1 00:00:04,559 --> 00:00:08,959 Speaker 1: What role will AI play in the future of fake 2 00:00:09,080 --> 00:00:12,200 Speaker 1: news and misinformation? And what does this have to do 3 00:00:12,320 --> 00:00:17,639 Speaker 1: with your brain's internal models, or with voice passwords, or 4 00:00:17,680 --> 00:00:22,560 Speaker 1: with what I'm calling the tall intelligence problem? And why 5 00:00:22,600 --> 00:00:25,960 Speaker 1: do I believe that these earliest days of AI are 6 00:00:26,120 --> 00:00:31,600 Speaker 1: actually it's golden age and we're quickly heading for a Balkanization. 7 00:00:35,200 --> 00:00:38,320 Speaker 1: Welcome to Inner Cosmos with me David Eagleman. I'm a 8 00:00:38,400 --> 00:00:43,160 Speaker 1: neuroscientist and author at Stanford and in these episodes we 9 00:00:43,240 --> 00:00:47,720 Speaker 1: sail deeply into our three pound universe to understand why 10 00:00:47,800 --> 00:00:58,720 Speaker 1: and how our lives look the way they do now. 11 00:00:58,760 --> 00:01:01,200 Speaker 1: In the last two episodes, we talked about the notion 12 00:01:01,320 --> 00:01:05,959 Speaker 1: of truth versus misinformation. Two weeks ago we covered the 13 00:01:06,040 --> 00:01:10,000 Speaker 1: question about truth in the media, and last week was 14 00:01:10,040 --> 00:01:14,800 Speaker 1: specifically about truth on the Internet. Today's episode is about 15 00:01:14,920 --> 00:01:21,680 Speaker 1: truth versus misinformation and artificial intelligence. So if you happened 16 00:01:21,760 --> 00:01:23,759 Speaker 1: to listen to the last two episodes, you'll know I've 17 00:01:23,760 --> 00:01:28,320 Speaker 1: been arguing that the position of truth is not as 18 00:01:28,400 --> 00:01:30,880 Speaker 1: simple as many pundits have made it out to be. 19 00:01:31,440 --> 00:01:35,120 Speaker 1: A lot of people have said truth is declining and 20 00:01:35,200 --> 00:01:40,839 Speaker 1: misinformation is on the rise. It's achieving ascendency so much 21 00:01:40,880 --> 00:01:44,560 Speaker 1: so that the Oxford English Dictionary in twenty sixteen coined 22 00:01:44,600 --> 00:01:48,800 Speaker 1: the term post truth, implying from the structure of the 23 00:01:48,840 --> 00:01:52,120 Speaker 1: word that there used to be a time where people 24 00:01:52,240 --> 00:01:58,200 Speaker 1: operated on truth, whereas now, regretfully, people's beliefs are predicated 25 00:01:58,280 --> 00:02:02,200 Speaker 1: on emotion and person old beliefs. So I've made my 26 00:02:02,360 --> 00:02:06,200 Speaker 1: arguments in those episodes about why I think that position 27 00:02:06,360 --> 00:02:09,480 Speaker 1: is so specious, and what I gave was a simple 28 00:02:09,680 --> 00:02:14,120 Speaker 1: historical analysis that demonstrates beyond any doubt that people have 29 00:02:14,320 --> 00:02:18,960 Speaker 1: never operated on anything but emotion and personal beliefs, and 30 00:02:19,040 --> 00:02:24,160 Speaker 1: the idea that that has recently changed belies nothing but 31 00:02:24,280 --> 00:02:29,360 Speaker 1: an appalling ignorance of history. And specifically in the last episode, 32 00:02:29,400 --> 00:02:33,200 Speaker 1: I argued that we're actually in a much better position 33 00:02:33,320 --> 00:02:36,880 Speaker 1: now because of the Internet, as it prevents the one 34 00:02:37,000 --> 00:02:43,760 Speaker 1: thing that has historically proven itself worse than misinformation, which 35 00:02:43,840 --> 00:02:49,400 Speaker 1: is censorship, having someone else decide for you what you 36 00:02:49,520 --> 00:02:53,320 Speaker 1: are allowed to see. And if you want to know more, 37 00:02:53,440 --> 00:02:56,200 Speaker 1: please go back and listen to those episodes where I 38 00:02:56,240 --> 00:03:01,600 Speaker 1: give examples of the USSR controlling a total clamp down 39 00:03:02,040 --> 00:03:06,119 Speaker 1: