1 00:00:01,440 --> 00:00:04,800 Speaker 1: Hey, welcome to Science Stuff, a production of iHeartRadio. I'm 2 00:00:04,840 --> 00:00:09,440 Speaker 1: horehit Champ, and today we're talking about artificial intelligence, or AI, 3 00:00:09,920 --> 00:00:13,800 Speaker 1: and specifically images and videos made by AI, which some 4 00:00:13,880 --> 00:00:17,599 Speaker 1: people call AI slop, and how it affects us, our brains, 5 00:00:17,680 --> 00:00:21,040 Speaker 1: our society, and our sense of reality. We're going to 6 00:00:21,079 --> 00:00:24,320 Speaker 1: be talking to an AI expert about this, a neuroscientist, 7 00:00:24,720 --> 00:00:28,560 Speaker 1: and somebody who's looked into the dangers of AI generated 8 00:00:28,600 --> 00:00:33,560 Speaker 1: pictures of cute, fuzzy animals. Yeah, you think they're perfectly safe, 9 00:00:33,600 --> 00:00:39,760 Speaker 1: but they're not. So ask chat GPT to clear your schedule, 10 00:00:40,120 --> 00:00:43,360 Speaker 1: claude your way to our show and mid journey with 11 00:00:43,440 --> 00:00:46,720 Speaker 1: us as we answer the question what is AI slop 12 00:00:47,200 --> 00:00:54,880 Speaker 1: doing to us? Enjoy? Hey? Everyone, So you've probably noticed 13 00:00:54,920 --> 00:00:58,800 Speaker 1: that AI is everywhere, and you've probably seen and heard 14 00:00:58,920 --> 00:01:02,120 Speaker 1: images and videos made by AI programs in your social 15 00:01:02,160 --> 00:01:06,800 Speaker 1: media feed, online, and even on TV, sometimes without knowing 16 00:01:06,840 --> 00:01:10,000 Speaker 1: it was made by AI. And this trend has exploded 17 00:01:10,319 --> 00:01:13,120 Speaker 1: in just the last few years. So I was curious 18 00:01:13,120 --> 00:01:16,280 Speaker 1: to know two things. First, what happened a few years 19 00:01:16,319 --> 00:01:19,319 Speaker 1: ago that led to this sudden explosion of AI content 20 00:01:19,440 --> 00:01:23,040 Speaker 1: and apps and services. Do we suddenly discover a magical 21 00:01:23,120 --> 00:01:27,600 Speaker 1: algorithm or alien artifact that suddenly gives machines this much power. 22 00:01:28,720 --> 00:01:31,720 Speaker 1: And second, I wanted to know what effect all this 23 00:01:31,880 --> 00:01:36,240 Speaker 1: fake AI made content has on our brains and how 24 00:01:36,280 --> 00:01:39,840 Speaker 1: we see the world. So we'll answer both questions, and 25 00:01:39,920 --> 00:01:43,360 Speaker 1: we'll start with the first. What happened a few years 26 00:01:43,360 --> 00:01:47,240 Speaker 1: ago that started this recent explosion in AI? And is 27 00:01:47,280 --> 00:01:50,240 Speaker 1: there a limit to it? To answer these questions, I 28 00:01:50,360 --> 00:01:52,640 Speaker 1: reached out to this guy I met at lunch the 29 00:01:52,680 --> 00:01:57,520 Speaker 1: other day, who I'm pretty sure is human. Well, thank 30 00:01:57,560 --> 00:01:59,080 Speaker 1: you doctor Tam for joining us. 31 00:01:59,200 --> 00:02:01,640 Speaker 2: Thank you JORGEV for having me. So my name is 32 00:02:01,680 --> 00:02:06,080 Speaker 2: doctor ed Tam. I am a postdoctoral fellow at Stanford University. 33 00:02:06,320 --> 00:02:09,560 Speaker 2: A lot of my work concerns the uncertainty around AI 34 00:02:10,120 --> 00:02:13,040 Speaker 2: and what are the limitations of AI and how we 35 00:02:13,120 --> 00:02:14,959 Speaker 2: can potentially overcome those. 36 00:02:15,000 --> 00:02:16,960 Speaker 1: And how can I make sure you're real and not 37 00:02:17,040 --> 00:02:17,440 Speaker 1: an AI. 38 00:02:19,280 --> 00:02:23,440 Speaker 2: Well, we've met in person, so you know that. 39 00:02:23,919 --> 00:02:28,800 Speaker 1: I met someone who looks like you a role I know, 40 00:02:28,880 --> 00:02:30,400 Speaker 1: although it was over at lunch. 41 00:02:30,520 --> 00:02:33,960 Speaker 2: So yes, a robot that can eat and joke and drink. 42 00:02:36,120 --> 00:02:38,440 Speaker 1: Okay, doctor ed Tam is going to tell me what 43 00:02:38,560 --> 00:02:41,960 Speaker 1: happened a few years ago that made AI swimmingly explode 44 00:02:41,960 --> 00:02:44,280 Speaker 1: in what it's able to do. But to do that, 45 00:02:44,520 --> 00:02:46,760 Speaker 1: we have to go back to where it all started, 46 00:02:47,040 --> 00:02:50,280 Speaker 1: which is when scientists had the idea for AI in 47 00:02:50,360 --> 00:02:51,040 Speaker 1: the first place. 48 00:02:52,960 --> 00:02:56,160 Speaker 2: It goes back to like the last century at least. 49 00:02:56,400 --> 00:02:59,200 Speaker 3: So the idea is that the human brain is composed 50 00:02:59,200 --> 00:03:02,640 Speaker 3: of cells that are called neurons, and these neurons connect 51 00:03:02,639 --> 00:03:05,760 Speaker 3: with each other and these different productivity patterns in the 52 00:03:05,800 --> 00:03:08,760 Speaker 3: brain gives us the rich functionality of the human brain. 53 00:03:09,000 --> 00:03:12,600 Speaker 3: So back in the nineteen hundred, scientists we're thinking, can 54 00:03:12,639 --> 00:03:18,120 Speaker 3: we somehow mimic the biological human neural network with a computer, 55 00:03:19,000 --> 00:03:21,160 Speaker 3: And so there are some early models that kind of 56 00:03:21,200 --> 00:03:23,240 Speaker 3: try to tries to do this. One of the earliest 57 00:03:23,280 --> 00:03:25,920 Speaker 3: and famous one is called the perceptron, and it's just 58 00:03:25,960 --> 00:03:28,520 Speaker 3: this kind of like three layer no network with some 59 00:03:29,040 --> 00:03:31,720 Speaker 3: nonlinear function and coded. So that was kind of like 60 00:03:31,760 --> 00:03:34,360 Speaker 3: the first model, but it was already kind of like 61 00:03:34,400 --> 00:03:35,480 Speaker 3: a mini breakthrough. 62 00:03:35,880 --> 00:03:38,200 Speaker 1: It sounds like a transformer like the toys. 63 00:03:38,640 --> 00:03:42,400 Speaker 3: Oh yes, yes, megatron. 64 00:03:43,680 --> 00:03:45,400 Speaker 2: I don't know why they call it the perceptron. 65 00:03:46,400 --> 00:03:50,080 Speaker 1: So, starting in the nineteen forties, scientists and engineers had 66 00:03:50,120 --> 00:03:52,720 Speaker 1: an idea of how the human brain worked. It was 67 00:03:52,760 --> 00:03:55,840 Speaker 1: a humongous network of neurons connected to each other, and 68 00:03:55,880 --> 00:03:58,720 Speaker 1: so the scientists and engineers wondered if we could make 69 00:03:58,880 --> 00:04:03,320 Speaker 1: an artificial brain. So they programmed on a computer how 70 00:04:03,360 --> 00:04:07,360 Speaker 1: a neuron reacts. It takes signals from other neurons, weighs 71 00:04:07,400 --> 00:04:10,840 Speaker 1: each input depending on some numbers or parameters, and then 72 00:04:11,040 --> 00:04:13,200 Speaker 1: if it all adds up to a certain amount, it 73 00:04:13,280 --> 00:04:17,400 Speaker 1: fires off a single signal to other neurons. And at 74 00:04:17,440 --> 00:04:21,040 Speaker 1: first these networks with layers of neurons were pretty simple, 75 00:04:21,400 --> 00:04:28,800 Speaker 1: but amazingly they could easily recognize patterns, meaning like recognize 76 00:04:28,960 --> 00:04:31,640 Speaker 1: numbers or images kind. 77 00:04:31,520 --> 00:04:35,160 Speaker 3: Of that's exactly right, like telling apart of human digits, 78 00:04:35,360 --> 00:04:37,159 Speaker 3: can you tell a part a cat from a dog. 79 00:04:37,200 --> 00:04:39,680 Speaker 3: There are all these image benchmark data sets that people 80 00:04:39,839 --> 00:04:42,200 Speaker 3: work on, and then it turns out when you stack 81 00:04:42,279 --> 00:04:45,080 Speaker 3: them on top of each other, it can encode really 82 00:04:45,160 --> 00:04:46,200 Speaker 3: really rich patterns. 83 00:04:47,720 --> 00:04:50,960 Speaker 1: So things progress somewhat slowly. Over the next few decades, 84 00:04:51,279 --> 00:04:53,960 Speaker 1: people started to make bigger and bigger networks of these 85 00:04:54,040 --> 00:04:57,520 Speaker 1: artificial neurons. Instead of a few neurons, it started to 86 00:04:57,520 --> 00:05:01,200 Speaker 1: become dozens, and then hundreds, and then thousands of these 87 00:05:01,200 --> 00:05:05,240 Speaker 1: neurons and dozens of these layers, and it was pretty powerful. 