WEBVTT - Part Two: AI Is Coming for Your Children 0:00:03.720 --> 0:00:10.000 What's soulless my children's fiction? I don't know. I don't 0:00:10.000 --> 0:00:12.400 know how to open this one. We're coming back on 0:00:12.480 --> 0:00:17.000 my investigation, my two part investigation into the unsettling and 0:00:17.079 --> 0:00:22.720 the moderately evil world of ai children's book Grifters. My 0:00:22.880 --> 0:00:29.240 guest is Ben Bohlan from Ridiculous History and Boy basically 0:00:29.560 --> 0:00:33.599 all of the podcasts that helped invent podcasting as an industry. 0:00:33.640 --> 0:00:36.400 Thank you Ben for being on the show and talking 0:00:36.400 --> 0:00:41.040 with us today. Ben, you mentioned a book called Ploto 0:00:41.479 --> 0:00:46.280 in our last episode, which is this weird algorithmic novel 0:00:46.360 --> 0:00:48.919 that a guy wrote in a book in the twenties 0:00:49.000 --> 0:00:51.400 where he lays out the fourteen hundred and sixty two 0:00:51.479 --> 0:00:54.480 possible book book plots. And we wanted to start talking 0:00:54.480 --> 0:00:56.400 about this because you found your copy of Plato. 0:00:57.480 --> 0:01:02.120 Yeah, yes, yeah, thank you for having me, Robert Sophie 0:01:02.680 --> 0:01:06.880 Fellow fans of Behind the Bastards. Let's give just a 0:01:06.920 --> 0:01:11.680 short excerpt from the forward of Plato by William Wallace. 0:01:11.800 --> 0:01:16.880 Cooked Yeah. It begins, picture a man and a woman 0:01:17.319 --> 0:01:20.840 walking through a thick fog in London. The year is 0:01:20.959 --> 0:01:25.160 nineteen twenty six. They are in love and they are miserable. 0:01:27.000 --> 0:01:35.080 Well and then already, Yeah, I'm hooked. Tell me more. Yeah, 0:01:35.160 --> 0:01:38.520 you know what Greenlit get, You get exactly two seasons 0:01:38.520 --> 0:01:41.600 on Netflix. Make sure to end it on a on 0:01:41.640 --> 0:01:44.160 a cliffhanger that we won't resolve because we have to 0:01:44.160 --> 0:01:45.720 pay you more if we do season three. 0:01:45.959 --> 0:01:53.040 Yeah, Robert owns Netflix now, Yeah, I fucking knew it. 0:01:53.640 --> 0:01:55.920 I knew it as you you know, I traded in 0:01:56.000 --> 0:01:59.880 my Pontiac Aztec, which which provided me with almost twice 0:01:59.880 --> 0:02:03.120 as much money as I needed to buy Netflix. So yeah, no, 0:02:03.200 --> 0:02:06.640 I now own Netflix and no longer have a death 0:02:06.680 --> 0:02:10.520 trap car. They let me keep the ASTech they paid 0:02:10.560 --> 0:02:13.040 me not to sell it to them. Sorry, I'm just 0:02:13.080 --> 0:02:16.799 seeing need of advertising. I'm hearing needive advertising for Pontiac 0:02:16.840 --> 0:02:21.399 here for Big Pontiac. And yeah, being an advertiser on podcasts, 0:02:22.000 --> 0:02:27.880 didn't manufacture Pontiactiac. We apparently made real cars once. 0:02:28.520 --> 0:02:33.079 And I'm wondering, I'm wondering, did you write the beginning 0:02:33.120 --> 0:02:37.440 of this show? Did you? I guess we should go 0:02:38.040 --> 0:02:42.840 public with this. Part one of this week's series was 0:02:43.080 --> 0:02:47.040 written in entirely by chat GPT. Is that correct? 0:02:47.440 --> 0:02:50.400 Yeah? I actually had meant to write part two in 0:02:50.560 --> 0:02:54.560 chat GPT. But chat GPT was so horrified by the 0:02:54.600 --> 0:02:58.200 task of writing behind the Bastards that it attempted to 0:02:58.320 --> 0:03:03.320 hijack my Pontiac as Tech, which which then drove itself 0:03:03.400 --> 0:03:08.000 into a mailbox and detonated on impact. So ri I 0:03:08.080 --> 0:03:13.079 p chat GPT. You know it was. It contained billions 0:03:13.080 --> 0:03:15.160 and billions of lines of texts, but it did not 0:03:15.360 --> 0:03:19.320 contain the ability to safely pilot a Pontiac vehicle track. 0:03:19.400 --> 0:03:21.480 Now if I was, if I was a child, I 0:03:21.480 --> 0:03:26.880 would think this is an intriguing story, perhaps with some 0:03:26.880 --> 0:03:27.960 some hidden treasure. 