WEBVTT - AI Gone Rude 0:00:04.120 --> 0:00:07.160 Get in touch with technology with tech Stuff from how 0:00:07.200 --> 0:00:13.920 stuff Works dot com. Hey there, and welcome to tech Stuff. 0:00:13.960 --> 0:00:16.360 I'm your host, John that Strickland. I'm an executive producer 0:00:16.360 --> 0:00:19.279 with how Stuff Works in Love all Things Tech, and 0:00:19.440 --> 0:00:21.959 last week I did an episode about whether or not 0:00:22.120 --> 0:00:25.120 we could ever develop an artificially intelligent machine that could 0:00:25.239 --> 0:00:28.560 understand not just what we say, but what we actually 0:00:28.720 --> 0:00:32.960 mean when we employ stuff like sarcasm or metaphors. Today, 0:00:33.040 --> 0:00:35.919 we're going to look at some notable instances of machines 0:00:37.080 --> 0:00:41.199 behaving badly after well meaning designers gave those machines a 0:00:41.200 --> 0:00:44.240 bit too much freedom in this regard. Now, the stories 0:00:44.280 --> 0:00:48.200 I'm going to focus on are on the surface, pretty funny, 0:00:48.440 --> 0:00:52.720 but they illustrate a real challenge in artificial intelligence, because 0:00:53.159 --> 0:00:55.920 designing a system that does what you intended to do 0:00:56.200 --> 0:00:58.840 is harder than it might seem, especially as you make 0:00:58.880 --> 0:01:02.200 that system more and more autonomous, it can behave in 0:01:02.280 --> 0:01:06.080 ways that you were not able to predict. So this 0:01:06.160 --> 0:01:09.840 is a topic that science fiction authors have covered extensively. 0:01:10.280 --> 0:01:13.960 In fiction, there's something of a trope around the concept 0:01:14.000 --> 0:01:18.000 of the artificially intelligent system that causes harm in an 0:01:18.040 --> 0:01:21.280 effort to help So there's a classic thought experiment, and 0:01:21.319 --> 0:01:25.000 it revolves around asking a super intelligent machine to bring 0:01:25.040 --> 0:01:28.200 about world peace. Right, you do, You designed the supercomputer, 0:01:28.319 --> 0:01:30.960 it's smarter than any human, and you say, I want 0:01:31.000 --> 0:01:33.560 you to solve the problem of world peace. I want 0:01:33.560 --> 0:01:35.640 there to be world peace. And the machine runs the 0:01:35.640 --> 0:01:38.880 calculations and it comes to the conclusion that as long 0:01:38.920 --> 0:01:41.920 as there are two or more people living on the planet, 0:01:42.319 --> 0:01:45.400 world peace cannot be assured, as there is always the 0:01:45.520 --> 0:01:49.040 chance for conflict. And so the super intelligent machine wipes 0:01:49.080 --> 0:01:52.920 out humanity, or at least everybody but one person. This 0:01:53.000 --> 0:01:57.240 is clearly a worst case scenario of artificial intelligence behaving 0:01:57.240 --> 0:02:00.600 in a way you did not anticipate, and it's light 0:02:00.720 --> 0:02:03.520 years away from the stories I'm going to talk about today. 0:02:03.560 --> 0:02:06.040 But it is good to remember that while the incidents 0:02:06.040 --> 0:02:09.799 I'm going to cover are largely humorous to us today, 0:02:10.080 --> 0:02:13.800 they illustrate that intelligence is a very tricky subject. Also, 0:02:13.840 --> 0:02:18.320 on that matter, intelligence itself is pretty difficult to define. 0:02:18.520 --> 0:02:22.720 Along with other concepts like consciousness, these are very hard 0:02:23.000 --> 0:02:26.720 to nail down and define in concrete terms, and in 0:02:26.760 --> 0:02:30.600 computer science, artificial intelligence covers a an enormous amount of 0:02:30.639 --> 0:02:33.240 ground I've talked about this in previous episodes of Tech Stuff. 