on the press and even on copying machines, and China 40 00:03:06,160 --> 00:03:09,400 Speaker 1: controlling which websites you are allowed to go to, and 41 00:03:09,639 --> 00:03:14,000 Speaker 1: Nikolai Chichescu of Romania controlling even the weather reports, and 42 00:03:14,280 --> 00:03:18,919 Speaker 1: Saddam Hussein of Iraq outlawing maps of Baghdad, and on 43 00:03:19,200 --> 00:03:24,200 Speaker 1: and on. Just imagine if President Trump or President Biden 44 00:03:24,600 --> 00:03:29,120 Speaker 1: had one hundred percent control over what you are or 45 00:03:29,160 --> 00:03:33,840 Speaker 1: are not allowed to see. This was the situation and 46 00:03:33,960 --> 00:03:37,440 Speaker 1: so many recent historical examples that we examined in the 47 00:03:37,520 --> 00:03:41,360 Speaker 1: last two episodes, from left wing communism in China and 48 00:03:41,400 --> 00:03:45,120 Speaker 1: the USSR, to right wing Nazism or fascism in Germany 49 00:03:45,200 --> 00:03:49,440 Speaker 1: or Italy, to theocratic dictatorships like we find in Iran. 50 00:03:50,080 --> 00:03:54,760 Speaker 1: In all cases the government decides what is the proper 51 00:03:54,800 --> 00:03:58,520 Speaker 1: material for you to see or Nazi and in all 52 00:03:58,640 --> 00:04:03,280 Speaker 1: cases uniformly, this has been disastrous and has led to 53 00:04:03,320 --> 00:04:07,600 Speaker 1: the starvation or execution of tens of millions of people. 54 00:04:08,040 --> 00:04:11,960 Speaker 1: So even if you don't like hearing other people's opinions, 55 00:04:12,280 --> 00:04:15,480 Speaker 1: and you deep down believe that they're all misinformed and 56 00:04:15,520 --> 00:04:19,560 Speaker 1: you know much better, the fact is that our society 57 00:04:19,640 --> 00:04:23,680 Speaker 1: will survive for longer if we put up with those people. 58 00:04:24,320 --> 00:04:27,799 Speaker 1: Rather than imagining we'd be better off if we could 59 00:04:27,960 --> 00:04:32,599 Speaker 1: just legislate or take up arms to have them agree 60 00:04:32,600 --> 00:04:35,240 Speaker 1: with us. So that's where we are so far in 61 00:04:35,279 --> 00:04:38,359 Speaker 1: this series. But today I'm going to dive into the 62 00:04:38,520 --> 00:04:43,039 Speaker 1: third part because we've just arrived in this technological era 63 00:04:43,720 --> 00:04:49,360 Speaker 1: which got here with unexpected speed and success of artificial intelligence, 64 00:04:49,400 --> 00:04:53,560 Speaker 1: and specifically of generative AI, which is a type of 65 00:04:53,680 --> 00:04:59,240 Speaker 1: artificial neural network that creates new content like beautiful passages 66 00:04:59,240 --> 00:05:05,000 Speaker 1: of writing, or shockingly great images or expert level music. 67 00:05:05,760 --> 00:05:11,000 Speaker 1: Its success blossomed suddenly mostly because of the increasing availability 68 00:05:11,040 --> 00:05:15,240 Speaker 1: of large data sets combined with more powerful computing hardware, 69 00:05:15,440 --> 00:05:20,599 Speaker 1: and as a result, generative AI now produces results that 70 00:05:20,640 --> 00:05:26,240 Speaker 1: are indistinguishable from human generated content like news articles, or 71 00:05:26,279 --> 00:05:31,000 Speaker 1: photographs or voices. And the question today is what will 72 00:05:31,040 --> 00:05:37,000 Speaker 1: AI mean for the future of disinformation or truth? Has 73 00:05:37,240 --> 00:05:43,480 Speaker 1: AI thrown a wrench irreversibly into this game of truth telling? 