0:03:28.560 --> 0:03:31.840 Yeah, yeah, we could. We could do a whole children's 0:03:31.840 --> 0:03:35.800 book about the Pontiac az Tech from Breaking Bad And 0:03:35.840 --> 0:03:38.120 what the fact that Walt has to drive such a 0:03:38.120 --> 0:03:41.400 piece of shit car says about his character? Really some 0:03:41.480 --> 0:03:45.160 of the best, most masterful character building. These ais could 0:03:45.240 --> 0:03:49.000 never do such effective character building as setting up the 0:03:49.160 --> 0:03:53.480 desperation of an impoverished chemistry teacher's life by showing him 0:03:53.520 --> 0:03:55.160 drive a Pontiac as Tech. 0:03:56.400 --> 0:03:59.960 Pontiac, are you unhappy? Take it on the road? 0:04:00.400 --> 0:04:04.760 Yeah, Pontiac, you won't live long in this car. 0:04:05.960 --> 0:04:07.760 God. Also, Breaking Bads really good. 0:04:08.040 --> 0:04:10.800 It is quite good, as good, as good as the 0:04:10.840 --> 0:04:15.000 Aztecs was a shitty car. So it when we left off, 0:04:15.600 --> 0:04:18.760 we were talking about we went through some really remarkable 0:04:18.960 --> 0:04:23.479 looking trannosaur images for this terrible coloring book. Now, in 0:04:23.480 --> 0:04:26.839 addition to making three legged t rexes, most AI image 0:04:26.880 --> 0:04:31.080 generators struggle to keep characters consistent across multiple images within 0:04:31.120 --> 0:04:33.040 a single book. So if you're like, you know, you've 0:04:33.080 --> 0:04:35.560 got like twenty pages, right, you need twenty illustrations for 0:04:35.600 --> 0:04:37.800 this book, and your character is this little girl in 0:04:37.839 --> 0:04:41.400 a zoo or whatever, you can give it the same input, 0:04:41.560 --> 0:04:44.280 like describe the little girl the same in each prompt, 0:04:45.320 --> 0:04:47.960 but it's really hard to get it to actually do 0:04:48.120 --> 0:04:53.080 the exact same girl in each illustration. Right. There are ways, 0:04:53.080 --> 0:04:55.920 there's whole guides to like keeping characters consistent, but it's 0:04:56.000 --> 0:04:59.359 it's a thing that like isn't easy. And most of 0:04:59.400 --> 0:05:02.600 the creators that I followed just kind of ignored this 0:05:02.800 --> 0:05:04.400 because it is kind of a pain in the ass 0:05:04.480 --> 0:05:06.440 to do, and they kind of trusted that what they 0:05:06.480 --> 0:05:09.839 were putting out, like the different illustrations looked close enough 0:05:10.120 --> 0:05:13.599 that like the parents buying these books wouldn't notice. And 0:05:13.680 --> 0:05:15.680 Sophy's going to show you. These are two pages from 0:05:15.680 --> 0:05:19.160 like a children's book about a little girl at a zoo. 0:05:19.240 --> 0:05:23.599 It's like bad, it's not about anything, but you can see, 0:05:23.640 --> 0:05:25.359 like the little girl, this is supposed to be the 0:05:25.360 --> 0:05:29.240 same character. But that's like those are clearly different little 0:05:29.320 --> 0:05:32.520 girls in both images. Oh oh yeah, one of them 0:05:32.560 --> 0:05:34.760 ass kind of curly hair, one of them asked straight hair. 0:05:35.360 --> 0:05:38.360 They're both done slightly different styles, but it is they 0:05:38.360 --> 0:05:41.360 are kind of close enough that unless you're really looking, 0:05:41.400 --> 0:05:45.480 you might not notice it. Most of the books I've seen, 0:05:45.880 --> 0:05:49.080 the consistency is even worse than that. And the laziest 0:05:49.080 --> 0:05:51.440 example of this I found is from the comic book 0:05:51.480 --> 0:05:54.039 the Adventure comic book that we talked about last episode, 0:05:54.520 --> 0:05:59.000 which is titled Treasures Beyond Gold, even though no treasures 0:05:59.000 --> 0:06:02.880 gold or otherwise actually make it into the book. Now, 0:06:02.960 --> 0:06:06.200 the author of this, Chris Heidorn, who are Christian Hidorn 0:06:06.800 --> 0:06:09.919 prevert like, does very technical prompts for his images, but 0:06:09.960 --> 0:06:12.359 this still means that he's just asking the machine to 0:06:12.480 --> 0:06:16.799 draw an attractive Western man or an attractive young Asian woman, 0:06:16.920 --> 