88 00:05:05,760 --> 00:05:08,240 Speaker 1: You could give the machine a whole bunch of images 89 00:05:08,279 --> 00:05:12,240 Speaker 1: of dogs and they could learn to recognize dogs. You 90 00:05:12,240 --> 00:05:14,920 Speaker 1: could give it a whole bunch of human fingerprints or faces, 91 00:05:15,160 --> 00:05:18,200 Speaker 1: and they could recognize fingerprints and faces. 92 00:05:18,760 --> 00:05:21,400 Speaker 3: So that was maybe in like the twenty tens or 93 00:05:21,400 --> 00:05:24,280 Speaker 3: something like that. When you go on Facebook, you know how, 94 00:05:24,360 --> 00:05:27,640 Speaker 3: if you post a picture on Facebook, it actually tags 95 00:05:27,640 --> 00:05:30,360 Speaker 3: your face and it tags the faces of your friends, 96 00:05:30,880 --> 00:05:32,560 Speaker 3: and I remember that being like magic. 97 00:05:32,680 --> 00:05:34,960 Speaker 2: I was like, wow, like it recognizes me. 98 00:05:35,880 --> 00:05:38,440 Speaker 3: And that's the type of AI back then ten years 99 00:05:38,480 --> 00:05:39,840 Speaker 3: ago that can do that. 100 00:05:41,000 --> 00:05:45,720 Speaker 1: So ten years ago, AIS were really good at recognizing things, 101 00:05:46,120 --> 00:05:49,039 Speaker 1: but that's kind of all they could do. They couldn't 102 00:05:49,040 --> 00:05:52,359 Speaker 1: make anything up or talk to you or write anything 103 00:05:52,440 --> 00:05:56,400 Speaker 1: for you. They were smart, but they were passive. But 104 00:05:56,400 --> 00:06:00,040 Speaker 1: then in twenty seventeen, some engineers at Google discovered a 105 00:06:00,200 --> 00:06:05,000 Speaker 1: special way to make these networks of neurons that changed everything. 106 00:06:05,680 --> 00:06:07,800 Speaker 1: And it kind of started when they said, you know what, 107 00:06:08,200 --> 00:06:11,719 Speaker 1: maybe we should stop trying to copy the human brain. 108 00:06:12,480 --> 00:06:14,840 Speaker 3: So at the very beginning, people try to come up 109 00:06:14,839 --> 00:06:18,640 Speaker 3: with very simple computer models, like the perceptron that tries 110 00:06:18,680 --> 00:06:22,000 Speaker 3: to mimic the human brain, and then gradually this neural 111 00:06:22,120 --> 00:06:26,320 Speaker 3: network or deep learning morphed into something completely different that 112 00:06:26,560 --> 00:06:28,520 Speaker 3: is no longer trying to mimic the human brain. 113 00:06:29,839 --> 00:06:32,760 Speaker 1: Yeah, you can imagine that if the History of the 114 00:06:32,839 --> 00:06:35,440 Speaker 1: Human Race was a movie, this might be the part 115 00:06:35,480 --> 00:06:38,599 Speaker 1: of the movie that foreshadows that it's not gonna end. Well, 116 00:06:42,760 --> 00:06:46,279 Speaker 1: what the engineers at Google had invented was something called 117 00:06:46,600 --> 00:06:50,919 Speaker 1: the transformer. In this case, it's not related to the 118 00:06:51,080 --> 00:06:51,760 Speaker 1: robot toys. 119 00:06:51,960 --> 00:06:54,520 Speaker 2: It is not related to the robot toys at all. 120 00:06:55,400 --> 00:06:56,880 Speaker 2: I guess I've been in the AI. 121 00:06:56,680 --> 00:06:59,200 Speaker 3: World for so long that transformers used to me in 122 00:06:59,240 --> 00:07:01,360 Speaker 3: the movie and me Tron and all that souff now 123 00:07:01,440 --> 00:07:03,039 Speaker 3: means something completely different to me. 124 00:07:03,880 --> 00:07:05,159 Speaker 1: They reprogrammed your brain. 125 00:07:05,320 --> 00:07:08,880 Speaker 2: They reprogrammed my brain. Yeah, okay. 126 00:07:09,000 --> 00:07:12,480 Speaker 1: What this transformer way of connecting neural networks that the 127 00:07:12,520 --> 00:07:16,120 Speaker 1: Google engineers invented in twenty seventeen does is that it 128 00:07:16,280 --> 00:07:21,080 Speaker 1: uses an old ticnique from language translation called attention. And 129 00:07:21,240 --> 00:07:25,080 Speaker 1: basically what it means is that when an AI is learning, say, 130 00:07:25,240 --> 00:07:27,760 Speaker 1: a piece of text from a book, it doesn't just 131 00:07:27,880 --> 00:07:31,720 Speaker 1: learn what each word means. It learns the context of 132 00:07:31,760 --> 00:07:34,720 Speaker 1: the word. For each word, it sort of learns what 133 00:07:34,800 --> 00:07:38,560 Speaker 1: word it usually has after it and before it, and 134 00:07:38,640 --> 00:07:41,480 Speaker 1: it also learns what word it usually has two words 135 00:07:41,520 --> 00:07:44,520 Speaker 1: after it and two words before it, and so on 136 00:07:44,920 --> 00:07:47,560 Speaker 1: and so on. Let's here's an example. So like the 137 00:07:47,560 --> 00:07:50,880 Speaker 1: word sit from sitting, usually it comes a few words 138 00:07:50,920 --> 00:07:52,560 Speaker 1: before chair or sofa. 139 00:07:52,320 --> 00:07:55,720 Speaker 3: Right right right, So I sit on a sofa. I 140 00:07:55,760 --> 00:07:58,920 Speaker 3: sit on a sofa is very logical, very sensible. The 141 00:07:59,040 --> 00:08:02,480 Speaker 3: key work that is important in terms of me understanding 142 00:08:02,600 --> 00:08:05,600 Speaker 3: what sofa means is the I in the sentence. So 143 00:08:05,640 --> 00:08:07,280 Speaker 3: it's not a dog sitting on the sofa, it's not 144 00:08:07,440 --> 00:08:10,679 Speaker 3: cat sitting on the sofa, it's I that is sitting 145 00:08:10,720 --> 00:08:13,560 Speaker 3: on the sofa. And so the attention mechanism allows you 146 00:08:13,840 --> 00:08:17,040 Speaker 3: to kind of learn that type of correlation. And there 147 00:08:17,080 --> 00:08:20,720 Speaker 3: are technical details about position encodings, but the main thing 148 00:08:20,800 --> 00:08:22,880 Speaker 3: is just you want to learn the meaning of a 149 00:08:22,920 --> 00:08:27,080 Speaker 3: word in the context of the words that surround it. 150 00:08:27,160 --> 00:08:30,000 Speaker 1: This sort of simple and kind of common sense sounding 151 00:08:30,040 --> 00:08:34,320 Speaker 1: trick is basically what revolutionized AI. You can apply it 152 00:08:34,360 --> 00:08:42,400 Speaker 1: to text, images, music videos, and it's what makes chat GPT, Gemini, Grock, 153 00:08:42,800 --> 00:08:46,800 Speaker 1: mid Journey, Claude, Nano, Banana, Copilot and all those AI 154 00:08:46,880 --> 00:08:51,080 Speaker 1: agents out there work. And that's because using this transformer 155 00:08:51,200 --> 00:08:54,640 Speaker 1: architecture and paying attention to the context of the data 156 00:08:55,040 --> 00:08:57,640 Speaker 1: clout AIS to teach themselves. 157 00:09:00,120 --> 00:09:02,680 Speaker 2: Yeah, so it's called self supervised learning. 158 00:09:03,040 --> 00:09:04,160 Speaker 1: Oh that sounds scary. 159 00:09:04,559 --> 00:09:09,640 Speaker 2: Yeah. Yeah. So traditional machine learning or AI systems, they 160 00:09:09,640 --> 00:09:13,520 Speaker 2: rely on supervised learning, which means that a human has 161 00:09:13,559 --> 00:09:16,280 Speaker 2: to tell you what is right or what is wrong. Now, 162 00:09:16,320 --> 00:09:20,160 Speaker 2: while that's great, it's not very scalable because a human 163 00:09:20,200 --> 00:09:22,679 Speaker 2: can only tell you what's right or wrong for a 164 00:09:22,800 --> 00:09:25,400 Speaker 2: thousand things or two thousand things or whatnot. You can't 165 00:09:25,400 --> 00:09:28,360 Speaker 2: scale it to a billion things. Whereas with this self 166 00:09:28,400 --> 00:09:33,280 Speaker 2: supervised paradigm, you don't need human supervision. You just need 167 00:09:33,320 --> 00:09:38,280 Speaker 2: some large purposes of unstructured data, and then the machine 168 00:09:38,320 --> 00:09:41,280 Speaker 2: teaches itself just by predicting what the next token is 169 00:09:41,480 --> 00:09:44,360 Speaker 2: and seeing whether that matches reality in terms of the 170 00:09:44,400 --> 00:09:45,480 Speaker 2: real data. 171 00:09:45,679 --> 00:09:48,360 Speaker 1: What doctor Tam saying is that this new type of 172 00:09:48,440 --> 00:09:52,319 Speaker 1: AI could teach itself by predicting what the next word 173 00:09:52,760 --> 00:09:55,520 Speaker 1: or part of an image should be according to its 174 00:09:55,559 --> 00:10:00,240 Speaker 1: context and then checking with the real thing. And do 175 00:10:00,400 --> 00:10:05,080 Speaker 1: ability of AIS to learn contexts and teach themselves, combined 176 00:10:05,120 --> 00:10:08,959 Speaker 1: with big companies putting billions of dollars into huge computer 177 00:10:09,040 --> 00:10:13,200 Speaker 1: farms meant that you could suddenly have ais teach themselves 178 00:10:13,480 --> 00:10:15,240 Speaker 1: the entire Internet. 