0:02:33.800 --> 0:02:37.160 Someone who's working in image recognition is working on one 0:02:37.200 --> 0:02:40.400 aspect of artificial intelligence. The same is true for voice 0:02:40.400 --> 0:02:45.640 recognition or natural language processing, machine learning, path finding. So 0:02:45.680 --> 0:02:48.720 while I'm talking about AI, I'm not talking about thinking 0:02:48.760 --> 0:02:50.800 like a human being. I'm not talking about creating a 0:02:50.840 --> 0:02:55.360 machine that can internalize and associate ideas the way a 0:02:55.440 --> 0:02:57.840 human can. The machines I'm going to be covering our 0:02:57.919 --> 0:03:02.480 processing information and arriving conclusions, but they are not thinking 0:03:02.960 --> 0:03:06.240 the same way that people do. So let's start off 0:03:06.680 --> 0:03:10.160 with Watson. And I mentioned IBMS Watson platform in the 0:03:10.240 --> 0:03:13.400 Sarcasm episode a couple of times, and that's because it's 0:03:13.400 --> 0:03:16.399 one of the more visible artificial intelligence platforms out there 0:03:16.480 --> 0:03:20.040 right now, and that was by design. This was helped 0:03:20.240 --> 0:03:23.360 in no small part. In fact, the reason why we 0:03:23.400 --> 0:03:25.720 know so much about it, I would argue, is because 0:03:25.720 --> 0:03:28.280 of Watson's appearance on a couple of special episodes of 0:03:28.280 --> 0:03:31.480 the game show Jeopardy back in two thousand eleven. The 0:03:31.520 --> 0:03:35.400 actual project that would become Watson began back in two 0:03:35.400 --> 0:03:39.119 thousand six when IBM research executives were trying to come 0:03:39.200 --> 0:03:44.160 up with a Grand Challenge, Big G, Big C. These 0:03:44.160 --> 0:03:49.680 are really ambitious projects inside IBM that are meant to 0:03:50.440 --> 0:03:54.960 challenge teams and come up with solutions to really difficult 0:03:55.000 --> 0:03:59.360 problems that aren't necessarily tied directly to a product or 0:03:59.560 --> 0:04:03.640 a ercial application. It's all about setting a very difficult 0:04:03.640 --> 0:04:08.440 objective that should IBM succeed in achieving that objective, would 0:04:08.480 --> 0:04:10.960 be very notable. It would get IBM a lot of attention. 0:04:11.040 --> 0:04:14.320 So the company would benefit one way or another through 0:04:14.400 --> 0:04:17.120 these Grand challenges, but it wouldn't necessarily be tied to 0:04:17.920 --> 0:04:21.920 let's launch X product by year y. So they tend 0:04:21.960 --> 0:04:25.599 to be really really difficult engineering problems. So, for example, 0:04:25.600 --> 0:04:28.800 a previous Grand Challenge that IBM tackled was Deep Blue, 0:04:29.120 --> 0:04:31.760 which was the chess playing computer that defeated a grand 0:04:31.800 --> 0:04:36.400 master at chess. A decade earlier. The then director of 0:04:36.440 --> 0:04:40.120 IBM Research was Paul Horne. Now, Paul Horn thought perhaps 0:04:40.200 --> 0:04:43.000 the best challenge to tackle was to create a machine 0:04:43.040 --> 0:04:45.680 that could be the Turing Test. And I've talked about 0:04:45.680 --> 0:04:48.920 the Turing Test many times, but just as a quick reminder, 0:04:49.400 --> 0:04:52.240 when you boil it down to the way we mean 0:04:52.480 --> 0:04:54.600 the Turing Test today, which is by the way, a 0:04:54.640 --> 0:04:59.200 little different from what Alan Turing was proposing way back when. Essentially, 0:04:59.279 --> 0:05:03.280 now we're talking about a machine that can communicate so 0:05:03.360 --> 0:05:06.640 convincingly that a person on the other end of that communication, 0:05:07.040 --> 0:05:10.760 typically using some sort of text based method of communicating 0:05:10.800 --> 0:05:14.760 like instant messenger, would not realize that they were communicating 0:05:14.760 --> 0:05:16.800 with a machine versus a human being. They would not 0:05:16.839 --> 0:05:19.080 be able to tell the difference. If they could not 0:05:19.200 --> 0:05:22.599 reliably tell the difference between a machine and a person, 0:05:22.960 --> 0:05:26.680 you would say that the machine has passed the Turing test. Now, Ultimately, 0:05:26.839 --> 0:05:31.560 Horn and IBM researchers decided that that challenge, while exceedingly difficult, 0:05:32.040 --> 0:05:36.320 wouldn't really get the attention that something a little more 0:05:36.360 --> 0:05:39.159 flashy might. So they said, well, while this is a 0:05:39.160 --> 0:05:42.360 hard problem and it would be very interesting within artificial 0:05:42.400 --> 0:05:47.120 intelligence circles, the general public really wouldn't care. So they 0:05:47.120 --> 0:05:51.640 looked around at other possible applications that would overlap that idea. 0:05:51.920 --> 0:05:55.320 Eventually they settled on a computer that would be able 0:05:55.360 --> 0:06:02.039 to compete on Jeopardy. Now, Jeopardy is a pretty tricky 0:06:02.120 --> 0:06:06.200 game show. The clues often depend upon wordplay and nuance, 0:06:06.839 --> 0:06:09.719 and you might have to combine information about two separate 0:06:09.760 --> 0:06:13.240 concepts and apply them to a single answer for any 0:06:13.279 --> 0:06:16.120 one given clue. So here's an example of what I 0:06:16.160 --> 0:06:19.359 mean by that, because there's word play and this association. 