0:06:21.080 like those are what he plugs in. And yeah, it's 0:06:21.120 --> 0:06:23.440 it's not like the so one of the for example, 0:06:23.520 --> 0:06:25.200 like one of the one of the prompts. He's got 0:06:25.240 --> 0:06:27.960 is like slash imagine blend of comic book art and 0:06:28.000 --> 0:06:31.160 line art and full natural colors. Attractive Western man in 0:06:31.160 --> 0:06:34.360 his early thirties with short cropped brown hair and stubble beard, 0:06:34.560 --> 0:06:37.799 shirt in beige color, walking through a bustling Southeast Asian 0:06:37.839 --> 0:06:41.839 market reading treasure map. And this results in a comic 0:06:41.880 --> 0:06:44.840 book where every single page, both of our main two 0:06:44.920 --> 0:06:49.200 characters are completely different people often drawn in decent styles, 0:06:49.480 --> 0:06:51.240 so you can see different styles. You can see in 0:06:51.279 --> 0:06:53.680 these two different images from two different pages. In the 0:06:53.680 --> 0:06:56.600 first one, she looks like, I don't know, like a 0:06:58.320 --> 0:07:00.560 you know, you've got like the the late character, and 0:07:00.600 --> 0:07:03.880 she's kind of like got a T shirt and what 0:07:03.960 --> 0:07:07.200 looks like a bandolier a sashel around her shoulder. She's 0:07:07.240 --> 0:07:10.400 got long straight hair. The male lead looks like Dean 0:07:10.480 --> 0:07:16.080 Winchester from Supernatural, you know, and he's got like a 0:07:16.120 --> 0:07:18.600 green over. He's not actually wearing baths like she's like 0:07:18.600 --> 0:07:20.640 the thing said, but like and then in the second image, 0:07:20.640 --> 0:07:22.960 from like a page or two later, she's been like 0:07:23.120 --> 0:07:25.840 animed up like twenty percent. Her eyes are like three 0:07:25.880 --> 0:07:30.400 times as large, and then he's gotten like fifteen percent 0:07:30.520 --> 0:07:34.000 Shaggy from Scooby Doo added to him, like, they're not 0:07:34.120 --> 0:07:37.800 the same people. She's got lip filler, she's wearing like 0:07:39.520 --> 0:07:42.120 what you might call it a sleeveless shirt now and 0:07:42.200 --> 0:07:46.000 again like an animate up a little bit. He's like, like, 0:07:46.160 --> 0:07:48.920 they're not the same people. Like, they're very clearly, very 0:07:48.960 --> 0:07:50.280 different looking characters. 0:07:50.440 --> 0:07:53.480 They lost their they also they lost their gear. 0:07:53.800 --> 0:07:56.480 Yeah, they're wearing totally different clothing and equipment. 0:07:57.000 --> 0:08:00.280 Yeah, and you can tell from the facial structure, like 0:08:00.320 --> 0:08:03.720 the the the curvature of the job. You know what 0:08:03.800 --> 0:08:06.120 it is. You know what it is if you're a 0:08:06.320 --> 0:08:10.800 parent thumbing through something like this. And again, everybody tune 0:08:10.840 --> 0:08:13.640 into part one if you haven't listened yet, if your 0:08:13.720 --> 0:08:17.000 parent thumbing through something like this or encountering this stuff 0:08:17.000 --> 0:08:23.200 with a cursory look. Then the two images of people, 0:08:23.680 --> 0:08:27.160 they they look like they could be related to each other. 0:08:27.600 --> 0:08:32.800 But to your point, Robert, very clearly not the same folks. 0:08:32.480 --> 0:08:35.840 Not the same folks. And like obviously there's the normal 0:08:35.920 --> 0:08:39.280 like weirdness like it. You know, the Dean Winchester version 0:08:39.320 --> 0:08:41.200 of the character looks better, but his neck is like 0:08:41.280 --> 0:08:46.560 cocked to an angle. Yeah, she's like a little kid, 0:08:46.640 --> 0:08:47.480 and the second one. 0:08:47.960 --> 0:08:50.920 Her face off the size of a child's face. 0:08:51.320 --> 0:08:56.840 And then yeah, but this is so Yeah. 0:08:55.720 --> 0:08:58.920 It's weird. It's weird. And again, an adult who like 0:08:59.280 --> 0:09:02.360 actually looked at this would kind of recognize a couple 0:09:02.360 --> 0:09:04.400 of pages in, oh, this is like some weird, shitty 0:09:04.440 --> 0:09:07.960 AI art thing. But again, these books, I think can 0:09:08.000 --> 0:09:10.280 be damaging to little kids. And this is what gets 0:09:10.280 --> 0:09:12.560 me into the actual educational research that I did for 0:09:12.640 --> 0:09:14.160 this investigation, because. 