WEBVTT - The Algorithm 0:00:01.120 --> 0:00:04.240 The views and opinions expressed in this podcast are solely 0:00:04.240 --> 0:00:07.560 those of the authors and participants and do not necessarily 0:00:07.600 --> 0:00:11.520 represent those of iHeart Media, Tenderfoot TV, or their employees. 0:00:12.320 --> 0:00:16.720 This series contains discussions of violence and sexual violence. Listener 0:00:16.720 --> 0:00:24.800 discretion is advised. Last time an algorithm. Lorie Townsend more 0:00:24.840 --> 0:00:27.800 in the death of her daughter, Africa Hardy, don't think 0:00:27.880 --> 0:00:30.960 that you know everything about child, because there's something that 0:00:31.000 --> 0:00:34.280 they're not telling you. After moving to Chicago when she 0:00:34.440 --> 0:00:41.599 was nineteen, Africa started escorting. In October, detectives found her 0:00:42.200 --> 0:00:46.320 strangled in the bathtub of an Indiana motel. I just 0:00:46.360 --> 0:00:49.839 think a lot of it, and I think it could 0:00:49.880 --> 0:00:54.840 have been prevented. Years earlier, journalist Thomas Hargrove had learned 0:00:54.840 --> 0:01:01.040 about the concept of linkage blindness. Most connected murders go unrecognized, 0:01:01.360 --> 0:01:04.160 and I kept that in the back of my mind. 0:01:05.000 --> 0:01:09.640 And then when Hargrove discovered an FBI database that tracked homicides, 0:01:09.959 --> 0:01:13.399 it gave him an idea, could we teach a computer 0:01:13.720 --> 0:01:20.800 to identify connected cases to find serial killings? From my 0:01:20.959 --> 0:01:26.120 Heart Media and Tenderfoot TV, this is algorithm I'm ben 0:01:26.200 --> 0:01:35.479 Keebrick I mentioned in the last episode that Africa's case 0:01:35.720 --> 0:01:38.880 flipped what I knew about crime on its head this 0:01:39.040 --> 0:01:44.120 episode you'll see why. Let's jump back to my conversation 0:01:44.360 --> 0:01:48.480 with Thomas Hargrove. So, at what point in my um 0:01:48.920 --> 0:01:51.360 you were saying that we shouldn't ignore her. Grove told 0:01:51.360 --> 0:01:53.920 me that back in two thousand four, when he was 0:01:53.960 --> 0:01:57.920 working as an investigative reporter in Washington, d C. He 0:01:57.960 --> 0:02:01.960 had come across a database with information about homicides all 0:02:02.000 --> 0:02:06.320 across the US. He wondered whether an algorithm could find 0:02:06.440 --> 0:02:11.920 patterns within that data and to text serial killers. Maybe 0:02:12.000 --> 0:02:15.160 it could even be used to detect active serial killers 0:02:15.200 --> 0:02:18.720 who had not yet been caught. I did not know 0:02:19.200 --> 0:02:23.400 that such an algorithm was possible. I was going on faith. 0:02:24.120 --> 0:02:26.280 I did believe that it was important, and I did 0:02:26.400 --> 0:02:28.880 believe that it could save lives, but I didn't know 0:02:28.960 --> 0:02:32.080 that it would work. I begged my editors, let me 0:02:32.120 --> 0:02:35.680 try this. Let's do a project looking at murder, with 0:02:35.840 --> 0:02:38.680 the understanding that the actual goal is to see if 0:02:38.720 --> 0:02:42.160 we could create a computer program that would identify serial murder. 0:02:42.760 --> 0:02:46.239 But remember this was back in two thousand four, years 0:02:46.320 --> 0:02:49.720 before people were talking about things like machine learning or 0:02:49.960 --> 0:02:53.200 the power of big data. Because Harder was on the 0:02:53.240 --> 0:02:56.440 cutting edge of these ideas, it was sometimes difficult to 0:02:56.520 --> 0:02:59.920 get other people to understand him or to take him serious. 0:03:00.120 --> 0:03:05.240 Lee My editors recognized that this could be something very cool, 0:03:05.919 --> 0:03:09.400 that this could make news, but it's hard to commit 0:03:09.720 --> 0:03:14.320 to a project you don't know upfront whether it's possible. Ultimately, 0:03:14.560 --> 0:03:17.839 his editors told him that trying to design an algorithm 0:03:18.000 --> 0:03:23.360 sounded too risky. It's normal for story pitches, especially ambitious 0:03:23.360 --> 0:03:27.240 ones like this, to get rejected, but what's unusual is 0:03:27.280 --> 0:03:31.280 that Hargrove didn't give up For the next six years. 