179 00:10:16,559 --> 00:10:21,040 Speaker 3: So the idea is that very surprising things happen when 180 00:10:21,200 --> 00:10:25,600 Speaker 3: you scale something to really really make sizes. They perform 181 00:10:25,880 --> 00:10:30,080 Speaker 3: way way better in terms of producing realistic samples and texts. 182 00:10:30,679 --> 00:10:33,199 Speaker 3: So there is some type of phase transition in terms 183 00:10:33,240 --> 00:10:36,920 Speaker 3: of how intelligent or how realistic the samples that it 184 00:10:37,000 --> 00:10:39,079 Speaker 3: generates will look like something. 185 00:10:39,920 --> 00:10:44,400 Speaker 1: Click. It suddenly activated some potential in these systems that 186 00:10:44,720 --> 00:10:50,520 Speaker 1: experts were not expecting. Right. So t wecap ais were 187 00:10:50,559 --> 00:10:53,719 Speaker 1: originally meant to copy the human brain, and they were 188 00:10:53,800 --> 00:10:58,480 Speaker 1: really good at just recognizing things patterns, faces, numbers. But 189 00:10:58,520 --> 00:11:02,280 Speaker 1: then some engineers at Google decided, you know what, let's 190 00:11:02,320 --> 00:11:05,800 Speaker 1: forget the human brain and just push this through the limit. 191 00:11:06,120 --> 00:11:08,680 Speaker 1: And what they came up with created ais they could 192 00:11:08,840 --> 00:11:14,040 Speaker 1: understand context, and more importantly, they could teach themselves. And 193 00:11:14,200 --> 00:11:17,800 Speaker 1: once that happened, they scaled up, gubbled up the Internet 194 00:11:18,000 --> 00:11:20,680 Speaker 1: and made a leap that's making a lot of people, 195 00:11:21,240 --> 00:11:25,320 Speaker 1: uh concerned right now. So when we come back, we're 196 00:11:25,360 --> 00:11:28,360 Speaker 1: going to talk about how these new ais can affect 197 00:11:28,400 --> 00:11:31,480 Speaker 1: your brain, and how they could even make endangered species 198 00:11:31,520 --> 00:11:35,800 Speaker 1: of animals go to think, don't go anywhere, We'll be 199 00:11:36,000 --> 00:11:51,720 Speaker 1: right back. Hey, welcome back. We're talking about AI slop 200 00:11:52,120 --> 00:11:55,080 Speaker 1: or AI made images and videos that are swamping the 201 00:11:55,120 --> 00:11:58,680 Speaker 1: Internet right now, and what effect it's having on us 202 00:11:59,280 --> 00:12:02,840 Speaker 1: so far. Talking about what happened in twenty seventeen that 203 00:12:03,040 --> 00:12:06,280 Speaker 1: suddenly made these AI systems be able to talk to 204 00:12:06,320 --> 00:12:09,520 Speaker 1: you and make the kind of realistic images everyone sees 205 00:12:09,640 --> 00:12:12,920 Speaker 1: in their social media feeds. Now we're gonna talk about 206 00:12:12,920 --> 00:12:16,200 Speaker 1: the effect it's having on our brains. Is it possible 207 00:12:16,240 --> 00:12:20,720 Speaker 1: for AI generated content to rewire how we think? To 208 00:12:20,760 --> 00:12:22,559 Speaker 1: tell us about that, I reached out to my friend, 209 00:12:22,760 --> 00:12:27,240 Speaker 1: neuroscientist Dwayne Godwin, a professor at Wake Forest University and 210 00:12:27,280 --> 00:12:29,560 Speaker 1: the co author of the book Out of Your Mind, 211 00:12:29,840 --> 00:12:33,160 Speaker 1: which he co wrote with me. Doctor Godwin recently wrote 212 00:12:33,160 --> 00:12:35,960 Speaker 1: an article about this topic online, so I thought he 213 00:12:36,080 --> 00:12:38,920 Speaker 1: was the perfect person to give us the neural perspective 214 00:12:39,080 --> 00:12:45,760 Speaker 1: on this topic. Here's my conversation with doctor Dwayne Godwin. Hey, Dwayne, 215 00:12:45,760 --> 00:12:46,200 Speaker 1: how are you? 216 00:12:46,559 --> 00:12:48,320 Speaker 4: I'm doing well? Orgey, how are you? 217 00:12:48,800 --> 00:12:50,920 Speaker 1: I'm good? Now? Just to make sure this is the 218 00:12:50,960 --> 00:12:53,040 Speaker 1: real you, right, I'm not talking to an AI version 219 00:12:53,080 --> 00:12:53,600 Speaker 1: of you. 220 00:12:53,640 --> 00:12:55,880 Speaker 4: No, I AMZ I'm the real deal man. 221 00:12:57,679 --> 00:13:00,959 Speaker 1: And is there an intelligence artificial or is organic? 222 00:13:01,200 --> 00:13:04,000 Speaker 4: It's augmented by caffeine. 223 00:13:05,320 --> 00:13:08,920 Speaker 1: Augmented intelligence? There you go. Well, I wonder if there's 224 00:13:08,920 --> 00:13:11,240 Speaker 1: a way for people to know if this conversation is 225 00:13:11,320 --> 00:13:14,400 Speaker 1: generated by one of those you know, chat GBT things 226 00:13:14,400 --> 00:13:15,520 Speaker 1: that mix podcasts. 227 00:13:15,679 --> 00:13:17,760 Speaker 4: You know, if we were really sneaky, we might have 228 00:13:18,000 --> 00:13:20,760 Speaker 4: generated an AI podcast of the two of us going 229 00:13:20,800 --> 00:13:24,880 Speaker 4: back and forth and then tested our audience in their 230 00:13:24,920 --> 00:13:28,199 Speaker 4: ability to tell the difference. But we're not that sneaky. 231 00:13:28,720 --> 00:13:29,600 Speaker 1: Of course, this is. 232 00:13:29,520 --> 00:13:30,120 Speaker 4: The real thing. 233 00:13:30,240 --> 00:13:32,320 Speaker 1: All right, let's get into it. I guess the first 234 00:13:32,400 --> 00:13:34,960 Speaker 1: question I had was how good are we at telling 235 00:13:35,080 --> 00:13:36,640 Speaker 1: is something is AI generated? 236 00:13:36,960 --> 00:13:41,400 Speaker 4: We're not very good. People tend to overestimate their AI radar. 237 00:13:41,760 --> 00:13:45,600 Speaker 4: A lot of obvious tells, like you know, extra fingers, 238 00:13:45,920 --> 00:13:49,920 Speaker 4: robotic phrasing, those are getting all smoothed out. And we're 239 00:13:49,960 --> 00:13:53,520 Speaker 4: looking at these things often on very tiny screens, and 240 00:13:53,640 --> 00:13:58,280 Speaker 4: so what's plausible quickly becomes good enough or believable wow, 241 00:13:58,440 --> 00:14:01,160 Speaker 4: And the result of that is there's kind of a 242 00:14:01,200 --> 00:14:04,640 Speaker 4: confidence gap. We feel very certain that we can tell 243 00:14:04,720 --> 00:14:07,640 Speaker 4: fake stuff when we see it, but often we're not 244 00:14:07,679 --> 00:14:08,920 Speaker 4: doing much better than chance. 245 00:14:09,320 --> 00:14:11,800 Speaker 1: Are there any studies that kind of tell us how 246 00:14:11,880 --> 00:14:14,319 Speaker 1: good we are at spotting fake ais? 247 00:14:14,520 --> 00:14:16,959 Speaker 4: Yeah, there have been studies that have been done. There 248 00:14:17,080 --> 00:14:20,680 Speaker 4: was a study where people are essentially two groups. They're 249 00:14:20,800 --> 00:14:25,440 Speaker 4: presented with both real images and AI generated images, and 250 00:14:25,480 --> 00:14:28,400 Speaker 4: it turns out that they're about fifty percent you know, 251 00:14:28,440 --> 00:14:30,800 Speaker 4: it's like fifty to fifty that they can tell the difference. 252 00:14:30,920 --> 00:14:32,560 Speaker 1: Basically, they can't tell the difference. 253 00:14:32,720 --> 00:14:35,400 Speaker 4: Yeah, so they're really about at the level of chance. 254 00:14:35,680 --> 00:14:39,080 Speaker 1: Yikes. Okay, and this was probably done a few couple 255 00:14:39,120 --> 00:14:40,080 Speaker 1: of years ago, right. 