0:09:14.160 --> 0:09:18.640 There's a yeah literary you were talking about this in 0:09:18.760 --> 0:09:24.120 part one, right, this is triggered or inspired deep dive 0:09:24.360 --> 0:09:25.640 into a sort of. 0:09:27.200 --> 0:09:30.679 Yeah, literary educational theory because like I wanted to know, 0:09:31.040 --> 0:09:33.960 is it like bad for kids to get handed nonsense 0:09:34.000 --> 0:09:38.640 books that aren't like, aren't about anything at all, and 0:09:38.679 --> 0:09:41.240 where the art is like not actually art, like where 0:09:41.240 --> 0:09:43.440 there's no intentionality behind it. It's just kind of like 0:09:43.480 --> 0:09:47.640 clip art placed more or less randomly and often slightly 0:09:47.720 --> 0:09:51.880 warped by you know, a machine hallucination. And yeah, there's 0:09:51.920 --> 0:09:55.640 actually there's a substantial body of scientific research into what 0:09:55.880 --> 0:10:01.439 is referred to as emergent literacy. Right. Emergent literacy is 0:10:01.520 --> 0:10:04.240 the these are the reading and writing skills that a 0:10:04.400 --> 0:10:09.200 child possesses or builds before they are can formally reader write. 0:10:09.320 --> 0:10:12.880 So when you are sitting down with your six month 0:10:12.920 --> 0:10:15.160 eight month old kid and you're going over a storybook 0:10:15.240 --> 0:10:17.240 that kid can't read, right, they can't like look at 0:10:17.240 --> 0:10:20.319 words and recognize what the individual words are. But because 0:10:20.320 --> 0:10:22.640 you're reading the story, they are starting to pick up 0:10:22.720 --> 0:10:26.319 on aspects of how stories are structured, what a story 0:10:26.320 --> 0:10:28.560 can tell, what characters are. These are all things that 0:10:28.559 --> 0:10:32.360 they are picking up that aren't literacy but also are 0:10:32.400 --> 0:10:38.120 a crucial building block to literacy. Right, this is emergent literacy. Right. 0:10:38.120 --> 0:10:40.880 That's this is a critical part of kids learning how 0:10:40.920 --> 0:10:44.840 to read and learning how to appreciate reading. Right, It's 0:10:44.880 --> 0:10:47.880 why Like when I was a kid, my mom like 0:10:48.200 --> 0:10:51.079 the honestly, like ninety percent of her parenting strategy is 0:10:51.120 --> 0:10:53.400 basically make sure he always has a book in his hands. 0:10:54.080 --> 0:11:01.360 Oh nice, were you able to did you have curation? Curate? 0:11:03.280 --> 0:11:04.239 Did you have autotomy? 0:11:04.840 --> 0:11:06.959 When I was too young to pick my because you know, 0:11:07.000 --> 0:11:08.960 at a certain age, like you know, I was six 0:11:09.000 --> 0:11:10.560 months of a year old or whatever, I'm not really 0:11:10.679 --> 0:11:12.400 picking my own books. She's just like you were like, 0:11:14.040 --> 0:11:15.760 but it was also like I did have a lot 0:11:15.760 --> 0:11:19.160 of like my grandma had basically every national geographic and 0:11:19.200 --> 0:11:21.400 like I would see ones with pirates or dinosaurs, and 0:11:21.480 --> 0:11:24.680 so I had like a lot of that shit. No 0:11:24.800 --> 0:11:27.240 as I and as I like was younger. You know, 0:11:27.280 --> 0:11:29.800 when I was in second grade, I found my dad 0:11:29.840 --> 0:11:32.000 had a copy of The Lost World checked out from 0:11:32.000 --> 0:11:34.440 the library, and I like demanded he renew it because 0:11:34.440 --> 0:11:36.200 there was a dinosaurs goll on the front. And so 0:11:36.240 --> 0:11:38.560 in second grade I read The Lost World, which is 0:11:38.600 --> 0:11:40.640 not a book that's for second graders. But my mom's 0:11:40.640 --> 0:11:44.079 attitude was like, like I had, my TV access was 0:11:44.120 --> 0:11:46.319 super restricted. I couldn't watch like Rin and Stimpy or 0:11:46.360 --> 0:11:48.520 The Simpsons as a little kid. But like, if it 0:11:48.600 --> 0:11:50.680 was a book, it didn't matter if there was fucking, 0:11:50.800 --> 0:11:53.920 if there was murder, if there was like sex crimes. 