0:03:31.360 --> 0:03:35.560 He kept pitching the story six years, and he says 0:03:35.640 --> 0:03:38.520 that his editors always considered it but would end up 0:03:38.560 --> 0:03:44.040 assigning him to another safer story. Then, in two thousand ten, 0:03:44.560 --> 0:03:48.000 Hargrove got his big break. He had just published a 0:03:48.040 --> 0:03:55.160 provocative piece called Saving Babies, exposing sudden infant death. Hargrove 0:03:55.280 --> 0:03:59.080 had analyzed data from Corners offices all across the US 0:03:59.760 --> 0:04:03.080 and found that many times when infant deaths were listed 0:04:03.080 --> 0:04:07.800 as sid's sudden infant death syndrome, the evidence actually showed 0:04:07.920 --> 0:04:12.480 that the infants had died of accidental cufifocations. Most of 0:04:12.520 --> 0:04:17.680 the time, babies don't die mysteriously. They die from avoidable 0:04:17.800 --> 0:04:23.080 accidents and unsafe sleeping conditions. They got covered up because 0:04:23.120 --> 0:04:26.440 of what was intended to be a very kindly diagnosis 0:04:26.520 --> 0:04:30.080 called SIDS. That was meant to be a kindness to parents. 0:04:30.800 --> 0:04:35.760 There was a mistake. This project prompted a national conference 0:04:35.839 --> 0:04:40.160 to re examine infant death. It caused the creation of 0:04:40.200 --> 0:04:44.720 a sudden Infant Death monitoring system by the CDC. I mean, 0:04:44.760 --> 0:04:49.600 it was tremendously successful. It had got rave attention. Finally, 0:04:49.640 --> 0:04:52.840 after six years, my stock was high enough in the 0:04:52.880 --> 0:04:55.640 newsroom that I was able to get them to look 0:04:55.680 --> 0:05:02.479 at murder. My editors that, okay, Tom, you've got a year. 0:05:04.120 --> 0:05:07.960 I could do a national reporting project looking at unsolved murder, 0:05:08.440 --> 0:05:11.760 with the understanding that what we were really about was 0:05:11.800 --> 0:05:16.320 to try to develop a statistical means to identify overlooked 0:05:16.520 --> 0:05:22.720 serial murder. Could we teach a computer to find serial killings? 0:05:25.040 --> 0:05:29.800 After years of dreaming about making this algorithm, now Hargrove 0:05:29.839 --> 0:05:33.960 had this chance to try and make it work. Hargrove 0:05:34.160 --> 0:05:38.040 starting to design a computer program, one that could comb 0:05:38.040 --> 0:05:42.320 through the five hundred thousand supplemental homicide reports and find 0:05:42.400 --> 0:05:46.919 patterns between victims, patterns that might suggest the work of 0:05:46.920 --> 0:05:53.360 a serial killer. So what is an algorithm? These days? 0:05:53.480 --> 0:05:57.320 Algorithms dictate much of our digital lives. They determine what 0:05:57.400 --> 0:06:02.040 TV shows get suggested to us, post we see on Facebook, 0:06:02.640 --> 0:06:07.480 what route our GPS takes us on. Often these algorithms 0:06:07.480 --> 0:06:10.360 are black boxes. We're not sure what data is being 0:06:10.400 --> 0:06:13.120 fed into them and how they're deciding what to spit 0:06:13.160 --> 0:06:18.920 back out. But at their core, algorithms aren't mysterious. They're 0:06:18.920 --> 0:06:22.680 not even really that complicated. Algorithms are just sets of 0:06:22.680 --> 0:06:27.120 instructions used to accomplish a specific task. You could say 0:06:27.279 --> 0:06:30.119 a cake recipe is an algorithm for baking a cake. 0:06:30.839 --> 0:06:34.080 You wouldn't say that, but you could. But while following 0:06:34.080 --> 0:06:37.760 an algorithm might be simple, designing a whole new algorithm 0:06:37.800 --> 0:06:40.839 is trickier. It wasn't like Cargroove was just baking a 0:06:40.920 --> 0:06:43.240 cake from scratch. It was more like he was trying 0:06:43.279 --> 0:06:47.960 to invent a brand new dessert. What we were doing 0:06:48.240 --> 0:06:52.080 with the algorithm was the process literally of trial and error. 