256 00:14:40,160 --> 00:14:43,080 Speaker 4: Yeah, that's right, And so things are only better now. 257 00:14:43,160 --> 00:14:47,320 Speaker 4: So you can imagine that right now, with the quality 258 00:14:47,560 --> 00:14:50,760 Speaker 4: of generative AI, that this is even going to be 259 00:14:50,800 --> 00:14:52,480 Speaker 4: more true. 260 00:14:52,760 --> 00:14:55,000 Speaker 1: Yes, we all like to think we're good at spotting 261 00:14:55,040 --> 00:14:58,600 Speaker 1: fake images or videos. I mean, we've all seen bad 262 00:14:58,600 --> 00:15:01,920 Speaker 1: computer graphic effects movies, so we think we know what 263 00:15:02,040 --> 00:15:04,920 Speaker 1: fake looks like. But actually the sign says that we 264 00:15:04,960 --> 00:15:08,360 Speaker 1: are not much better than guessing randomly if something is 265 00:15:08,400 --> 00:15:12,160 Speaker 1: fake or not. And according to doctor Godwin, there are 266 00:15:12,320 --> 00:15:16,240 Speaker 1: three reasons our brains are kind of helpless against this 267 00:15:16,440 --> 00:15:20,280 Speaker 1: onslaught of fake content coming at us on a daily basis. 268 00:15:20,920 --> 00:15:23,720 Speaker 1: The first reason is that your brain sort of wants 269 00:15:23,760 --> 00:15:28,600 Speaker 1: to believe things are real. I guess, as a neuroscientists 270 00:15:28,600 --> 00:15:31,120 Speaker 1: and as a psychologist, why do you think our brains 271 00:15:31,240 --> 00:15:32,600 Speaker 1: are so easy to fool? 272 00:15:33,040 --> 00:15:35,640 Speaker 4: Well, a couple of reasons. You know, our brains are 273 00:15:35,680 --> 00:15:40,880 Speaker 4: built for speed and survival and not for fact checking. 274 00:15:41,080 --> 00:15:43,360 Speaker 1: Right, And I guess that means out in the wild, 275 00:15:43,520 --> 00:15:45,880 Speaker 1: if you think you see a bear or a tiger 276 00:15:45,960 --> 00:15:47,840 Speaker 1: coming at you, your brain is not going to have 277 00:15:47,840 --> 00:15:49,920 Speaker 1: wanted like fact check that there really is a tiger 278 00:15:49,960 --> 00:15:52,400 Speaker 1: coming at you. You want your brain to just react 279 00:15:52,400 --> 00:15:54,400 Speaker 1: to the idea that maybe there's a tiger coming at you. 280 00:15:54,520 --> 00:15:57,720 Speaker 4: Yes, that's sedaptive. You're going to go with bear if 281 00:15:57,720 --> 00:15:59,000 Speaker 4: it looks anything. 282 00:16:00,640 --> 00:16:02,920 Speaker 1: Good. Because I guess we don't have all information all 283 00:16:02,960 --> 00:16:05,800 Speaker 1: the time, so our brains are sort of have trained to, like, 284 00:16:05,960 --> 00:16:07,840 Speaker 1: do as much as you can with as little information 285 00:16:07,880 --> 00:16:08,480 Speaker 1: as possible. 286 00:16:08,600 --> 00:16:11,880 Speaker 4: Yeah, there's a concept in psychology called heuristics, which is 287 00:16:11,960 --> 00:16:16,120 Speaker 4: basically making decisions on the basis of very limited information, 288 00:16:16,200 --> 00:16:19,000 Speaker 4: and our brains are doing that constantly, so you know, 289 00:16:19,040 --> 00:16:21,200 Speaker 4: we have to make those very quick decisions, so we 290 00:16:21,240 --> 00:16:23,560 Speaker 4: can pass our genes onto the next generation. 291 00:16:24,040 --> 00:16:26,200 Speaker 1: I see, and so you can exploit that. I guess 292 00:16:26,280 --> 00:16:28,960 Speaker 1: because we see something that sort of looks real, your 293 00:16:28,960 --> 00:16:31,520 Speaker 1: mind will basically make the conclusion that it is real. 294 00:16:31,720 --> 00:16:34,480 Speaker 4: Yeah, it will create the story that it is real. 295 00:16:34,720 --> 00:16:36,960 Speaker 4: You know, what is the purpose of our brains. Our 296 00:16:37,000 --> 00:16:41,080 Speaker 4: brains are prediction machines, and so we've learned certain facts 297 00:16:41,120 --> 00:16:44,520 Speaker 4: about the world, about how things normally go. So if 298 00:16:44,520 --> 00:16:47,280 Speaker 4: you're in the woods and you encounter something it seems 299 00:16:47,360 --> 00:16:49,960 Speaker 4: like a bear, your brain already knows that that's probably 300 00:16:50,000 --> 00:16:52,440 Speaker 4: not a great thing, and so you're going to extrapolate 301 00:16:52,560 --> 00:16:55,640 Speaker 4: that story and you're going to want to avoid that 302 00:16:55,680 --> 00:16:57,600 Speaker 4: bear as much as possible, right. 303 00:16:57,760 --> 00:17:01,360 Speaker 1: I guess we kind of had a natural mechanism against that, 304 00:17:01,840 --> 00:17:04,320 Speaker 1: which was this idea of the Uncanny Valley. 305 00:17:04,520 --> 00:17:07,240 Speaker 4: Yeah, I'm familiar with the Uncanny Valley. It's that sort 306 00:17:07,280 --> 00:17:09,639 Speaker 4: of eerie feeling that you get when you see something 307 00:17:09,680 --> 00:17:11,760 Speaker 4: that's trying to be real but it's not real. It 308 00:17:11,800 --> 00:17:14,440 Speaker 4: makes me uncomfortable because I know that it's not supposed 309 00:17:14,480 --> 00:17:15,960 Speaker 4: to look like that exactly. 310 00:17:16,119 --> 00:17:18,199 Speaker 1: Yeah, So, like we did have a brain mechanism to 311 00:17:18,280 --> 00:17:22,000 Speaker 1: defense against fake things. Right, But it seems like AI 312 00:17:22,040 --> 00:17:25,440 Speaker 1: today has basically blown past that defense, like it's now 313 00:17:25,920 --> 00:17:27,320 Speaker 1: past the uncanny valley. 314 00:17:27,600 --> 00:17:31,119 Speaker 4: Yeah, it has jumped across the uncanny valley for sure. 315 00:17:32,240 --> 00:17:35,520 Speaker 1: In other words, our brain has a built in BS 316 00:17:35,560 --> 00:17:39,479 Speaker 1: meter called the uncanny valley. But now AI content is 317 00:17:39,520 --> 00:17:42,600 Speaker 1: so good it's blown past that last line of defense 318 00:17:42,840 --> 00:17:45,800 Speaker 1: or your brain. Okay. The second reason our brains are 319 00:17:45,880 --> 00:17:49,640 Speaker 1: easy to fool by AI generated content is that it's 320 00:17:49,680 --> 00:17:52,639 Speaker 1: not built to withstand the constant stream of it we 321 00:17:52,800 --> 00:17:56,879 Speaker 1: get every day. The second thing you said was repetition. 322 00:17:57,200 --> 00:18:00,560 Speaker 4: Yeah, that's right. So your resistance is being reduced in 323 00:18:00,640 --> 00:18:03,640 Speaker 4: two ways. So first is just the repetition of seeing 324 00:18:03,720 --> 00:18:05,280 Speaker 4: the same thing over and over. 325 00:18:05,640 --> 00:18:08,919 Speaker 1: Meaning like, if I'm scrolling through these online feeds and 326 00:18:09,359 --> 00:18:11,520 Speaker 1: most of what I see has that sort of feeling 327 00:18:11,520 --> 00:18:13,959 Speaker 1: of an AI generated thing, then my brain is just 328 00:18:14,000 --> 00:18:16,439 Speaker 1: gonna normalize it. Is that kind of what you're saying, Like, 329 00:18:16,480 --> 00:18:17,520 Speaker 1: it's yeah, it is. 330 00:18:17,800 --> 00:18:21,080 Speaker 4: There's another more insidious part of that, which is, let's 331 00:18:21,200 --> 00:18:24,680 Speaker 4: use the you know, the cute animal AI generated cute 332 00:18:24,720 --> 00:18:29,040 Speaker 4: animal videos you'll see in the comments underneath those people 333 00:18:29,119 --> 00:18:32,840 Speaker 4: reacting to it as if it's real, and so part 334 00:18:32,920 --> 00:18:36,920 Speaker 4: of that is, Hey, everybody else thinks it's real, and 335 00:18:37,119 --> 00:18:39,879 Speaker 4: I'm a person too, shouldn't I be thinking that this 336 00:18:40,040 --> 00:18:43,840 Speaker 4: is real? And so there's this sort of desire on 337 00:18:43,960 --> 00:18:47,399 Speaker 4: the part of people to reach a consensus with your group, 338 00:18:48,080 --> 00:18:50,879 Speaker 4: and so if you start identifying with that group, then 339 00:18:51,000 --> 00:18:54,640 Speaker 4: your brain might be bent toward making a decision that, hey, 340 00:18:54,760 --> 00:18:57,760 Speaker 4: maybe this is real, even if you might have had 341 00:18:57,800 --> 00:19:01,040 Speaker 4: initial doubts. Yeah, maybe you start to fall into line. 342 00:19:01,720 --> 00:19:04,159 Speaker 1: Like we all like to think we're independent thinkers and 343 00:19:04,200 --> 00:19:07,159 Speaker 1: critical thinkers, but part of our brain is sort of 344 00:19:07,240 --> 00:19:09,359 Speaker 1: wired to kind of basically follow the mob. 345 00:19:09,800 --> 00:19:12,639 Speaker 4: Yeah, a little bit of mob mentality. Yeah, we like 346 00:19:12,680 --> 00:19:17,040 Speaker 4: to think of ourselves as logical beings, but the reality 347 00:19:17,200 --> 00:19:22,840 Speaker 4: is we're all challenged by biases, and those biases can 348 00:19:22,960 --> 00:19:26,560 Speaker 4: be very adaptive in the case of things like heuristics. 