0:11:54.280 --> 0:11:56.200 As long as it was a book, it was okay. 0:11:56.280 --> 0:11:58.800 That was like my mom's attitude. If he's reading it's fine. 0:11:59.600 --> 0:12:02.400 I feel. I read Stephen King's It when he was 0:12:02.480 --> 0:12:03.520 probably too young. 0:12:03.760 --> 0:12:05.440 That's a fucked up book for a kid. 0:12:05.520 --> 0:12:11.320 Sure, it's not ideal. And my and my parents were like, hey, 0:12:11.840 --> 0:12:15.120 they were bragging. They were like, hey, look at our 0:12:15.200 --> 0:12:20.320 kid reads. Uh, this stuff on his own. What a 0:12:20.360 --> 0:12:23.480 self starter? What a literate child? 0:12:24.000 --> 0:12:26.200 And I actually think we're joking about it. I actually 0:12:26.200 --> 0:12:29.680 think like kids reading about fucked up shit is good 0:12:29.720 --> 0:12:32.040 for them in a way that like maybe kids watching 0:12:32.160 --> 0:12:35.480 fucked up shit in movies or TV isn't because interesting. 0:12:35.679 --> 0:12:38.240 There's the degree of we're going to talk about this, 0:12:38.400 --> 0:12:42.800 Like watching TV or a movie is much more of 0:12:42.840 --> 0:12:45.120 a one way street, especially for a kid. You know, 0:12:45.200 --> 0:12:47.520 as adults you kind of get the ability to sort 0:12:47.559 --> 0:12:51.000 of interact with and analyze it more. I do think 0:12:51.040 --> 0:12:54.200 that like watching TV or a movie is more of 0:12:54.240 --> 0:12:56.000 a one way thing than like when you are reading, 0:12:56.040 --> 0:12:58.960 as we'll talk about, it's a dialogue between you and 0:12:59.040 --> 0:13:02.800 the book, right, like you are actively constructing meaning alongside 0:13:02.840 --> 0:13:07.280 that work. Anyway, this gets us back to like emergent storytelling, 0:13:07.320 --> 0:13:10.360 because a lot of aspects of emergent storytelling are things 0:13:10.440 --> 0:13:14.920 like understanding that the illustrations in a book are carrying 0:13:14.960 --> 0:13:19.560 aspects of character and aspects of the story. And so 0:13:20.200 --> 0:13:22.720 little kids, very little kids, one year olds, two years 0:13:22.880 --> 0:13:26.240 all year olds earlier than you'd expect, have already started 0:13:26.240 --> 0:13:28.760 to realize that when they are looking at a story book, 0:13:29.360 --> 0:13:31.800 what you're reading them is not the whole story. The 0:13:31.840 --> 0:13:35.800 illustrations are part of the story. One study on emergent 0:13:35.840 --> 0:13:39.120 reading strategies I read by Judith Leisiker and Elizabeth Hopper 0:13:39.160 --> 0:13:43.040 of Purdue University noted emergent reading strategies such as wordless 0:13:43.040 --> 0:13:45.679 book reading are often seen as precursors to the meaning 0:13:45.720 --> 0:13:48.800 making that comes later during print reading. I actually found 0:13:49.080 --> 0:13:51.720 one study where like they would read kids a story, 0:13:52.200 --> 0:13:53.960 and then they would hand them a copy of that 0:13:54.040 --> 0:13:56.040 story book without any text on it, and they would 0:13:56.080 --> 0:13:59.520 ask the kids to write the story, and the kids 0:13:59.559 --> 0:14:02.920 would write more detailed stories than the original versions because 0:14:02.960 --> 0:14:06.480 they're taking things that they recognize from the illustrations and 0:14:06.559 --> 0:14:09.240 adding that in when they recreate the story, you know, 0:14:11.080 --> 0:14:13.400 which is really interesting to me. And that's what scares 0:14:13.440 --> 0:14:16.080 me about a lot of these mid journey created children's books, 0:14:16.120 --> 0:14:19.120 because when you've got a story that somebody is wanting 0:14:19.160 --> 0:14:23.920 to tell transmit information through their story, their illustrations are 0:14:23.960 --> 0:14:27.400 transmitting information too. That is not None of these illustrations 0:14:27.440 --> 0:14:30.240 and these abooks are transmitting