0:06:52.560 --> 0:06:57.840 The ability to experiment with data was critical. If you're 0:06:57.839 --> 0:07:00.719 trying to come up with a new dessert, might experiment 0:07:00.800 --> 0:07:03.640 with ingredients until you come up with a combination that 0:07:03.680 --> 0:07:07.440 tastes good. Hargrove knew what ingredients he had to play with. 0:07:08.120 --> 0:07:11.559 That was the data from the supplemental homicide report, which 0:07:11.640 --> 0:07:14.560 listed the weapon that was used he was shot with 0:07:14.600 --> 0:07:17.880 a handgun, the time and location of the murder January, 0:07:18.840 --> 0:07:21.840 and the age, race, and sex of the victim. Victim 0:07:21.920 --> 0:07:25.520 is a blackmail eighteen years old. But as he experimented 0:07:25.600 --> 0:07:29.440 with how an algorithm might sort through that data, Hargrove 0:07:29.600 --> 0:07:31.560 needed a way to check if he was on the 0:07:31.640 --> 0:07:35.240 right track. So we decided to test each new prototype 0:07:35.240 --> 0:07:37.760 of his algorithm to see if it could detect a 0:07:37.840 --> 0:07:43.080 known serial killer. We used as our test bed the 0:07:43.240 --> 0:07:49.040 forty eight victims of serial killer Gary Ridgeway, so called 0:07:49.040 --> 0:07:53.160 Green River Killer. He was convicted convicted in a court 0:07:53.160 --> 0:07:56.840 of law of murdering forty eight girls and women in 0:07:56.880 --> 0:08:01.360 the Seattle area. The question was, we'd teach a computer 0:08:01.960 --> 0:08:05.200 a process that would tell us that something god awful 0:08:05.320 --> 0:08:09.560 happened in Seattle. At the time, Gary Ridgeway was thought 0:08:09.640 --> 0:08:13.760 to be the most prolific serial killer in America. Her 0:08:13.840 --> 0:08:16.480 group figured if he could create an algorithm that could 0:08:16.520 --> 0:08:20.960 detect Ridgeway. Maybe it would also detect other serial killers 0:08:21.080 --> 0:08:26.160 that had gone unrecognized. To motivate himself, Hargrove stuck up 0:08:26.160 --> 0:08:29.960 a picture of Gary Ridgeway in his office. It was 0:08:30.200 --> 0:08:33.560 one of his booking photographs. He was glowering and looked 0:08:33.600 --> 0:08:37.760 like it was glowering at me. Under that picture, I 0:08:37.880 --> 0:08:43.959 typed the headline, what do serial victims look like? Statistically? 0:08:45.160 --> 0:08:48.240 For over two decades, Ridgeway had killed women and dumped 0:08:48.240 --> 0:08:52.720 their bodies along the Green River outside Seattle. In most cases, 0:08:52.840 --> 0:08:56.080 by the time anyone had discovered those corpses, they were 0:08:56.120 --> 0:09:00.920 already skeletal. Ridgeway has snuffed out does sense of lives. 0:09:01.480 --> 0:09:05.160 He'd killed women and girls with their own hopes and dreams. 0:09:06.000 --> 0:09:09.600 But Hargrove needed to approach their deaths like the algorithm would, 0:09:10.080 --> 0:09:13.480 looking at their cases by their numbers, using only the 0:09:13.600 --> 0:09:18.960 data from the FBI's supplemental homicide reports. So what did 0:09:19.040 --> 0:09:23.920 ridgeways victims look like? Statistically? And more broadly, what do 0:09:24.000 --> 0:09:28.240 a serial killer's victims look like? To find out, I 0:09:28.320 --> 0:09:31.760 reached out to one of Hargrove's collaborators, Dr Mike A. Mott. 0:09:32.520 --> 0:09:37.079 A Mott is a forensic psychology professor at Radford University, 0:09:37.240 --> 0:09:40.920 and he curates the world's largest database of serial killers 0:09:40.960 --> 0:09:43.640 and their victims. If you look at how serial killers 0:09:43.640 --> 0:09:47.720 are portrayed in movies or on TV, the stereotypes not 0:09:47.840 --> 0:09:52.480 really very consistent with with the back Are these also 0:09:52.640 --> 0:09:56.000 stereotypes that law enforcement might have when they're trying to 0:09:56.040 --> 0:10:01.120 decide could this murder possibly be a serial murder? Well, 0:10:01.240 --> 0:10:04.800 law enforcement certainly has the same stereotypes. When we've done 0:10:04.840 --> 0:10:09.480 presentations to law enforcement groups and even clinical psychologists that 0:10:09.520 --> 0:10:27.920 are police psychologists, they're very surprised at the results. I'm 0:10:27.960 --> 0:10:31.640 talking to Professor Mike al Mott about what serial murders 0:10:31.679 --> 0:10:36.760 look like statistically, because statistical differences between serial killings and 0:10:36.840 --> 0:10:40.600 other homicides are the kind of signatures and algorithm could 0:10:40.679 --> 0:10:44.720 use to detect a serial killer. Professor Amatt has built 0:10:44.760 --> 0:10:49.400 a database with