349 00:19:27,840 --> 00:19:30,840 Speaker 1: And then the third reason our brains are easy targets 350 00:19:30,880 --> 00:19:33,520 Speaker 1: for AI slob is that a lot of that SLOB 351 00:19:33,680 --> 00:19:37,280 Speaker 1: is made to play with our emotions. And then you 352 00:19:37,280 --> 00:19:40,639 Speaker 1: said the third one was that we tend to believe 353 00:19:40,760 --> 00:19:42,920 Speaker 1: things that kind of affect our emotions. 354 00:19:43,320 --> 00:19:46,879 Speaker 4: Yes, that's probably one of the most insidious things, is 355 00:19:46,920 --> 00:19:52,120 Speaker 4: that things that ring our emotional bell tend to make 356 00:19:52,160 --> 00:19:56,479 Speaker 4: it through without the kind of critical thought processing that 357 00:19:56,520 --> 00:20:00,960 Speaker 4: we would normally apply to something that's just factual or techation. 358 00:20:01,880 --> 00:20:06,640 Speaker 4: So think about politics. Politics is very polarizing. So if 359 00:20:06,720 --> 00:20:10,200 Speaker 4: you see something in your social media feed and it 360 00:20:10,359 --> 00:20:14,879 Speaker 4: confirms what your emotions tell you should be true based 361 00:20:14,920 --> 00:20:18,639 Speaker 4: on your political leanings, then that we'll get through the 362 00:20:18,720 --> 00:20:23,080 Speaker 4: normal cognitive criticism that you would apply to that information. 363 00:20:23,800 --> 00:20:26,640 Speaker 1: Yikes, Like, we let our emotions get in the way 364 00:20:26,680 --> 00:20:31,199 Speaker 1: of our critical thinking, so our brains are ill equipped 365 00:20:31,280 --> 00:20:34,320 Speaker 1: to deal with all this AI slot because it tends 366 00:20:34,359 --> 00:20:37,560 Speaker 1: to believe what it sees. We're inundated with this content 367 00:20:38,000 --> 00:20:41,600 Speaker 1: and it's all meant to trigger our emotions and things 368 00:20:41,720 --> 00:20:45,040 Speaker 1: doctor Godwin says are just getting worse. 369 00:20:46,400 --> 00:20:50,840 Speaker 4: They're getting worse because it's getting better. There's a lot 370 00:20:50,960 --> 00:20:55,600 Speaker 4: of money being put into making these things basically indistinguishable 371 00:20:55,600 --> 00:20:59,160 Speaker 4: from reality, and we were already suffering under the idea 372 00:20:59,359 --> 00:21:02,880 Speaker 4: that we could possibly tell the difference, but it's becoming 373 00:21:02,920 --> 00:21:05,600 Speaker 4: a parent that we cannot, and it's going to be 374 00:21:05,680 --> 00:21:06,440 Speaker 4: even harder. 375 00:21:06,920 --> 00:21:09,120 Speaker 1: It kind of seems like it's also getting worse because 376 00:21:09,480 --> 00:21:12,480 Speaker 1: it seems like people are more willing to post fake 377 00:21:12,600 --> 00:21:16,280 Speaker 1: AI generated things, like just the other day, the White 378 00:21:16,359 --> 00:21:20,800 Speaker 1: House official account posted some AI altered image of some 379 00:21:20,960 --> 00:21:24,840 Speaker 1: activist protester, and they were totally unapologetic about it. 380 00:21:25,200 --> 00:21:29,679 Speaker 4: Yeah, the tools are more readily available, and your ability 381 00:21:29,720 --> 00:21:32,600 Speaker 4: to manipulate those tools doesn't even require that you have 382 00:21:32,680 --> 00:21:34,480 Speaker 4: any sort of computer science degree. 383 00:21:34,840 --> 00:21:36,480 Speaker 1: Yeah, it's just an app on your phone. 384 00:21:36,880 --> 00:21:39,600 Speaker 4: Yeah. This stuff is scalable too, So it's not just 385 00:21:39,720 --> 00:21:41,639 Speaker 4: that a person can go in and do like a 386 00:21:41,680 --> 00:21:45,159 Speaker 4: one off. A single person could generate a lot of 387 00:21:45,200 --> 00:21:48,800 Speaker 4: this content and get it out there, and even use 388 00:21:49,040 --> 00:21:53,119 Speaker 4: AI to generate bots that would distribute that content for you. 389 00:21:53,760 --> 00:21:54,040 Speaker 2: Yeah. 390 00:21:54,119 --> 00:21:56,480 Speaker 4: So it's only going to get worse, I think. 391 00:21:57,520 --> 00:22:00,040 Speaker 1: Yeah, it's not looking so good for us organic and 392 00:22:00,080 --> 00:22:04,639 Speaker 1: they naturally made squishy intelligences out here. But according to 393 00:22:04,680 --> 00:22:07,679 Speaker 1: doctor Godwin, there are some things you can do to 394 00:22:07,720 --> 00:22:11,480 Speaker 1: protect your neurons. What can we do to protect our 395 00:22:11,560 --> 00:22:14,800 Speaker 1: brains from the defect of this AI slut? 396 00:22:15,240 --> 00:22:20,119 Speaker 4: Well, it's hard. I wish I had an easy answer, 397 00:22:20,160 --> 00:22:23,240 Speaker 4: because what I've told you so far it's pretty hopeless. Right, 398 00:22:23,440 --> 00:22:24,480 Speaker 4: it sounds hopeless. 399 00:22:24,640 --> 00:22:26,800 Speaker 1: Maybe I should look it up and chat GBT. 400 00:22:26,720 --> 00:22:28,680 Speaker 4: But I think there are things that we can do, 401 00:22:29,040 --> 00:22:31,600 Speaker 4: and part of it is just slowing down. The thing 402 00:22:31,680 --> 00:22:34,560 Speaker 4: we can't do is there's no way we're going to 403 00:22:34,600 --> 00:22:39,239 Speaker 4: become a forensic AI detector. So what we have to 404 00:22:39,280 --> 00:22:44,640 Speaker 4: do is really add some friction, build some simple verification habits, 405 00:22:45,000 --> 00:22:48,639 Speaker 4: and try to shrink our exposure to this fire hose 406 00:22:48,880 --> 00:22:52,560 Speaker 4: of bad stuff. I think a good rule is if 407 00:22:52,600 --> 00:22:55,880 Speaker 4: it hits you emotionally in the first two seconds, then 408 00:22:56,040 --> 00:23:00,640 Speaker 4: just pause and then just ask the question, Okay, made this, 409 00:23:01,000 --> 00:23:03,760 Speaker 4: where did it come from? And can I confirm that 410 00:23:03,840 --> 00:23:07,080 Speaker 4: this is real? Somewhere boring and reputable. 411 00:23:07,600 --> 00:23:10,000 Speaker 1: I see, look for sources that you can trust, and 412 00:23:10,200 --> 00:23:13,800 Speaker 1: maybe shift your consumption from the wild west of social 413 00:23:13,840 --> 00:23:17,560 Speaker 1: media to some of these sources that you trust more 414 00:23:17,680 --> 00:23:20,680 Speaker 1: and that are known to give you objective truth. 415 00:23:21,200 --> 00:23:24,920 Speaker 4: That's right. So if something seems to perfectly confirm your 416 00:23:24,960 --> 00:23:29,199 Speaker 4: own biases, that's the time to tell yourself, you know, 417 00:23:29,760 --> 00:23:32,280 Speaker 4: I need to dig a little deeper before I accept this. 418 00:23:32,760 --> 00:23:35,560 Speaker 1: I see, check other websites, maybe even the ones that 419 00:23:35,600 --> 00:23:36,359 Speaker 1: you don't agree with. 420 00:23:36,680 --> 00:23:39,720 Speaker 4: That's right. Part of it is if it's real, then 421 00:23:39,760 --> 00:23:43,080 Speaker 4: it should exist somewhere other than your social media feed. 422 00:23:43,680 --> 00:23:46,080 Speaker 4: And I think the other part of this is there 423 00:23:46,119 --> 00:23:49,840 Speaker 4: is an algorithm that is trying to cater information to you. 424 00:23:49,920 --> 00:23:53,760 Speaker 4: It's called a social media feed for a reason, because 425 00:23:53,800 --> 00:23:55,680 Speaker 4: it's being fed to you. 426 00:23:56,080 --> 00:24:00,760 Speaker 1: Right, I see, ye realize that it's a feed. 427 00:24:01,200 --> 00:24:04,240 Speaker 4: So you have a choice. In x dot com, for example, 428 00:24:04,359 --> 00:24:07,000 Speaker 4: it'll have a curated feed, and then it'll have a 429 00:24:07,080 --> 00:24:10,320 Speaker 4: feed of people that you follow, and I always turn 430 00:24:10,400 --> 00:24:14,800 Speaker 4: off the curated feed and only go for the accounts 431 00:24:14,800 --> 00:24:16,719 Speaker 4: that I have actually chosen to follow. 