information. They are there to 0:14:30.400 --> 0:14:33.600 tick a box, but like the characters aren't interacting. They 0:14:33.600 --> 0:14:35.680 don't even match what they're supposed to match with the 0:14:35.680 --> 0:14:38.520 prompt says like it's all off, Like it may look 0:14:39.000 --> 0:14:41.640 like a human drawing, Like these look like competent drawings, 0:14:41.680 --> 0:14:44.920 but they're not drawings of anything. Nothing is being revealed 0:14:44.920 --> 0:14:48.200 in the faces of these characters, in their physical positioning 0:14:48.240 --> 0:14:51.080 and the actions that they're shown taking part in it, Like, 0:14:51.160 --> 0:14:54.160 none of that is actually present here because there's not 0:14:54.480 --> 0:15:01.400 a person driving the artwork. And that's really small. Children 0:15:01.440 --> 0:15:04.360 are info vacuums. They are hoovering up observations about the 0:15:04.360 --> 0:15:07.240 world at a terrifying pace, and before they can read, 0:15:07.600 --> 0:15:10.160 they come to understand things like story structure and the 0:15:10.200 --> 0:15:12.920 meaning of words and phrases by studying the illustrations that 0:15:12.960 --> 0:15:16.960 accompany text. By breaking the illustrative part of a story book, 0:15:17.280 --> 0:15:19.880 you are breaking the way in which kids learn to 0:15:20.000 --> 0:15:23.000 read at a fundamental level before they even under Like 0:15:23.040 --> 0:15:28.479 the precursors to literacy are shattered by not showing them 0:15:28.640 --> 0:15:32.280 actual illustrations. Like there's a real danger of that here. 0:15:33.080 --> 0:15:36.480 The fact that these are so disjointed and wrong could 0:15:36.520 --> 0:15:39.840 fuck up the way kids sort of are understanding these 0:15:39.880 --> 0:15:44.240 stories on a very fundamental level. Yeah, and that's really frightening. 0:15:44.280 --> 0:15:46.480 To me, that's a real risk. 0:15:47.400 --> 0:15:50.880 That's not even I mean, we can call it a risk. 0:15:51.080 --> 0:15:57.440 But this is as earlier establishment. This is a thing 0:15:57.560 --> 0:16:01.480 that is happening. Yeah, there is the injurious potential short, 0:16:01.560 --> 0:16:05.880 but that potential has to a degree been actualized. 0:16:05.160 --> 0:16:08.080 You know, and some number of kids have had these 0:16:08.080 --> 0:16:09.720 books handed to them already. 0:16:10.880 --> 0:16:14.240 And and what what what would be then? I think 0:16:14.320 --> 0:16:20.000 everybody probably is thinking the same question, what would the 0:16:21.040 --> 0:16:24.200 risk be? Like? What is the what is the worst 0:16:24.360 --> 0:16:29.840 case scenario? Does a does a child uh read a 0:16:30.240 --> 0:16:33.520 does like a latchkey kid, uh sit alone with their 0:16:33.800 --> 0:16:39.920 uh fantastically terrible uh children's stories and they what They 0:16:39.960 --> 0:16:42.560 go to a museum one day and they say, Hey, 0:16:44.080 --> 0:16:47.640 the t rex skeleton is wrong? Where are the thumbs? 0:16:48.000 --> 0:16:50.840 I think that's I think that's like one a specific 0:16:51.120 --> 0:16:52.760 thing that could happen as a result of these like 0:16:52.800 --> 0:16:55.000 weird coloring books. I think the scarier thing that could 0:16:55.000 --> 0:16:56.640 happen as the result of stories one of them is 0:16:56.640 --> 0:17:01.240 that like kids who become readers, who come to love 0:17:01.320 --> 0:17:06.120 reading and fiction and thus writing and then create you know, culture, right, 0:17:06.240 --> 0:17:08.600 Like large aspects of our culture created by kids who 0:17:08.600 --> 0:17:11.520 love to read as kids and then become writers. That 0:17:11.800 --> 0:17:15.120 a necessary part of that is loving and understanding stories, 0:17:15.160 --> 0:17:19.160 and a necessary part of that is into gradually integrating 0:17:19.200 --> 0:17:21.680 the illustrations, which are the first things that you start 0:17:21.720 --> 0:17:25.680 to recognize in story books as a kid, to words 0:17:25.720 --> 0:17:27.280 and the way that words work, and the