information on thousands of serial killers, and 0:10:49.400 --> 0:10:52.920 he told me that there are many misconceptions about these killers, 0:10:52.960 --> 0:10:56.880 even within law enforcement and the true crime community. The 0:10:56.960 --> 0:10:59.840 stereotypes about the profile of a serial killer or pro 0:11:00.040 --> 0:11:03.320 file of the victims is not really very consistent with 0:11:03.360 --> 0:11:06.800 the facts. If we want to catch serial killers, we 0:11:06.880 --> 0:11:09.560 need to know who they really are in the reality 0:11:09.600 --> 0:11:12.360 of their crimes. I want you to stop for a 0:11:12.400 --> 0:11:15.440 moment and try to conjure up a mental image of 0:11:15.440 --> 0:11:19.760 a typical serial killers victim. How old are they, what 0:11:19.840 --> 0:11:24.319 do they look like? Now, imagine all of the victims 0:11:24.360 --> 0:11:27.040 of a single killer. What would you guess they have 0:11:27.160 --> 0:11:30.960 in common with one another? If we look at the 0:11:31.000 --> 0:11:36.280 victims of serial killers, of the victims or women or 0:11:36.480 --> 0:11:39.760 forty nine or men, so there's not really much of 0:11:39.760 --> 0:11:43.760 a difference there. One of the stereotypes about serial killers 0:11:43.800 --> 0:11:45.520 is that they have a type, you know, kinde of 0:11:45.600 --> 0:11:48.600 victim that they're going to kill. The most consistent profile 0:11:48.640 --> 0:11:50.640 we can have is really in the age category of 0:11:50.679 --> 0:11:54.920 who they're killing. About are going to kill people that 0:11:54.960 --> 0:11:57.880 are in the same age categories, so it's children, or 0:11:57.880 --> 0:12:02.559 its teens, or its adults or elderly. Serial Killers are 0:12:02.640 --> 0:12:06.280 less consistent when it comes to other attributes of their victims. 0:12:06.360 --> 0:12:10.280 For example, only six kill victims there are all male 0:12:10.600 --> 0:12:13.400 or all female. And then if you look at race, 0:12:13.480 --> 0:12:17.440 for example, of serial killers only kill somebody of the 0:12:17.520 --> 0:12:21.240 same race. And while the stereotypical victim of a serial 0:12:21.320 --> 0:12:24.439 killer is white, actually a third of victims or people 0:12:24.440 --> 0:12:29.440 of color. For Gary Ridgeway, the serial killer that Hargrove 0:12:29.480 --> 0:12:32.640 picked as his test case, all his known victims were 0:12:32.679 --> 0:12:35.839 young women, but they were diverse in terms of race. 0:12:36.640 --> 0:12:40.080 In an interview with an FBI agent, Ridgeway says he 0:12:40.160 --> 0:12:44.600 targeted prostitutes out of convenience, and studies show that there's 0:12:44.600 --> 0:12:49.000 actually been a dramatic rise in serial killers targeting sex workers. 0:12:49.760 --> 0:12:53.000 In the seventies, prostitutes are thought to have made up 0:12:53.040 --> 0:12:57.719 around six of the female victims of serial killers. By 0:12:57.720 --> 0:13:01.120 the two thousands, more than two thirds of women murdered 0:13:01.160 --> 0:13:05.800 by serial killers were sex workers. Many serial killings are 0:13:05.960 --> 0:13:09.679 sexual in nature. About one third of serial killers rape 0:13:09.720 --> 0:13:14.240 at least one of their victims. Ridgeway told detectives he'd 0:13:14.280 --> 0:13:17.560 have sex with his victims before he strangled them, and 0:13:17.600 --> 0:13:20.520 when asked why he chose to choke all his victims, 0:13:20.720 --> 0:13:24.640 Ridgeway replied because that was more personal and more rewarding 0:13:24.880 --> 0:13:30.079 than shooting them. And compared to typical murderers, serial killers 0:13:30.160 --> 0:13:32.760 are more likely to use methods that are up close 0:13:32.840 --> 0:13:38.280 and personal, like strangulation or bludgending. So if we're looking 0:13:38.360 --> 0:13:44.840 at victims in the US about were shot, were strangled 0:13:46.040 --> 0:13:50.640 were stabbed and ten percent were blugend and serial killers 0:13:50.760 --> 0:13:53.560 tend to be consistent in the method they used to kill. 0:13:54.240 --> 0:13:57.440 Amont says that two thirds of serial killers use only 0:13:57.480 --> 0:14:01.320 a single means to kill their victims. So what are 0:14:01.360 --> 0:14:05.600 the takeaways from