432 00:24:17,200 --> 00:24:20,440 Speaker 1: Yeah, well that's a good one. Well, hopefully people will listen. 433 00:24:20,640 --> 00:24:27,280 Speaker 1: But here's the twist, ending, Dwayne, this was AI generated. Yes, No, 434 00:24:27,480 --> 00:24:30,040 Speaker 1: we're just kidding. No, we actually had this conversation. How 435 00:24:30,040 --> 00:24:32,280 Speaker 1: can we prove it? Dwayne, say something that an AI 436 00:24:32,320 --> 00:24:32,960 Speaker 1: would never say? 437 00:24:33,720 --> 00:24:37,359 Speaker 4: What an AI would never say? Never listen to an AI? 438 00:24:38,119 --> 00:24:45,120 Speaker 1: All right, let me come back. We're going to talk 439 00:24:45,160 --> 00:24:48,480 Speaker 1: about another way in which AI slob is affecting us, 440 00:24:48,640 --> 00:24:51,879 Speaker 1: one that is kind of unexpected. That is through the 441 00:24:52,000 --> 00:24:57,080 Speaker 1: insidious use of pictures of cute animals. You think they're harmless, 442 00:24:57,119 --> 00:24:59,639 Speaker 1: but actually they could lead to a lot of death 443 00:24:59,760 --> 00:25:03,320 Speaker 1: and destruction. So stay with us. We'll be right back. 444 00:25:16,960 --> 00:25:20,879 Speaker 1: Welcome back. I am not an AI or am I 445 00:25:22,280 --> 00:25:25,879 Speaker 1: We're talking about the effects AI slop or AI generated 446 00:25:25,960 --> 00:25:29,040 Speaker 1: content has on us. And so far we learn how 447 00:25:29,080 --> 00:25:32,120 Speaker 1: AI slop works and why it's so easy for our 448 00:25:32,200 --> 00:25:35,879 Speaker 1: brains to believe it's real. Now we're going to explore 449 00:25:35,920 --> 00:25:39,320 Speaker 1: how we can affect our actions by focusing on one 450 00:25:39,359 --> 00:25:44,320 Speaker 1: specific example, which is pictures and videos of cute animals. 451 00:25:46,920 --> 00:25:49,480 Speaker 1: You've probably seen this in your social media feed. Some 452 00:25:49,600 --> 00:25:53,359 Speaker 1: cute little hummingbirds taking shelter inside a flower, or a 453 00:25:53,440 --> 00:25:57,200 Speaker 1: cute kitten plucking gummy bears off a gummy bear tree, 454 00:25:57,680 --> 00:26:02,280 Speaker 1: or a cow scratching itself with the room. Actually one 455 00:26:02,280 --> 00:26:05,000 Speaker 1: of those is real, but the fact that you don't 456 00:26:05,040 --> 00:26:07,320 Speaker 1: know which of them it is is kind of the 457 00:26:07,359 --> 00:26:10,280 Speaker 1: whole problem, and, according to the next person I talked to, 458 00:26:10,880 --> 00:26:15,840 Speaker 1: can kind of be a serious problem with potentially dire consequences. 459 00:26:16,200 --> 00:26:20,240 Speaker 1: Ketterina Zimmer is a science journalist who writes about biology 460 00:26:20,320 --> 00:26:23,320 Speaker 1: and the environment, and she recently published an article about 461 00:26:23,400 --> 00:26:28,680 Speaker 1: aislop in Atmos magazine. Well, thank you, Katrina for joining us. 462 00:26:29,080 --> 00:26:30,159 Speaker 5: Thanks so much for having me. 463 00:26:30,560 --> 00:26:32,240 Speaker 1: Now, how can I be sure that you're real and 464 00:26:32,320 --> 00:26:33,360 Speaker 1: not AI generated? 465 00:26:34,760 --> 00:26:37,040 Speaker 5: That's a good question. I guess you'll have to trust 466 00:26:37,119 --> 00:26:37,439 Speaker 5: me on this. 467 00:26:37,480 --> 00:26:37,680 Speaker 2: One. 468 00:26:38,359 --> 00:26:41,679 Speaker 1: Oh no, Well, you've written a lot about the impact 469 00:26:41,720 --> 00:26:45,720 Speaker 1: of AI and AI generated images on us, but from 470 00:26:45,760 --> 00:26:48,040 Speaker 1: a different angle, So tell us about the article that 471 00:26:48,080 --> 00:26:48,520 Speaker 1: you wrote. 472 00:26:48,680 --> 00:26:52,920 Speaker 6: Yes, the story began sometime last year when I started 473 00:26:52,960 --> 00:26:57,760 Speaker 6: seeing this wave of AI generated images of animals on Instagram, 474 00:26:57,960 --> 00:27:01,280 Speaker 6: which is my poison and of social media use. 475 00:27:01,800 --> 00:27:04,680 Speaker 5: And so for context, I'm a big animal lover and. 476 00:27:04,760 --> 00:27:08,760 Speaker 6: I write journalistic articles about nature and wildlife for living, 477 00:27:09,000 --> 00:27:11,520 Speaker 6: So my algorithms know that and they show me a 478 00:27:11,560 --> 00:27:12,560 Speaker 6: lot of animal content. 479 00:27:12,920 --> 00:27:16,160 Speaker 5: But these AI generated images started to. 480 00:27:16,840 --> 00:27:21,439 Speaker 6: Worry me a bit because they were usually very captivating, 481 00:27:21,800 --> 00:27:25,760 Speaker 6: often very beautiful images, but they're not real and I 482 00:27:25,760 --> 00:27:28,840 Speaker 6: guess I could only tell that they were AI generated 483 00:27:28,920 --> 00:27:33,760 Speaker 6: because they were depicting animals doing things that they usually don't, 484 00:27:33,920 --> 00:27:37,600 Speaker 6: or moving not quite the way you'd expect, or sometimes 485 00:27:37,600 --> 00:27:42,160 Speaker 6: they had this like glossy photoshop sheen to them that's 486 00:27:42,280 --> 00:27:44,480 Speaker 6: characteristic of some AI imagery. 487 00:27:45,359 --> 00:27:46,600 Speaker 5: And what alarmed me. 488 00:27:47,160 --> 00:27:49,680 Speaker 6: Was that a lot of people, judging by the comments 489 00:27:49,680 --> 00:27:52,520 Speaker 6: sections on some of these posts, didn't seem to know 490 00:27:52,680 --> 00:27:55,159 Speaker 6: that they were a fake because they look so realistic. 491 00:27:55,600 --> 00:27:57,399 Speaker 1: Wow, what would they say? So? 492 00:27:57,520 --> 00:28:02,800 Speaker 6: One example is an I generated video of a pair 493 00:28:02,880 --> 00:28:06,760 Speaker 6: of hummingbirds sheltering from the rain inside a rose. I 494 00:28:06,800 --> 00:28:11,080 Speaker 6: asked an ornithologist about this. Hummingbirds do not do that, 495 00:28:11,200 --> 00:28:15,640 Speaker 6: and people are still commenting things like oh, how cute lovebirds, 496 00:28:15,960 --> 00:28:20,320 Speaker 6: how adorable, and it's got three point one million likes. 497 00:28:20,560 --> 00:28:23,879 Speaker 1: Wow, So millions of people were believing that these were 498 00:28:23,920 --> 00:28:24,760 Speaker 1: real hummingbirds. 499 00:28:25,080 --> 00:28:28,800 Speaker 6: Yeah, And as a truth abiding journalist, I felt it 500 00:28:28,840 --> 00:28:31,840 Speaker 6: was important to call this out for what it is, 501 00:28:32,080 --> 00:28:36,640 Speaker 6: misinformation and also explore what impact does it have when 502 00:28:36,800 --> 00:28:39,840 Speaker 6: you have millions of people believing this kind of imagery 503 00:28:39,920 --> 00:28:44,320 Speaker 6: is real? How could this affect our perceptions of nature? 504 00:28:44,960 --> 00:28:48,440 Speaker 1: Amazing? What are some other examples of PI generated images 505 00:28:48,480 --> 00:28:49,720 Speaker 1: of animals you've seen out there? 506 00:28:49,920 --> 00:28:54,880 Speaker 6: Yeah, so there are different genres of this kind of imagery. 507 00:28:54,480 --> 00:28:56,640 Speaker 5: So birds in beautiful settings. 508 00:28:56,840 --> 00:29:01,000 Speaker 6: I've seen a lot of videos showing animals like bears 509 00:29:01,080 --> 00:29:05,880 Speaker 6: or sometimes seals being rescued from the ocean. They're also 510 00:29:06,320 --> 00:29:10,440 Speaker 6: AI concoctions of animals that don't exist, like there are 511 00:29:10,480 --> 00:29:15,240 Speaker 6: some like tadpole type creatures with big googly eyes that 512 00:29:15,600 --> 00:29:18,520 Speaker 6: I saw going viral as well. And there are also 513 00:29:18,560 --> 00:29:22,200 Speaker 6: a lot of videos showing like eagles or lions attacking 514 00:29:22,760 --> 00:29:25,040 Speaker 6: children or dogs, that kind of thing. 