way that 0:17:27.280 --> 0:17:30.800 words tell stories and describe characters and plot, and this 0:17:30.960 --> 0:17:33.879 kind of can break that. The risk is that like, 0:17:34.400 --> 0:17:37.600 and kids are you know, potentially kind of fragile here, 0:17:37.640 --> 0:17:42.040 Like you could damage the ability of children to appreciate 0:17:42.160 --> 0:17:45.840 reading and to appreciate stories. And maybe kids who would 0:17:45.840 --> 0:17:48.520 have been readers, who would have cared about this stuff, 0:17:48.600 --> 0:17:52.000 or who would have wanted to create things won't because 0:17:52.520 --> 0:17:56.800 they're kind of at a very early stage. Their understanding 0:17:56.840 --> 0:17:59.720 of what reading is for is broken because it's not 0:17:59.800 --> 0:18:04.240 being They're not having a conversation, you know, between an 0:18:04.280 --> 0:18:08.280 author and an illustrator in themselves and constructing meaning from that. 0:18:08.800 --> 0:18:15.240 They are having this simulacrum of a story. This this nonsense. 0:18:15.280 --> 0:18:19.160 These the fucking potato chips of of of like even 0:18:19.200 --> 0:18:21.520 worse than potato chips. If you ever you know, you know, 0:18:21.600 --> 0:18:23.080 did you ever read good Omens? 0:18:23.520 --> 0:18:26.200 Yes? Yeah, Terry Pratchett and Neil. 0:18:26.000 --> 0:18:28.600 Pratchett and Neil Game and one of the it's you know, 0:18:28.640 --> 0:18:30.920 it's it's a book. Anti Christ is like the hero 0:18:31.119 --> 0:18:32.760 kind of and there's like the four Horsemen of the 0:18:32.800 --> 0:18:36.360 Apocalypse our characters in it, and one of them, Famine 0:18:36.920 --> 0:18:39.520 is like kind of a little counterintuitively at first, he's 0:18:39.560 --> 0:18:42.200 like his big plot, he's like running this some fast 0:18:42.200 --> 0:18:45.560 food franchise. The food has any nutrition in it, Like 0:18:45.600 --> 0:18:48.000 there's nothing. It's actually you can starve to death eating 0:18:48.000 --> 0:18:50.879 this food. Like That's what I think of when I 0:18:50.920 --> 0:18:53.080 think of these books and what they can do to 0:18:53.280 --> 0:18:57.960 like kid's abilit like developing literacy. I find that unsettling, 0:18:59.240 --> 0:19:01.960 AI advocate when you talk about how fucked up and 0:19:02.000 --> 0:19:05.200 wrong a lot of this looks. We'll talk about like, well, 0:19:05.200 --> 0:19:06.879 you know, this is just mid journey version I know 0:19:06.880 --> 0:19:09.040 three or four or whatever, or this is you know, 0:19:09.119 --> 0:19:11.359 version three or four of chatch Ept and it's only 0:19:11.400 --> 0:19:13.000 going to get better. Look at how much better it 0:19:13.040 --> 0:19:15.240 is now than it was. You know, before you knew 0:19:15.240 --> 0:19:17.960 these things existed. It's going to get so much better. 0:19:18.000 --> 0:19:20.040 You know, eventually it'll be seamless. You won't be able 0:19:20.040 --> 0:19:23.320 to tell. That's actually not a guarantee. Nobody knows that. 0:19:23.640 --> 0:19:25.600 For one thing, we're kind of off the map here, 0:19:25.640 --> 0:19:28.520 like one of the big Like there's aspects of how 0:19:28.560 --> 0:19:31.960 these things work that are kind of like unclear even 0:19:31.960 --> 0:19:34.120 to the people making them in aspects of how good 0:19:34.240 --> 0:19:37.120 can they be. And one of the things I will say, 0:19:37.359 --> 0:19:40.400 I heard this from somebody who's who's in the industry recently, 0:19:40.600 --> 0:19:44.720 was like, you know, like, how do you tell whether 0:19:44.800 --> 0:19:46.960 or not the model is getting more intelligent? And he's like, 0:19:47.000 --> 0:19:48.320 I don't know, how do you tell if people are 0:19:48.320 --> 0:19:50.280 more or less intelligent? We don't have a good agreement 0:19:50.359 --> 0:19:52.959 on that, like a like IQ's bullshit, Like, yeah, that's 0:19:52.960 --> 0:19:55.639 actually pretty good point, you know, like you're you're not wrong. 