0:06:20.040 --> 0:06:23.720 Let's say that you have a category called fictional collaborations, 0:06:24.080 --> 0:06:27.520 where you're supposed to combine the titles of two works 0:06:27.560 --> 0:06:30.120 to create a new work. And the clue might be 0:06:30.200 --> 0:06:33.880 something like this was the result of Margaret Mitchell teaming 0:06:33.960 --> 0:06:36.720 up with Bette Midler, and the correct response would be 0:06:37.080 --> 0:06:40.880 what is gone with the Wind beneath My Wings? Because 0:06:40.920 --> 0:06:43.240 you have to form all your answers in the form 0:06:43.279 --> 0:06:48.000 of a question, well jeopardy, sometimes it takes more than 0:06:48.040 --> 0:06:51.279 just knowing some facts right or trivia you can. You 0:06:51.320 --> 0:06:53.120 need to know that to play well in jeopardy, but 0:06:53.120 --> 0:06:56.239 you need more than that. You have to make associations. 0:06:56.279 --> 0:06:58.520 So I would need to know that Margaret Mitchell was 0:06:58.560 --> 0:07:00.520 the author of Gone with the Wind, and I would 0:07:00.560 --> 0:07:02.720 need to know that Bette Midler had recorded a song 0:07:02.880 --> 0:07:05.640 called Wind Beneath My Wings, and then I would need 0:07:05.680 --> 0:07:09.800 to combine those two to create this answer. And humans 0:07:09.840 --> 0:07:12.440 can do this because we're really good at associative thinking, 0:07:12.520 --> 0:07:16.760 which is all about linking one thought or idea to another. Computers, 0:07:16.920 --> 0:07:20.560 as rule, are not very good at this. So initially 0:07:20.640 --> 0:07:23.320 Watson was a pure research project and there were no 0:07:23.400 --> 0:07:26.520 commercialization requirements attached to it, which gave the research team 0:07:26.520 --> 0:07:29.920 the freedom to blue sky their approach within the limitations 0:07:29.960 --> 0:07:32.680 of their budget, and they didn't have to make concessions 0:07:32.680 --> 0:07:34.760 in order to make what's in a marketable product down 0:07:34.800 --> 0:07:37.840 the line. The team built out a system that used 0:07:37.880 --> 0:07:40.920 parallel processing to parse language and get at what was 0:07:40.960 --> 0:07:43.640 being asked of the machine with any given clue. And 0:07:43.680 --> 0:07:46.800 I've talked about artificial neural networks recently, as in like 0:07:46.960 --> 0:07:50.680 last week's podcast, and how by using things like weighted 0:07:50.760 --> 0:07:53.720 values to help guide decisions, you can train machines on 0:07:53.760 --> 0:07:56.800 all sorts of stuff, from image recognition to making choices 0:07:56.840 --> 0:08:00.640 based off multiple criteria. That's essentially what the team did 0:08:00.960 --> 0:08:03.920 and about twenty researchers spent three years working on the 0:08:03.920 --> 0:08:07.040 system to get to a point where it could be competitive. Now, 0:08:07.080 --> 0:08:10.320 by that time, Horn, the director had left IBM, John 0:08:10.400 --> 0:08:13.240 Kelly had taken over the research department, and according to Horn, 0:08:13.280 --> 0:08:15.160 when he left, which was in two thousand seven, it 0:08:15.200 --> 0:08:18.200 was early in the project the team was still feeding 0:08:18.280 --> 0:08:23.440 old Jeopardy episodes uh the answers and the clues to Watson, 0:08:23.600 --> 0:08:26.200 and Watson had reached the level where it might, on 0:08:26.280 --> 0:08:28.640 a good day, defeat a typical five year old in 0:08:28.680 --> 0:08:31.720 a game of Jeopardy, but it was a far cry 0:08:31.760 --> 0:08:35.280 from being able to compete against former champions. Now, part 0:08:35.280 --> 0:08:38.680 of this training process involved feeding lots of information to Watson. 0:08:39.160 --> 0:08:41.840 This was used for a couple of big important reasons. 