515 00:29:25,680 --> 00:29:29,200 Speaker 1: Why do you think people make these images and videos? 516 00:29:29,520 --> 00:29:32,400 Speaker 6: Yeah, I think there's a lot of motivations that feed 517 00:29:32,440 --> 00:29:38,840 Speaker 6: into this. The most obvious one is just wanting more likes, clicks, followers, 518 00:29:39,040 --> 00:29:42,400 Speaker 6: which is of course like the currency of status for 519 00:29:42,560 --> 00:29:43,200 Speaker 6: social media. 520 00:29:43,480 --> 00:29:46,280 Speaker 1: People just want to get popular on social media. 521 00:29:46,320 --> 00:29:50,400 Speaker 6: I think so in some cases, they're also financial incentives, 522 00:29:50,520 --> 00:29:55,320 Speaker 6: so social media platforms like TikTok will actually pay creators 523 00:29:55,360 --> 00:29:59,240 Speaker 6: for videos that go viral. I had a conversation with 524 00:29:59,480 --> 00:30:03,320 Speaker 6: one cre who makes AI generated videos, and he had 525 00:30:03,320 --> 00:30:07,000 Speaker 6: a motivation that really surprised me. For him, part of 526 00:30:07,040 --> 00:30:10,560 Speaker 6: the fun is deliberately testing whether people can recognize these 527 00:30:10,600 --> 00:30:14,280 Speaker 6: as fake. So he'll often make his videos to look 528 00:30:14,440 --> 00:30:17,720 Speaker 6: like they were shot from an iPhone, like there is 529 00:30:17,840 --> 00:30:21,320 Speaker 6: one showing a giraffe walking through a mall in Dubai. 530 00:30:22,120 --> 00:30:24,600 Speaker 1: Now you wrote a little bit about why we're vulnerable 531 00:30:24,640 --> 00:30:27,200 Speaker 1: to these images, like why are we prone to believe 532 00:30:27,240 --> 00:30:28,320 Speaker 1: these fake animals? 533 00:30:28,560 --> 00:30:31,000 Speaker 6: Yeah, I think a lot of the spoils down to 534 00:30:31,440 --> 00:30:36,920 Speaker 6: our growing lack of knowledge about animals among many societies today. 535 00:30:37,040 --> 00:30:39,840 Speaker 6: So a lot of us are living in cities, We're 536 00:30:39,880 --> 00:30:43,520 Speaker 6: being less exposed in nature. I think we're losing touch 537 00:30:43,760 --> 00:30:47,120 Speaker 6: on a large scale, and I mean a lot of us, 538 00:30:47,240 --> 00:30:51,080 Speaker 6: including me, I should say, struggle to identify species of 539 00:30:51,120 --> 00:30:55,000 Speaker 6: plants and animals around us. There is one study of 540 00:30:55,160 --> 00:30:59,200 Speaker 6: students at Oxford University that showed that half of them 541 00:30:59,480 --> 00:31:05,200 Speaker 6: couldn't five British bird species. If I recall correctly, those 542 00:31:05,240 --> 00:31:06,840 Speaker 6: were actually biology students. 543 00:31:06,960 --> 00:31:07,320 Speaker 1: Wow. 544 00:31:07,480 --> 00:31:08,600 Speaker 5: And I think we're. 545 00:31:08,560 --> 00:31:14,960 Speaker 6: Especially susceptible to biology related misinformation in particular because, I mean, 546 00:31:15,040 --> 00:31:19,360 Speaker 6: scientists are always discovering new species and new behaviors. Just 547 00:31:19,400 --> 00:31:23,360 Speaker 6: the other week, there's that story about Veronica, this Austrian 548 00:31:23,640 --> 00:31:26,880 Speaker 6: cow who used a broom to scratch herself that was 549 00:31:26,920 --> 00:31:32,480 Speaker 6: not AI generated because that information came from reputable news outlets. 550 00:31:32,560 --> 00:31:34,280 Speaker 1: For a second, I thought you were mentioning something that 551 00:31:34,400 --> 00:31:35,120 Speaker 1: was not real. 552 00:31:35,480 --> 00:31:37,120 Speaker 5: No, No, I'm pretty sure that was real. 553 00:31:37,320 --> 00:31:39,600 Speaker 1: This was a cow that grabbed a broom and scratched 554 00:31:39,640 --> 00:31:40,280 Speaker 1: itself with it. 555 00:31:40,720 --> 00:31:44,080 Speaker 6: Yeah, she had figured out to use a broom to 556 00:31:44,120 --> 00:31:48,840 Speaker 6: scratch herself in particular ways, and I think it was 557 00:31:49,000 --> 00:31:52,840 Speaker 6: described as the first case of tool use in cows 558 00:31:53,000 --> 00:31:53,880 Speaker 6: by scientists. 559 00:31:54,840 --> 00:31:58,160 Speaker 1: Bonus points. Do you if you correctly guessed the cows 560 00:31:58,160 --> 00:32:01,320 Speaker 1: scratching itself with a broom is the real image and 561 00:32:01,640 --> 00:32:04,760 Speaker 1: not AI generated. It's true, look it up on a 562 00:32:04,880 --> 00:32:07,720 Speaker 1: real new site. Okay, now let's get to how these 563 00:32:07,760 --> 00:32:12,400 Speaker 1: AI made images can affect us. And then in the 564 00:32:12,480 --> 00:32:15,040 Speaker 1: article you wrote about how this can affect how we 565 00:32:15,160 --> 00:32:16,960 Speaker 1: relate to animals, what did you learn. 566 00:32:17,080 --> 00:32:20,560 Speaker 6: Yeah, so a lot of these impacts are just unfolding 567 00:32:20,800 --> 00:32:23,880 Speaker 6: at the moment, so we don't have concrete data on this. 568 00:32:24,080 --> 00:32:27,440 Speaker 6: But the experts I spoke to had a few concerns. 569 00:32:27,680 --> 00:32:31,640 Speaker 6: One of them is that as we see more images 570 00:32:32,000 --> 00:32:34,840 Speaker 6: of a certain animal, that can make us more likely 571 00:32:34,920 --> 00:32:39,040 Speaker 6: to believe that that animal is in fact more abundant 572 00:32:39,160 --> 00:32:42,640 Speaker 6: in the real world. So there's some interesting research showing 573 00:32:42,720 --> 00:32:48,480 Speaker 6: that seeing images of endangered lions or giraffes can cause 574 00:32:48,560 --> 00:32:52,360 Speaker 6: us to overestimate their abundance in the wild. And I 575 00:32:52,360 --> 00:32:55,640 Speaker 6: don't know the exact numbers, but I think there's only 576 00:32:55,960 --> 00:32:58,840 Speaker 6: like a little over one hundred and ten thousand wild 577 00:32:58,880 --> 00:33:01,760 Speaker 6: giraffes in the world, and only a little more than 578 00:33:01,800 --> 00:33:04,920 Speaker 6: twenty thousand lions. But a lot of people don't know 579 00:33:04,960 --> 00:33:08,239 Speaker 6: that because we see these animals all the time in 580 00:33:08,320 --> 00:33:13,959 Speaker 6: travel commercials or children's toys cartoons, and so some of 581 00:33:14,000 --> 00:33:17,920 Speaker 6: the AI generated videos of birds that I've seen featured 582 00:33:17,960 --> 00:33:22,440 Speaker 6: birds of paradise, many of which are threatened by habitat loss. 583 00:33:22,680 --> 00:33:25,600 Speaker 6: So I think seeing these images a lot can cause 584 00:33:25,680 --> 00:33:28,880 Speaker 6: us to believe they're actually quite abundant and make us 585 00:33:28,960 --> 00:33:32,560 Speaker 6: blind to the conservation crisis these creatures are in. 586 00:33:32,880 --> 00:33:33,280 Speaker 2: I see. 587 00:33:35,160 --> 00:33:37,880 Speaker 1: So that's one way, hey, I slap can change our 588 00:33:37,960 --> 00:33:41,240 Speaker 1: view of animals. It can distort how abundant or how 589 00:33:41,280 --> 00:33:44,960 Speaker 1: safe from extinction they are. It can also distort our 590 00:33:45,040 --> 00:33:48,640 Speaker 1: view of how safe or dangerous so animals are. 591 00:33:50,480 --> 00:33:53,160 Speaker 6: And there are some other concerns and more related to 592 00:33:53,600 --> 00:33:59,520 Speaker 6: specific kinds of videos. So one expert mentioned that seeing 593 00:34:00,120 --> 00:34:05,120 Speaker 6: human like depictions of animals like polar bears could lead 594 00:34:05,200 --> 00:34:09,360 Speaker 6: us to believe that they're cute, fluffy, and approachable creatures, 595 00:34:10,080 --> 00:34:14,080 Speaker 6: and that kind of information could lead people to seek 596 00:34:14,120 --> 00:34:18,920 Speaker 6: out encounters with those animals in the wild. There actually 597 00:34:19,000 --> 00:34:22,640 Speaker 6: are cases of people in the US going out to 598 00:34:22,680 --> 00:34:27,280 Speaker 6: try and pet mountain lions and bears and unfortunately ending 599 00:34:27,400 --> 00:34:28,600 Speaker 6: they end up getting moulled. 