0:19:56.760 --> 0:19:59.720 But also just like the idea that these models will 0:19:59.720 --> 0:20:04.520 get better and better at storytelling, at creating fiction, at 0:20:04.520 --> 0:20:07.440 creating images to go with fiction, that is not a guarantee. 0:20:07.600 --> 0:20:09.399 One of the reasons why that's not a guarantee is 0:20:09.400 --> 0:20:12.800 that the popularity of AI tools means that the Internet 0:20:12.840 --> 0:20:15.160 at a very rapid pace, is being flooded with more 0:20:15.200 --> 0:20:17.720 AI generated stuff. Right, More and more of this is 0:20:17.720 --> 0:20:20.359 getting spat out on the Internet every day, And because 0:20:20.480 --> 0:20:24.520 new ais will be trained on this content. That means 0:20:24.520 --> 0:20:27.400 that new AIS and AIS that are updated to have 0:20:28.080 --> 0:20:30.959 more stuff from after twenty twenty two are going to 0:20:31.040 --> 0:20:34.280 be trained on stuff that they generated. So you are 0:20:34.359 --> 0:20:38.080 feeding AI art and text back into the model and 0:20:38.119 --> 0:20:41.960 feed it's a feedback. It can lead to what researchers 0:20:42.000 --> 0:20:44.720 call model collapse, right, which is the idea that like, well, 0:20:44.920 --> 0:20:47.960 the kind of these derangements, these messed up illustrations, like 0:20:48.000 --> 0:20:51.879 the faults. If you're feeding this back and retraining it 0:20:51.960 --> 0:20:54.920 on flawed stuff that had already put out, it's going 0:20:54.920 --> 0:20:58.280 to just keep exacerbating those flaws. A group of researchers 0:20:58.320 --> 0:21:02.119 published in the General ARCSIV described this as what happens 0:21:02.160 --> 0:21:05.159 when quote, the use of model generated content and training 0:21:05.200 --> 0:21:09.119 causes irreversible defects in the resulting models. Oh like that 0:21:09.240 --> 0:21:13.720 movie multiplicity exactly, like multiplicity. That's right, that's right. This 0:21:13.840 --> 0:21:19.040 is a multiplicity kind of situation. Yeah. And I find 0:21:19.040 --> 0:21:22.520 this particularly worrisome because the present models are already pretty 0:21:22.520 --> 0:21:26.879 full of defects. Take the storybook generated for a video 0:21:27.040 --> 0:21:30.359 called I Create a best selling children's book using AI 0:21:30.600 --> 0:21:33.320 in under an hour. Now. The creator of this video, 0:21:33.400 --> 0:21:36.720 whose name is Grayson Sands, looks like what you'd get 0:21:36.760 --> 0:21:39.080 if you fed mid journey the prompt what if Ron 0:21:39.119 --> 0:21:42.119 Weasley was a registered sex offender? And I don't know 0:21:42.160 --> 0:21:45.400 if i'll include that joke in the final article. It's 0:21:45.480 --> 0:21:48.959 mean and not proper for like a serious piece of journalism. 0:21:49.240 --> 0:21:50.640 But I don't like this guy. 0:21:51.440 --> 0:21:54.639 It's also it's also just I mean because we're an 0:21:54.640 --> 0:21:59.080 audio podcast, right, so it's also just for everyone playing 0:21:59.119 --> 0:22:04.159 along at home. Uh do look at do look it up? 0:22:04.480 --> 0:22:07.400 You know I hate black hair. 0:22:07.560 --> 0:22:11.440 He's got the big, like white framed sunglasses. He's got 0:22:11.440 --> 0:22:14.320 a leather jacket on. There's a guitar hung on his 0:22:14.400 --> 0:22:15.680 wall behind him for someone. 0:22:15.760 --> 0:22:19.600 He's got epaulets, which for some reason bothers me. It's 0:22:19.920 --> 0:22:24.400 it's like a Again, we're very not into body shaming. 0:22:24.680 --> 0:22:27.800 But this is not about his body shape or anything 0:22:27.800 --> 0:22:32.800 about this. This is why his aesthetic choices, his vibe. 0:22:33.240 --> 0:22:38.280 His vibe. His vibe looks off. So wait, in under 0:22:38.320 --> 0:22:41.800 an hour, we create a best selling children's book. 0:22:42.960 --> 0:22:45.119 Sounds like there's what he promises in the video. 0:22:45.640 --> 0:22:47.720 Sounds like there should be some air quotes around a 0:22:47.720 --> 0:22:49.400 couple of these things, Robert. 0:22:49.760 --> 0:22:52.479 All of them. So he decides he's going to generate 0:22:52.520 --> 0:22:55.600