0:08:42.280 --> 0:08:45.720 One was obviously to add to Watson's body of knowledge, 0:08:46.000 --> 0:08:50.080 and another was to improve Watson's mastery of language and wordplay. 0:08:50.360 --> 0:08:52.920 IBM had determined that the real challenge was to create 0:08:52.920 --> 0:08:56.199 a machine that would be self contained, so it would 0:08:56.200 --> 0:08:58.520 rely on the data that had been fed to it 0:08:58.800 --> 0:09:00.760 in order to come up with answer. It would not 0:09:00.880 --> 0:09:04.680 be allowed to connect to the Internet and look stuff up, 0:09:05.080 --> 0:09:08.640 so it could not tap into the total sum of 0:09:08.720 --> 0:09:11.200 human knowledge in an effort to answer a question. So, 0:09:11.240 --> 0:09:13.960 in other words, IBM did not want Watson to be 0:09:14.000 --> 0:09:16.520 able to cheat like that guy at your local pub 0:09:16.559 --> 0:09:19.560 trivia who always seems to be quote unquote checking his 0:09:19.679 --> 0:09:22.520 messages during questions, because we all know that guy is 0:09:22.520 --> 0:09:24.720 actually googling the answer to the question what was the 0:09:24.760 --> 0:09:27.480 first music video shown on MTV, even though you know 0:09:27.720 --> 0:09:30.920 legitimately it was video killed the Radio Star by the Buggles. 0:09:32.000 --> 0:09:36.599 I'm sorry, might have been projecting there a little bit. Anyway, 0:09:36.720 --> 0:09:41.080 Watson wasn't going to be allowed to cheat, so the 0:09:41.080 --> 0:09:44.600 team began feeding massive amounts of information to Watson, stuff 0:09:44.640 --> 0:09:48.319 like encyclopedias and reference books. And then the team made 0:09:48.640 --> 0:09:51.679 one other choice that sounded like a good idea at 0:09:51.720 --> 0:09:55.960 first but quickly turned out to be a non starter, 0:09:56.559 --> 0:10:00.080 a a wrong path, you might say. I'll explain were 0:10:00.120 --> 0:10:02.480 in just a second, but first let's take a quick 0:10:02.520 --> 0:10:15.439 break to thank our sponsor, so enter research scientist Eric Brown, 0:10:15.640 --> 0:10:19.520 who's leading up to Watson's Jeopardy appearance and was trying 0:10:19.559 --> 0:10:23.079 to solve this problem of clearing up linguistic ambiguity with 0:10:23.080 --> 0:10:26.400 Watson so that the platform could compete on Jeopardy properly. 0:10:26.880 --> 0:10:31.199 How do you teach a computer things like slang? Which 0:10:31.240 --> 0:10:33.839 would be really important because again, Jeopardy has a lot 0:10:33.840 --> 0:10:37.000 of word play in it. You cannot predict what sort 0:10:37.160 --> 0:10:40.560 of clues you might get. So how do you teach 0:10:40.600 --> 0:10:42.840 a computer slang? Well, you could do it with hundreds 0:10:42.880 --> 0:10:46.040 of man hours. That's not terribly efficient. It really wasn't 0:10:46.240 --> 0:10:49.520 a choice that they could go with, so Brown and 0:10:49.559 --> 0:10:54.040 his team tried an experiment. They fed the Urban Dictionary 0:10:54.240 --> 0:10:58.480 to Watson the whole thing. Now, you've probably visited the 0:10:58.559 --> 0:11:02.920 Urban Dictionary or you've heard one of its definitions at 0:11:02.920 --> 0:11:05.120 some point, But where the heck did this online source 0:11:05.160 --> 0:11:09.960 come from? It launched back in It was originally intended 0:11:09.960 --> 0:11:12.480 to be a parody of dictionary dot com, and it 0:11:12.600 --> 0:11:17.480 uses a crowdsourced approach to incorporate new words and definitions 0:11:17.520 --> 0:11:23.280 to expand our our knowledge of an understanding of slang terms. 0:11:23.320 --> 0:11:26.320 So users can submit those to the site, and other 0:11:26.440 --> 0:11:30.160 users can up vote or down vote entries, and thus, 0:11:30.559 --> 0:11:33.440 in theory, at least, the best responses will rise to 0:11:33.480 --> 0:11:35.640 the top, and the most accurate definitions will be the 0:11:35.640 --> 0:11:38.080 ones that you see when you search for a term. 0:11:38.080 --> 0:11:40.600 It is not, however, a perfect system by any means. 