question. Those are the best. 780 00:46:10,560 --> 00:46:12,400 Speaker 1: But you can get in touch with me, let me 781 00:46:12,440 --> 00:46:16,120 Speaker 1: know your experiences. The email address is tech stuff at 782 00:46:16,160 --> 00:46:18,440 Speaker 1: how stuff works dot com, or you can drop me 783 00:46:18,480 --> 00:46:21,279 Speaker 1: a line on Twitter or Facebook. The handle for the 784 00:46:21,280 --> 00:46:25,040 Speaker 1: show at both of those is tech stuff hs W. Remember, 785 00:46:25,520 --> 00:46:28,760 Speaker 1: normally you can watch me record shows live at twitch 786 00:46:28,800 --> 00:46:32,560 Speaker 1: dot tv slash tech stuff. I record on Wednesdays and Friday's. 787 00:46:32,760 --> 00:46:36,399 Speaker 1: Today's episode is a little bit outside the norm. There 788 00:46:36,480 --> 00:46:38,839 Speaker 1: is no one currently watching me live, so when I'm 789 00:46:38,840 --> 00:46:42,000 Speaker 1: doing my dance like I am right now, no one 790 00:46:42,080 --> 00:46:45,600 Speaker 1: can see. But most days you can see, and I 791 00:46:45,680 --> 00:46:48,520 Speaker 1: do the dance then too. So join me at twitch 792 00:46:48,520 --> 00:46:50,799 Speaker 1: dot tv slash tech stuff to watch the show live. 793 00:46:50,840 --> 00:46:53,600 Speaker 1: You get to see all the elements of the show 794 00:46:53,680 --> 00:46:57,720 Speaker 1: come together, and I will talk to you again. Really 795 00:47:03,280 --> 00:47:05,680 Speaker 1: for more on this and thousands of other topics, is 796 00:47:05,719 --> 00:47:16,240 Speaker 1: that how stuff works dot com.