a MOOTS data? Serial killers show some 0:14:05.920 --> 0:14:10.520 but not perfect consistency in terms of their victims age, race, 0:14:10.679 --> 0:14:15.320 and sex, and also in their method of killing. But 0:14:15.400 --> 0:14:18.240 it turns out that there's one more thing that serial 0:14:18.320 --> 0:14:22.120 homicides tend to have in common, a property that would 0:14:22.160 --> 0:14:26.360 be crucial for hard Groves algorithm. We spent the summer 0:14:26.400 --> 0:14:29.720 of two thousand and ten finding at least the hundred 0:14:29.840 --> 0:14:33.680 procedures that crashed and burned. Does the presence of a 0:14:33.760 --> 0:14:37.360 serial killer increase the rated which women are murdered? No? 0:14:38.200 --> 0:14:41.000 Does it increase the rate of which women are murdered 0:14:41.040 --> 0:14:46.400 through unusual means? No? And do you get any indications 0:14:46.440 --> 0:14:51.320 that you're getting closer? Yeah, So as we were progressing 0:14:51.560 --> 0:14:54.560 through the hundred and one things that don't work, we 0:14:54.560 --> 0:14:57.960 were starting to get closer. So um, the last thing 0:14:58.200 --> 0:15:02.520 was to look at what the clearance rate was for 0:15:02.840 --> 0:15:07.960 particular types of weapons. That term clearance rate refers to 0:15:08.000 --> 0:15:11.720 the percentage of cases that police end up arresting someone 0:15:11.800 --> 0:15:16.160 for the crime. Hard Group realized that Ridgeway had gotten 0:15:16.160 --> 0:15:19.040 away with his murders for so long that they've been 0:15:19.080 --> 0:15:22.840 listed as unsolved in the database. So if you looked 0:15:22.880 --> 0:15:26.640 at killings in Seattle that matched his method, the number 0:15:26.640 --> 0:15:30.240 of cases police had cleared was much lower than expected. 0:15:31.280 --> 0:15:36.360 The presence of an active serial killer often destroys the 0:15:36.440 --> 0:15:39.680 batting average for the local police department. They're able to 0:15:39.680 --> 0:15:43.560 solve most murders, but not that type of murder because 0:15:43.560 --> 0:15:48.000 there's a serial killer who's avoiding arrest hert group was 0:15:48.040 --> 0:15:51.400 making progress. He'd picked up the hint of a signal 0:15:51.480 --> 0:15:54.400 from Bridgeway in Seattle, but it wasn't enough to make 0:15:54.440 --> 0:15:58.280 the algorithm useful as a tool. He racked his brain 0:15:58.520 --> 0:16:00.800 trying to come up with a way to improve it. 0:16:01.960 --> 0:16:04.680 One day, near the end of the summer, he looked 0:16:04.760 --> 0:16:08.440 up at Ridgeway's mug shot and asked himself again, what 0:16:08.480 --> 0:16:13.160 do a serial killers victims look like? Statistically? As he 0:16:13.200 --> 0:16:16.720 talked through the project one last time with his research assistant, 0:16:16.800 --> 0:16:20.920 Liz Lucas, an idea struck him. As I was taking 0:16:21.040 --> 0:16:23.360 Liz to the airport because she had to go home 0:16:23.400 --> 0:16:26.360 to defend her master's thesis. I told her that what 0:16:26.560 --> 0:16:29.520 might work, and we're gonna try this next is a 0:16:29.600 --> 0:16:34.600 kind of cluster analysis. Instead of querying all of the data, 0:16:35.280 --> 0:16:40.240 we tried to assemble the data into smaller clusters according 0:16:40.280 --> 0:16:43.720 to the county where the murderer occurred, the gender of 0:16:43.760 --> 0:16:46.560 the victim, the method of killing, and at that time 0:16:46.600 --> 0:16:51.120 age group. Then we calculated what the clearance rate was 0:16:51.720 --> 0:16:55.640 for each cluster, how many of those murders were solved. 