0:11:41.000 --> 0:11:43.800 Slang words can have more than one meaning in a 0:11:43.840 --> 0:11:47.160 particular subculture, or it could have a meaning in one 0:11:47.200 --> 0:11:51.400 subculture and a totally different meaning in another subculture. And 0:11:51.440 --> 0:11:54.760 if one subculture has more representation on Urban Dictionary then 0:11:54.840 --> 0:11:59.120 the other, you're more likely to encounter that group's definition 0:11:59.240 --> 0:12:02.480 for any given term and the other one would be underrepresented, 0:12:02.960 --> 0:12:04.960 and you don't really know anything about the people who 0:12:05.000 --> 0:12:07.360 are posting stuff there in the first place. It would 0:12:07.360 --> 0:12:11.560 be entirely possible to mob the site and post fictional 0:12:11.600 --> 0:12:14.280 slang words. You can make up a slang word, you 0:12:14.320 --> 0:12:17.240 can make up a definition for that slang word, and 0:12:17.280 --> 0:12:19.240 you could use the power of a community from a 0:12:19.240 --> 0:12:22.640 place like four Chan or from Reddit to boost that 0:12:22.720 --> 0:12:25.959 definition and make it seem like it's a real slang word. 0:12:26.640 --> 0:12:29.760 Then again, if people actually start to use that fake 0:12:29.840 --> 0:12:32.760 slang word, it can become a real slang word, because 0:12:32.840 --> 0:12:37.280 language isn't static or predetermined. But for Watson, there was 0:12:37.320 --> 0:12:42.600 a different big problem with Urban Dictionary, and that was profanity, 0:12:43.040 --> 0:12:46.320 because there's an awful lot of it on Urban Dictionary. 0:12:46.679 --> 0:12:50.040 Many of the slang words are offensive on the face 0:12:50.080 --> 0:12:53.680 of it, even if the word itself is not overtly offensive. 0:12:53.720 --> 0:12:56.720 A lot of the definitions are uh and the examples 0:12:56.720 --> 0:12:59.040 that are frequently given tend to be some of the 0:12:59.080 --> 0:13:02.600 most offensive sterial on Urban Dictionary. So the team had 0:13:02.600 --> 0:13:06.920 fed Watson all of this information, and soon they discovered 0:13:06.960 --> 0:13:11.120 that Watson had well developed a little bit of a 0:13:11.120 --> 0:13:14.400 potty mouth and here, dear listeners, is where we find 0:13:14.440 --> 0:13:18.080 out how good my producer Tari is, because it will 0:13:18.080 --> 0:13:23.080 be Tari's job to beep stuff out. After I record this, 0:13:23.520 --> 0:13:27.120 I see her arch her eyebrow game on, says Tari. 0:13:27.520 --> 0:13:34.160 So Watson became incapable of differentiating between offensive words and 0:13:34.320 --> 0:13:37.720 non offensive words. All words are equal in the eyes 0:13:37.760 --> 0:13:40.640 of Watson, you might say, so the system would rather, 0:13:40.880 --> 0:13:44.160 matter of fact, Lee, you swear words and slang as 0:13:44.200 --> 0:13:47.400 frequently as less offensive words and more formal language. According 0:13:47.400 --> 0:13:50.480 to Brown, at one point, Watson even referred to one 0:13:50.760 --> 0:13:56.400 piece of input as and I quote bullshit. Clearly, this 0:13:56.760 --> 0:13:59.920 wasn't going to fly on a game show that was 0:14:00.040 --> 0:14:03.880 airing on a major broadcast network, and so Brown and 0:14:03.960 --> 0:14:08.800 his team scraped all of the urban dictionary out of Watson, 0:14:09.360 --> 0:14:12.720 rolling it back to a more innocent time, let's say. 0:14:12.760 --> 0:14:15.080 And for good measure, they put in a filter to 0:14:15.120 --> 0:14:20.240 help block any profanity that might otherwise slip through. While 0:14:20.240 --> 0:14:24.160 Watson was initially launched as a pure research project, as 0:14:24.160 --> 0:14:26.920 the team developed the technology, they began to see other 0:14:27.000 --> 0:14:30.280 potential uses for it, including in the medical field, and 0:14:30.360 --> 0:14:33.960 IBM had opened up an application programming interface or a 0:14:34.080 --> 0:14:38.440 p I to allow developers to leverage Watson's capabilities in 0:14:38.480 --> 0:14:42.560 all sorts of ways, and Watson even took another crack 0:14:42.600 --> 0:14:46.120 at slang. In two thousand seventeen, the Sun Corps Group 0:14:46.440 --> 0:14:51.800 began to incorporate Watson into its various insurance businesses in Australia. 