600 00:34:29,680 --> 00:34:31,560 Speaker 1: Are you trying to tell me that polar bears don't 601 00:34:31,560 --> 00:34:33,640 Speaker 1: really drink Coca cola and wear scars? 602 00:34:33,880 --> 00:34:36,799 Speaker 5: I'm afraid not. Sorry to bust that myth. 603 00:34:38,480 --> 00:34:40,360 Speaker 1: Say she didn't go up to a polar bear with 604 00:34:40,440 --> 00:34:41,520 Speaker 1: us pop drink. 605 00:34:42,200 --> 00:34:44,919 Speaker 5: Definitely not. You should never approach a polar bear. 606 00:34:46,520 --> 00:34:50,239 Speaker 1: Yes, fake images can make some animals seem safer than 607 00:34:50,320 --> 00:34:53,640 Speaker 1: they really are, and you can also have the opposite effect. 608 00:34:54,040 --> 00:34:57,560 Speaker 1: You can make some animals seem more dangerous than they 609 00:34:57,640 --> 00:34:58,120 Speaker 1: really are. 610 00:34:59,400 --> 00:35:03,040 Speaker 6: There are a lot of videos showing animals like eagles 611 00:35:03,239 --> 00:35:09,160 Speaker 6: or lions attacking humans or dots, which could make people 612 00:35:09,200 --> 00:35:12,360 Speaker 6: more likely to support the culling of those animals or 613 00:35:12,640 --> 00:35:15,319 Speaker 6: less likely to support their conservation. 614 00:35:17,000 --> 00:35:21,239 Speaker 1: Kindarina says. Another big genre of AI generated images or 615 00:35:21,320 --> 00:35:25,600 Speaker 1: videos are those of exotic or endangered animals walking around 616 00:35:25,920 --> 00:35:27,320 Speaker 1: in people's houses. 617 00:35:28,200 --> 00:35:32,719 Speaker 6: Seeing a lot of AI generated imagery also features exotic 618 00:35:32,920 --> 00:35:34,000 Speaker 6: animals being. 619 00:35:33,880 --> 00:35:36,360 Speaker 5: Kept as pets, such as Kappi virus. 620 00:35:36,640 --> 00:35:41,200 Speaker 6: Seeing that kind of imagery could make people more likely 621 00:35:41,320 --> 00:35:44,439 Speaker 6: to want to have them as pets and potentially even 622 00:35:44,560 --> 00:35:47,680 Speaker 6: fuel the capture of these animals in the wild. 623 00:35:48,880 --> 00:35:52,320 Speaker 1: But one of the worst consequences of all this AI slob, 624 00:35:52,680 --> 00:35:57,040 Speaker 1: Katerina argues, is that it makes us doubt what's actually real. 625 00:35:58,560 --> 00:36:02,319 Speaker 6: I want to add the big impact that worries me 626 00:36:02,520 --> 00:36:06,960 Speaker 6: personally is that as people begin to cotton on that 627 00:36:07,200 --> 00:36:10,920 Speaker 6: a lot of this imagery isn't real, they might start 628 00:36:10,920 --> 00:36:16,400 Speaker 6: to dismiss wildlife imagery altogether. So you could have photographers 629 00:36:16,480 --> 00:36:19,640 Speaker 6: spending years trying to get a significant image of a 630 00:36:19,640 --> 00:36:24,320 Speaker 6: bird doing something interesting, only to have people on social 631 00:36:24,400 --> 00:36:27,480 Speaker 6: media say, oh, that's fake news. I'm not going to 632 00:36:27,520 --> 00:36:31,279 Speaker 6: care about that. And the same thing could happen with 633 00:36:31,640 --> 00:36:36,200 Speaker 6: scientists who discover new species or new behaviors. And I 634 00:36:36,200 --> 00:36:39,040 Speaker 6: think that's really what worries me the most is that 635 00:36:39,120 --> 00:36:42,000 Speaker 6: it could further alienate us from the natural. 636 00:36:41,680 --> 00:36:44,680 Speaker 1: World because we don't know what to believe anymore, and 637 00:36:44,719 --> 00:36:48,920 Speaker 1: so therefore we will tend to not care about. 638 00:36:48,760 --> 00:36:50,080 Speaker 5: It, right exactly. 639 00:36:50,440 --> 00:36:53,239 Speaker 6: Yeah, And I think that's really sad because this is 640 00:36:53,280 --> 00:36:57,360 Speaker 6: a time when species really need our attention. So twenty 641 00:36:57,600 --> 00:37:01,279 Speaker 6: eight percent of plant and animal species that have been 642 00:37:01,320 --> 00:37:05,640 Speaker 6: studied are in some way threatened with extinction, so basically 643 00:37:05,640 --> 00:37:09,080 Speaker 6: a third. And to me, I think that's the saddest 644 00:37:09,080 --> 00:37:11,279 Speaker 6: thing about this is that not only are we not 645 00:37:11,400 --> 00:37:15,280 Speaker 6: paying attention to these real threats that animals are facing, 646 00:37:15,800 --> 00:37:19,680 Speaker 6: but that we're instead looking at these fake images of 647 00:37:19,719 --> 00:37:24,279 Speaker 6: them on our phones that don't reflect reality. It's like 648 00:37:25,239 --> 00:37:29,680 Speaker 6: the world's on fire and we're looking at these hummingbirds 649 00:37:29,719 --> 00:37:33,440 Speaker 6: and roses and going, oh, how cute. That's wonderful, and 650 00:37:33,480 --> 00:37:36,520 Speaker 6: there are actually some hummingbird species that are in need 651 00:37:36,640 --> 00:37:38,319 Speaker 6: of conservation attention. 652 00:37:38,719 --> 00:37:41,880 Speaker 1: Meaning that it's sort of detaching, is further from reality, 653 00:37:42,280 --> 00:37:46,440 Speaker 1: and the reality could be kind of bad out there. Yeah, exactly, 654 00:37:46,880 --> 00:37:48,799 Speaker 1: like the world's on fire and people are looking at 655 00:37:48,800 --> 00:37:53,040 Speaker 1: the pictures of the fire and going, oh, that's ais. 656 00:37:53,239 --> 00:37:57,680 Speaker 6: Yeah, exactly, all right. 657 00:37:58,239 --> 00:38:01,120 Speaker 1: Well, hopefully that gives you a pretty sense of why 658 00:38:01,120 --> 00:38:04,680 Speaker 1: these AI generated images are so powerful all of a sudden, 659 00:38:04,960 --> 00:38:07,880 Speaker 1: why our brain is so easily fooled by them, and 660 00:38:07,920 --> 00:38:11,520 Speaker 1: how even pictures and videos of cute animals can impact 661 00:38:11,680 --> 00:38:15,200 Speaker 1: our sense of reality. Some of you might be wondering, well, 662 00:38:15,440 --> 00:38:17,839 Speaker 1: what can we do about this, and the answer is 663 00:38:18,120 --> 00:38:21,600 Speaker 1: not that complicated. According to our experts. You should a 664 00:38:22,440 --> 00:38:25,040 Speaker 1: not believe everything you see on the Internet or your 665 00:38:25,080 --> 00:38:29,759 Speaker 1: social media feed and look for reliable and reputable sources. B. 666 00:38:30,560 --> 00:38:33,239 Speaker 1: You should keep pushing companies to do a better job 667 00:38:33,280 --> 00:38:38,240 Speaker 1: of labeling AI generated content, and c lobby your elected 668 00:38:38,280 --> 00:38:42,240 Speaker 1: officials to put in place more rules and regulations around 669 00:38:42,239 --> 00:38:45,440 Speaker 1: this stuff. But most definitely you should listen to this 670 00:38:45,560 --> 00:38:50,520 Speaker 1: podcast every week. Wait did I say that or did 671 00:38:50,800 --> 00:38:54,600 Speaker 1: AI say that? Either way, if you're hearing this and 672 00:38:54,719 --> 00:38:58,239 Speaker 1: you're human, thanks for joining us. See you next time 673 00:39:02,760 --> 00:39:06,200 Speaker 1: you've been listening to Science Stuff. The production of iHeartRadio 674 00:39:06,960 --> 00:39:09,920 Speaker 1: written and produced by me or hych Ham, credited by 675 00:39:10,000 --> 00:39:13,880 Speaker 1: Rose Seguda, executive producer Jerry Rowland, and audio engineer and 676 00:39:13,920 --> 00:39:17,320 Speaker 1: mixer Kasey Pegram and you can follow me on social media. 677 00:39:17,560 --> 00:39:20,200 Speaker 1: Just search for PhD Comics and the name of your 678 00:39:20,200 --> 00:39:23,120 Speaker 1: favorite platform. Be sure to subscribe to Science Stuff on 679 00:39:23,160 --> 00:39:27,000 Speaker 1: the iHeartRadio app, Apple Podcasts, or wherever you get your podcasts, 680 00:39:27,040 --> 00:39:30,000 Speaker 1: and please tell your friends we'll be back next Wednesday 681 00:39:30,040 --> 00:39:32,560 Speaker 1: with another episode.