0:16:56.520 --> 0:16:59.200 Up until this point, he tried looking at the data 0:16:59.360 --> 0:17:02.840 for all the orders in Seattle. His new idea was 0:17:02.920 --> 0:17:06.679 to further break down the data. Have the algorithm split 0:17:06.800 --> 0:17:10.760 up the homicides for each county into buckets of victims, 0:17:10.880 --> 0:17:13.720 and then rank these different buckets by the percentage of 0:17:13.760 --> 0:17:18.120 cases that were solved. And Hargrove wasn't just doing this 0:17:18.200 --> 0:17:21.000 for Seattle. He was doing this for data all across 0:17:21.040 --> 0:17:23.520 the US, and he was hoping that out of the 0:17:23.560 --> 0:17:27.920 thousands of clusters, Ridgeways Seattle killings would show up somewhere 0:17:27.960 --> 0:17:31.800 near the top of the list. Hargrove ran his new 0:17:31.840 --> 0:17:37.520 algorithm and waited anxiously for the result. When we did that, 0:17:38.000 --> 0:17:42.280 the Green River killings jumped out plane as Day came 0:17:42.320 --> 0:17:46.600 in third place most of the time seventies seven percent 0:17:46.680 --> 0:17:48.920 of the time, and arrest is made when a woman 0:17:49.040 --> 0:17:52.920 is killed in the cluster where the Green River killings 0:17:52.960 --> 0:17:58.359 were grouped. The solution rate was less than and that 0:17:58.640 --> 0:18:03.119 was our key. We're looking for groups of similar murders 0:18:03.160 --> 0:18:08.199 that have very low clearance rates. We had hundreds of 0:18:08.240 --> 0:18:12.760 results all over the country, highly suspicious clusters, and we 0:18:12.800 --> 0:18:35.280 started investigating them. Heart Grove was a static. It seemed 0:18:35.320 --> 0:18:39.200 like his algorithm might be working. It detected Gary Ridgeway, 0:18:39.359 --> 0:18:43.119 the Green River killer, but at the same time, he 0:18:43.280 --> 0:18:46.800 designed the algorithm to detect Ridgeway, and he kept tweaking 0:18:46.840 --> 0:18:49.879 it until he did. So maybe the algorithm had just 0:18:49.920 --> 0:18:54.160 detected Ridgeway by luck and it wouldn't generalize to other killers. 0:18:54.840 --> 0:18:58.720 In statistics, this is a problem called overfitting, and it's 0:18:58.760 --> 0:19:01.800 a common problem when scientists try to make algorithms that 0:19:01.880 --> 0:19:05.680 predict things. So to get around this problem of overfitting, 0:19:06.000 --> 0:19:08.840 people often train their algorithms with one set of data 0:19:09.240 --> 0:19:13.560 and then validate the algorithm by testing it with new data. So, 0:19:13.680 --> 0:19:16.840 since Hargrove had been using homicides in Seattle. To train 0:19:16.920 --> 0:19:19.919 his algorithm, he needed to look at other cities to 0:19:19.920 --> 0:19:23.800 see if it was really working. There were two larger 0:19:24.160 --> 0:19:29.920 suspicious clusters, and in first place was Los Angeles. They 0:19:29.960 --> 0:19:35.119 were a large group of almost entirely African American women 0:19:35.359 --> 0:19:39.040 who were killed by handguns, and those murders had a 0:19:39.200 --> 0:19:43.399 very low solution rate. So I assembled a spreadsheet of 0:19:43.400 --> 0:19:46.920 those murders and emailed them to the public relations department 0:19:46.920 --> 0:19:49.399 of the l a p D. Got one of the 0:19:49.600 --> 0:19:53.040 representatives on the phone, is there a chance that any 0:19:53.080 --> 0:19:56.560 of these could be serial murders? And he spent a 0:19:56.760 --> 0:19:59.800 minute or so looking through the files, and then he 0:19:59.800 --> 0:20:02.520 can back and said, what are you kidding? They're all 0:20:02.560 --> 0:20:07.440 serial killings? I said, what. In fact, in the es 0:20:08.040 --> 0:20:10.800 l ap D established what they hoped would be a 0:20:10.840 --> 0:20:15.040 secret task force. They didn't want to alarm the public, 0:20:15.760 --> 0:20:18.560 but they were exploring the possibility that there could be 0:20:18.600 --> 0:20:22.679 a serial killer active. They called it the South Side 0:20:22.680 --> 0:20:26.879 Slayer Task Force. Well, it turned out that the task 0:20:26.920 --> 0:20:29.520 force was misnamed. It should have been called the South 0:20:29.680 --> 0:20:34.080 Side Slayers Task Force. Because they had five they were 0:20:34.080 --> 0:20:36.880 all quite independent of each other. They didn't know each 0:20:36.880 --> 0:20:39.920 other but they were all killing women over a period 0:20:39.960 --> 0:20:44.560 of twenty years. You heard that right. During the eighties 0:20:44.600 --> 0:20:48.159 and nineties, Los Angeles had at least five different serial 0:20:48.280 --> 0:20:52.439 killers who were shooting, strangling, and sexually assaulting women in 0:20:52.440 --> 0:20:55.560 the area. This was happening at the height of the 0:20:55.640 --> 0:20:58.879 crack epidemic and at a time when homicides across the 0:20:58.960 --> 0:21:03.719 US where