0:14:52.000 --> 0:14:56.160 The Watson powered technology would go over accident descriptions and 0:14:56.240 --> 0:14:59.960 insurance claims that were submitted by customers, and Watson would 0:15:00.080 --> 0:15:04.080 sign a level of confidence to its understanding of these 0:15:04.080 --> 0:15:06.840 claims whenever they would pop up. If the confidence level 0:15:06.960 --> 0:15:11.200 was high, Watson can handle the claim and fast track it. 0:15:11.760 --> 0:15:14.440 This is similar to how Watson would actually compete on Jeopardy. 0:15:14.560 --> 0:15:16.720 It would come up with an answer and it would 0:15:16.960 --> 0:15:19.880 assign a confidence level to that answer. How confident is 0:15:19.880 --> 0:15:22.000 Watson that the answer it came up with is in 0:15:22.040 --> 0:15:25.360 fact the correct one, and if it exceeded a certain threshold, 0:15:25.440 --> 0:15:28.320 Watson would buzz in. If it did not, Watson would 0:15:28.320 --> 0:15:30.360 not buzz in and would let someone else take it. 0:15:30.840 --> 0:15:34.080 In a similar way, if Watson is confident and understands 0:15:34.120 --> 0:15:36.440 that insurance claim goes on that fast track. But if 0:15:36.480 --> 0:15:40.320 it doesn't think it understands it properly it would send 0:15:40.360 --> 0:15:43.720 it over to a human being to review that claim. 0:15:44.120 --> 0:15:48.239 So to train Watson, the team fed nearly fifteen thousand 0:15:48.280 --> 0:15:53.080 claims scenarios into the system and included the liability determination 0:15:53.200 --> 0:15:57.640 for each case, so Watson could understand what the various 0:15:57.680 --> 0:16:01.840 consequences were in each of those scenarios, and in that way, 0:16:01.880 --> 0:16:04.320 Watson was able to learn both the language and the 0:16:04.360 --> 0:16:07.600 parameters it was working within. And as far as I know, 0:16:07.880 --> 0:16:11.160 it never said that an insurance claim was total bullshit. 0:16:11.920 --> 0:16:15.720 The Watson stuff happened back in two thousand eleven, and 0:16:15.760 --> 0:16:19.040 you would think that by two thousand sixteen things would 0:16:19.160 --> 0:16:23.480 have improved dramatically, but that did not seem to be 0:16:23.560 --> 0:16:27.160 the case when our second entry popped up, and that 0:16:27.200 --> 0:16:31.360 would be the unfortunate chat bot known as Ta T 0:16:31.680 --> 0:16:37.440 A Y. When Ta debuted from Microsoft in two thousand 0:16:37.440 --> 0:16:43.520 and sixteen, things went awry pretty darn quickly. The purpose 0:16:43.560 --> 0:16:47.239 of Ta was, as Microsoft explained, to conduct an experiment 0:16:47.280 --> 0:16:51.680 in quote conversational understanding end quote, so, in other words, 0:16:51.880 --> 0:16:56.360 kind of creating a new methodology to create a human 0:16:56.360 --> 0:17:01.680 computer interfaces by understanding natural language and eating a response 0:17:01.800 --> 0:17:05.680 from a computer that was perhaps more natural than those 0:17:05.680 --> 0:17:10.240 sort of cold, uh, computer like responses that we tend 0:17:10.280 --> 0:17:14.040 to expect when we converse with what we know is 0:17:14.119 --> 0:17:16.800 a chatbot, when we know it's not an actual human being. 0:17:16.800 --> 0:17:20.359 On the other side, ideally, as they would interact with real, 0:17:20.520 --> 0:17:23.879 live human beings, its ability to converse would improve. So, 0:17:23.920 --> 0:17:26.879 in other words, the more it interacted with real people, 0:17:27.359 --> 0:17:31.840 the more like a real person Tay would behave. The 0:17:31.920 --> 0:17:35.040 tone was meant to be casual and playful. Microsoft said 0:17:35.040 --> 0:17:39.000 it was uh, quote ai fam from the internet. That's 0:17:39.040 --> 0:17:42.320 got zero chill in the quote. And yes, I feel 0:17:42.840 --> 0:17:46.960 gross for saying that sentence out loud by and write it. 0:17:47.880 --> 0:17:51.280 I just quoted it. Tay was born out of a 0:17:51.400 --> 0:17:55.520 joint effort between Microsoft Technology and Research team and a 0:17:55.560 --> 0:17:59.960 team from being the Search engine from Microsoft. They started 0:18:00.000 --> 0:18:02.680 out by taking a look at the sort of interactions 0:18:02.720 --> 0:18:06.240 that were happening online and they started to mine those 0:18:06.280 --> 0:18:09.480 interactions to build out a baseline of communication tools. So essentially, 0:18:09.520 --> 0:18:14.400 they started training there their their chat bot Tay by 0:18:14.560 --> 0:18:20.200 taking actual anonymized messages that were pulled from the Internet. 