peaking. The large number of overall homicides probably 0:21:03.760 --> 0:21:06.080 helped cover up the fact that there were these serial 0:21:06.160 --> 0:21:09.879 killers operating in l A. But it's also likely that 0:21:09.920 --> 0:21:13.119 the police didn't give these murders enough attention due to 0:21:13.200 --> 0:21:16.760 a phenomena that's sometimes called victim discounting, and this is 0:21:16.800 --> 0:21:21.760 the tendency to ignore crimes targeting marginalized groups. But when 0:21:21.800 --> 0:21:25.199 Hargrove found out about these l A serial killers, he 0:21:25.240 --> 0:21:28.080 took this as another sign that the algorithms seemed to 0:21:28.119 --> 0:21:31.959 be working. It had detected a group of confirmed serial 0:21:32.080 --> 0:21:36.520 killers that he hadn't even been aware of. Hargrove continued 0:21:36.560 --> 0:21:40.320 down his list of top clusters that the algorithm had identified. 0:21:41.080 --> 0:21:44.639 He called up the police department in Youngstown, Ohio, and 0:21:44.720 --> 0:21:47.520 left a long message where he tried to explain the 0:21:47.560 --> 0:21:50.919 algorithm and how it had detected a cluster of murders 0:21:50.920 --> 0:21:54.520 in Youngstown. Thirty minutes later, the phone rings the chief 0:21:54.560 --> 0:21:57.720 of detectives and you gotta give him credit, answering a 0:21:57.800 --> 0:22:00.359 voicemail like that, he's a young guy. He called back 0:22:00.440 --> 0:22:03.320 and said, that was the damnedest message I ever had 0:22:03.359 --> 0:22:05.560 on my phone. And so I went back and I 0:22:05.640 --> 0:22:10.080 interviewed my senior detectives and they told me something I 0:22:10.160 --> 0:22:13.240 did not know. We thought we had a serial killer 0:22:13.280 --> 0:22:15.840 in the nineties. We definitely thought we had one, and 0:22:15.880 --> 0:22:19.359 we never got him, and so we started a new investigation. 0:22:19.920 --> 0:22:24.119 He was attempting to locate the rape kids from those cases, 0:22:24.760 --> 0:22:27.760 to try to DNA type all of the rape kits 0:22:27.800 --> 0:22:32.040 they could find. Unfortunately, and this was very embarrassing, the 0:22:32.119 --> 0:22:35.719 rape kits had all been destroyed, the property rooms had 0:22:35.760 --> 0:22:39.400 been cleaned out, and he was not able to get 0:22:39.440 --> 0:22:44.679 any of the kids from his cases or from surrounding jurisdictions. 0:22:45.320 --> 0:22:49.520 There were other similar murders in neighboring jurisdictions, but they 0:22:49.560 --> 0:22:52.600 too did not retain the rape kids. So it was 0:22:52.640 --> 0:22:54.600 it was very sad, but they gave it the old 0:22:54.640 --> 0:22:59.159 college try again. The algorithm had identified a cluster of 0:22:59.240 --> 0:23:01.399 murders that seemed like it was the work of a 0:23:01.520 --> 0:23:05.560 serial killer. Police couldn't prove that the killings were connected, 0:23:06.119 --> 0:23:09.720 but the algorithms findings lined up with what police suspected, 0:23:10.440 --> 0:23:14.400 and the algorithm had inspired a reinvestigation of those cold cases. 0:23:15.480 --> 0:23:18.159 Hargrove felt like he was on the right track and 0:23:18.240 --> 0:23:21.119 that maybe the algorithm could be a useful tool for 0:23:21.240 --> 0:23:25.600 law enforcement. We selected ten major cities that appeared to 0:23:25.640 --> 0:23:30.960 have a suspicious number of algorithm identified murders. Gary was 0:23:31.000 --> 0:23:35.480 one of those ten. Gary, Indiana, the city right next 0:23:35.480 --> 0:23:39.120 door to where Africa Hardy would be strangled. Four years later, 0:23:39.920 --> 0:23:44.639 the algorithm flagged fifteen all unsolved murders in the Gary, 0:23:44.680 --> 0:23:48.600 Indiana area. They were all women who were strangled. Not 0:23:48.680 --> 0:23:53.240 one of the cases were solved, which is unusual. Seventy 0:23:53.280 --> 0:23:57.400 seven percent of female murders get cleared, but not one 0:23:57.520 --> 0:24:02.480 of these fifteen strangulation murder and Gary were cleared. I 0:24:02.600 --> 0:24:06.560 called the public relations officer for the Gary Police Department, 0:24:06.800 --> 0:24:09.560 gave him my name and said what we had found, 0:24:10.640 --> 0:24:13.160 and said, is there a chance that you're dealing with 0:24:13.240 --> 0:24:17.280 a serial killer? The next day the phone rings, it's 0:24:17.400 --> 0:24:21.200 uh as I recall His name was Captain Roberts, who said, UM, 0:24:21.240 --> 0:24:24.760 I've checked with our detectives and I can tell you 0:24:24.920 --> 0:24:30.200 definitively that there are no unsolved serial murders in Gary, Indiana, 0:24:30.960 --> 0:24:35.679 which is by definition an impossible statement to make. Unless 0:24:35.720 --> 0:24:39.239 you have no unsolved murders, you cannot claim you'd be 0:24:39.280 --> 0:24:44.240 definitively certain that there are no unsolved serial murders. Hargrove 0:24:44.320 --> 0:24:47.000 felt like the police were just blowing him off, so 0:24:47.040 --> 0:24:51.679 he started investigating the fifteen murders himself. Hargrove wanted to 0:24:51.720 --> 0:24:55.040 know whether the algorithm was working, and he thought that 0:24:55.119 --> 0:24:58.879 if he found evidence suggesting the murders were connected, maybe 0:24:59.000 --> 0:25:02.919 the police would start take came seriously. But looking up 0:25:02.920 --> 0:25:07.280 the information about the cases wasn't easy. The FBI supplemental 0:25:07.320 --> 0:25:10.720 homicide report didn't include the victim's names or even the 0:25:10.760 --> 0:25:13.960 exact date of their death. It just listed a month 0:25:14.000 --> 0:25:17.560 in a year. So define more details about the cases, 0:25:18.040 --> 0:25:21.919 Hargrove had to meticulously dig through old issues of local papers, 0:25:22.400 --> 0:25:24.679 and as he started to piece things together, he was 0:25:24.760 --> 0:25:28.280 unsettled by what he found. He was disturbed not just 0:25:28.320 --> 0:25:31.840 by how grotesque these murders were, but by the patterns 0:25:31.880 --> 0:25:35.000 that seemed to link them, and he became convinced that 0:25:35.080 --> 0:25:38.600 Northwest Indiana had a serial killer on the loose, a 0:25:38.720 --> 0:25:43.280 strangler who was targeting young women, women just like Africa Hardy. 0:25:44.440 --> 0:25:46.960 He tried to talk to Gary's chief of police, but 0:25:47.040 --> 0:25:50.280 he couldn't get through. I continued to suggest, have you 0:25:50.359 --> 0:25:54.720 really looked at these cases? Soon the police stopped returning 0:25:54.720 --> 0:25:59.040 his calls altogether, and Hargrove needed a response. He wasn't 0:25:59.080 --> 0:26:02.400 just playing arm chair detective. We were about to publish 0:26:02.400 --> 0:26:05.680 a story saying that Gary, Indiana has a serial killer 0:26:06.320 --> 0:26:10.600 and the police would not talk about it. We were 0:26:10.640 --> 0:26:13.439 afraid that the reason they weren't talking to us was 0:26:13.480 --> 0:26:15.719 because they were hot on the heels of solving it. 0:26:16.080 --> 0:26:17.920 That they had a suspect that we're trying to reel 0:26:18.000 --> 0:26:19.879 them in, and we were going to screw that up. 0:26:21.000 --> 0:26:25.200 We needn't have worried, but um, that was our fear. 0:26:25.920 --> 0:26:29.080 I even sent registered letters to the chief of police 0:26:29.119 --> 0:26:31.440 and to the mayor saying what we were about to do, 0:26:32.040 --> 0:26:35.600 and if there's any issue they have, or any conversation 0:26:35.640 --> 0:26:39.520 they want to have, for heaven's sakes, call me. This 0:26:39.600 --> 0:26:42.480 is the letter that I wrote to the chief of 0:26:42.520 --> 0:26:48.760 Police and Mayor in Gary, Indiana, and it goes dear Chief. 0:26:48.840 --> 0:26:52.560 Carter Scripts Hour news service based in Washington, d C 0:26:52.840 --> 0:26:58.720 is conducting a national reporting project looking into the thousand 0:26:58.800 --> 0:27:03.120 unsolved thomicides committed in the United States since night. As 0:27:03.160 --> 0:27:06.600 part of this project, we are investigating whether it's possible 0:27:06.640 --> 0:27:11.600 to spot victims of serial murder among these unsolved killings. 0:27:11.760 --> 0:27:17.720 Using the FBI's Supplementary Homicide Report. Using multivariate analysis, we've 0:27:17.760 --> 0:27:21.920 determined that Gary, Indiana has an elevated number of unsolved 0:27:22.000 --> 0:27:25.600 murders of women who were strangled in recent years. The 0:27:25.720 --> 0:27:29.080