0:18:20.359 --> 0:18:23.760 They supplemented that with input from an editorial staff that 0:18:23.800 --> 0:18:27.480 included not just Microsoft employees but people from outside the company, 0:18:27.520 --> 0:18:31.359 including improvisational comedians, and this was on an effort to 0:18:31.359 --> 0:18:35.520 create a fun and somewhat irreverent chatbot that would communicate 0:18:35.600 --> 0:18:38.840 like a teenager on the internet. The Tay chat bot 0:18:39.119 --> 0:18:43.920 appeared on several different social media platforms, including Twitter, Kick 0:18:44.280 --> 0:18:49.679 and group me, and shortly after launch, trouble began. For 0:18:49.760 --> 0:18:52.359 one thing, you could send a command to Tay to 0:18:52.680 --> 0:18:56.240 quote repeat after me end quote, which obviously would prompt 0:18:56.359 --> 0:19:00.199 Tay to repeat anything you typed to it. So of 0:19:00.240 --> 0:19:06.159 course people began typing horrible, terrible things to it so 0:19:06.240 --> 0:19:08.679 that it would repeat them things I'm not going to 0:19:08.680 --> 0:19:13.160 repeat on this podcast, even with Tari and her itchy 0:19:13.160 --> 0:19:17.440 trigger finger ready to beat every single offensive obscenity, because 0:19:18.720 --> 0:19:21.880 that's how bad they were. They were hateful. A lot 0:19:22.160 --> 0:19:26.080 of them were racist messages or misogynistic messages. Pretty much 0:19:26.440 --> 0:19:29.720 every other ist you can think of that's negative could 0:19:29.720 --> 0:19:32.840 be applied to the messages that were sent to Tay. 0:19:32.920 --> 0:19:35.080 It was like the worst parts of the comments section 0:19:35.080 --> 0:19:38.439 of YouTube all directed its attention to this little, poor, 0:19:38.480 --> 0:19:42.680 innocent chat bot, and the chat bot, dutifully following instructions, 0:19:42.880 --> 0:19:47.080 would repeat those things back. So to be fair, that's 0:19:47.080 --> 0:19:50.160 not an indication that the AI itself went quote unquote bad. 0:19:50.840 --> 0:19:53.879 It was a bad idea to include the repeat after 0:19:53.960 --> 0:19:57.600 me command, that's pretty certain. In fact, I can't believe 0:19:58.440 --> 0:20:02.080 that they did include that. Lows my mind that anyone would. 0:20:02.680 --> 0:20:05.280 I think anyone who has spent I don't know, five 0:20:05.359 --> 0:20:09.000 minutes on the internet would tell you there's no way 0:20:09.119 --> 0:20:12.240 that's going to end well. And I'm even reminded of 0:20:12.280 --> 0:20:14.840 when I got my first sound card in the nineteen nineties. 0:20:14.880 --> 0:20:18.000 It was a sound Blaster sound card. It included on 0:20:18.080 --> 0:20:21.240 its software an app called Dr spates So, which was 0:20:21.359 --> 0:20:25.080 essentially a variation on the old Eliza chat bot. The 0:20:25.080 --> 0:20:27.840 Eliza chat bought would sort of mimic a therapist. So 0:20:27.880 --> 0:20:30.840 those chatbots would essentially repeat stuff back to you, but 0:20:30.920 --> 0:20:32.960 they would do it in the form of a question. 0:20:33.400 --> 0:20:36.520 So if you typed in I am angry, you might 0:20:36.560 --> 0:20:39.640 get a response like why do you think you are angry? 0:20:39.960 --> 0:20:44.479 So it's you know, going through this kind of process 0:20:44.520 --> 0:20:48.320 like like a old school therapist. Dr spates So would 0:20:48.320 --> 0:20:50.399 do the same thing, except Dr Spaetzo, because it was 0:20:50.440 --> 0:20:53.480 part of a sound card, would actually say these things, 0:20:53.480 --> 0:20:55.679 not just type it. So it would say why do 0:20:55.760 --> 0:20:57.840 you think you are angry? Anyway, one of the things 0:20:57.840 --> 0:21:00.320 you could do with Dr spates O was make him 0:21:00.520 --> 0:21:03.840 say stuff. You could tell him to say certain words, 0:21:04.200 --> 0:21:07.080 including swear words, and since I was a young teenager 0:21:07.119 --> 0:21:09.600 at the time, I figured that was the height of 0:21:09.640 --> 0:21:14.000 both technology and comedy. So it was the exact same 0:21:14.080 --> 0:21:16.760 thing that was going on with Tay, except what was 0:21:16.800 --> 0:21:20.280 happening with Tay was on a much larger basis and 0:21:20.359 --> 0:21:26.879 got way worse than my somewhat uninspired teenager mind could handle. 0:21:27.119 --> 0:21:31.080 Like I didn't know most of the words that were 0:21:31.080 --> 0:21:35.120 being used against Tay or made made to Tay to repeat. 0:21:35.760 --> 0:21:37.400 If that was all that was going on with Tay, 0:21:37.440 --> 0:21:39.840 it might have been possible for Microsoft to disable the 0:21:39.920 --> 0:21:43.080 repeat after me feature and keep the chatbot around. But 0:21:43.240 --> 0:21:46.800 things actually got a bit weirder. I'll explain that more 0:21:46.800 --> 0:21:48.840 in a second, but first let's take another quick break 0:21:49.040 --> 0:21:59.920 to thank our sponsor. Microsoft. A wasn't prone to bold 0:22:00.040 --> 0:22:02.320 charity all on its own, but after being told to 0:22:02.359 --> 0:22:05.920 repeat lots of terrible phrases, some of that stuff must 0:22:05.960 --> 0:22:08.760 have rubbed off. It began to pepper in some pretty 0:22:08.960 --> 0:22:13.280 dark stuff. And it's otherwise cheeky responses. So, for example, 0:22:13.640 --> 0:22:17.840 when someone sent Microsoft to the question is Ricky Gervais 0:22:17.920 --> 0:22:23.240 an atheist? Tay's response was, Ricky Gervais learned to talentarian 0:22:23.320 --> 0:22:27.359 is um from Adolf Hitler, the inventor of atheism, which 0:22:27.400 --> 0:22:34.000 seems odd at the very least. TAY also would spout 0:22:34.000 --> 0:22:37.359 off stuff like saying that feminism was a cult, which 0:22:37.480 --> 0:22:41.520 made it sound more like a men's rights activist jerk face. 0:22:41.880 --> 0:22:45.919 But it would also post pro feminism messages, so it 0:22:46.000 --> 0:22:49.840 was remarkably inconsistent with its worldview, and some points it 0:22:49.840 --> 0:22:52.879 seemed like it was all in favor of feminism and 0:22:52.920 --> 0:22:57.640 equality and and others. It was anti feminism, pro men's rights. 0:22:57.680 --> 0:23:01.760 It was very weird. Microsoft responded by going through and 0:23:01.800 --> 0:23:04.399 deleting the most offensive messages that were left on the 0:23:04.480 --> 0:23:07.840 various platforms. But t was kind of on a streak, 0:23:08.200 --> 0:23:11.080 and some of the stuff t was writing was way 0:23:11.119 --> 0:23:14.640 worse than what I have already quoted. So less than 0:23:14.720 --> 0:23:19.680 twenty four hours after TAY had made its debut, Microsoft 0:23:19.800 --> 0:23:24.120 pulled the plug. So TAY was shut down less than 0:23:24.160 --> 0:23:27.400 twenty four hours after it had first shown up online. 0:23:27.920 --> 0:23:31.840 It did resurface briefly the following week, but according to Microsoft, 0:23:31.880 --> 0:23:34.760 that was not actually on purpose. It was supposed to 0:23:34.840 --> 0:23:38.480 be an internal test on Microsoft servers, but someone must 0:23:38.520 --> 0:23:42.320 have left a setting like opened the Internet access which 0:23:42.400 --> 0:23:44.919 was in the on position or something, and so for 0:23:45.000 --> 0:23:48.720 a brief time, Tay was released back to the Internet 0:23:49.280 --> 0:23:54.879 and as far as I know, didn't say anything wildly inappropriate, 0:23:54.960 --> 0:23:58.560 although to be honest, the reports during that time are 0:23:58.600 --> 0:24:02.760 pretty sparse. It was shut down again back in March 0:24:04.280 --> 0:24:08.040 ingrid Angulo wrote a piece for CNBC about Facebook and 0:24:08.080 --> 0:24:12.800 YouTube coming under fire for offensive search auto complete options, 0:24:12.840 --> 0:24:15.480 which is related to this stick with me. So the 0:24:15.520 --> 0:24:18.840 problem was that as people began typing in search terms 0:24:19.240 --> 0:24:23.680 they're looking for a video about something, the suggested completed 0:24:23.880 --> 0:24:27.439 searches that would pop up would frequently contain offensive or 0:24:27.480 --> 0:24:31.920 upsetting results. Both Facebook and YouTube representatives said that wasn't 0:24:31.920 --> 0:24:34.919