74 00:05:44,800 --> 00:05:47,680 Speaker 1: So let's start with the question of truth and AI. 75 00:05:48,040 --> 00:05:50,880 Speaker 1: If you ask the question to people on the street, 76 00:05:51,680 --> 00:05:54,760 Speaker 1: was Trump a good president? Obviously you're going to get 77 00:05:54,760 --> 00:05:58,279 Speaker 1: a range of responses. So the question of what is 78 00:05:58,480 --> 00:06:01,960 Speaker 1: truth here is a difficult one. Now, if you ask 79 00:06:02,160 --> 00:06:06,359 Speaker 1: chat GPT this question, it will tell you the question 80 00:06:06,600 --> 00:06:12,159 Speaker 1: is subjective and depends on individual perspectives and priorities. And 81 00:06:12,200 --> 00:06:16,120 Speaker 1: then it'll tell you supporters of Trump often highlight X, 82 00:06:16,240 --> 00:06:19,760 Speaker 1: y Z, and then it says critics point to concerns 83 00:06:19,800 --> 00:06:23,560 Speaker 1: such as ABC, and it does this sort of back 84 00:06:23,640 --> 00:06:26,479 Speaker 1: and forth for most questions that you ask it, even 85 00:06:26,520 --> 00:06:29,359 Speaker 1: if you had intended to get a yes or no answer. 86 00:06:29,920 --> 00:06:32,400 Speaker 1: And a lot of people that I've talked with find 87 00:06:32,480 --> 00:06:37,479 Speaker 1: this aspect of llms, these large language models, really annoying. 88 00:06:37,920 --> 00:06:41,000 Speaker 1: And that's because every question you ask to Barret or 89 00:06:41,080 --> 00:06:45,040 Speaker 1: chat GPT gives this sort of wishy washy answer and 90 00:06:45,360 --> 00:06:48,359 Speaker 1: tells you, look, there's this opinion on it, there's that 91 00:06:48,480 --> 00:06:51,560 Speaker 1: opinion on it. More discussion will have to take place, 92 00:06:52,080 --> 00:06:55,400 Speaker 1: and if you say yes or no, was Trump a 93 00:06:55,440 --> 00:06:59,279 Speaker 1: good president? It will tell you quote as a neutral 94 00:06:59,320 --> 00:07:03,240 Speaker 1: and objective AI, I don't have personal opinions, The evaluation 95 00:07:03,320 --> 00:07:06,680 Speaker 1: of whether Donald Trump was a good president is subjective 96 00:07:06,960 --> 00:07:11,040 Speaker 1: and depends on individual perspectives and priorities. And then it 97 00:07:11,160 --> 00:07:14,280 Speaker 1: ends by noting that it is essential to consider a 98 00:07:14,440 --> 00:07:18,280 Speaker 1: range of perspectives and weigh various aspects of his presidency 99 00:07:18,680 --> 00:07:22,040 Speaker 1: to form an informed opinion. And you know what it is, 100 00:07:22,080 --> 00:07:24,119 Speaker 1: annoying that it does that, and you know what, else, 101 00:07:24,760 --> 00:07:28,440 Speaker 1: it's right. This is exactly what we should do, given 102 00:07:28,480 --> 00:07:32,880 Speaker 1: that public opinion on Trump's presidency is polarized and assessments 103 00:07:33,280 --> 00:07:38,200 Speaker 1: vary widely based on political ideology and personal values. So 104 00:07:38,280 --> 00:07:41,720 Speaker 1: what Bard or chat GPT or other lms do is 105 00:07:41,840 --> 00:07:45,200 Speaker 1: actually quite genius, which is that they don't get caught 106 00:07:45,280 --> 00:07:48,720 Speaker 1: in the trap of giving a yes or no answer 107 00:07:49,160 --> 00:07:52,640 Speaker 1: to what we think is a yes or no question. Instead, 108 00:07:53,000 --> 00:07:56,880 Speaker 1: they are synthesizing the opinions of millions and millions of 109 00:07:56,920 --> 00:08:01,040 Speaker 1: people who have written things down, and therefore or it says, well, 110 00:08:01,160 --> 00:08:03,640 Speaker 1: some people think this, and some people think that, and 111 00:08:03,680 --> 00:08:07,600 Speaker 1: although some people find that annoying, it actually represents a 112 00:08:07,840 --> 00:08:12,680 Speaker 1: beautiful sort of fairness. For example, you can ask chat 113 00:08:12,720 --> 00:08:17,480 Speaker 1: GPT give me different points of view about abortion, and 114 00:08:17,520 --> 00:08:19,680 Speaker 1: it does a beautiful job with that. It gives you 115 00:08:19,760 --> 00:08:24,680 Speaker 1: the perspective from pro choice advocates about reproductive rights and 116 00:08:24,720 --> 00:08:28,040 Speaker 1: women having the right to make decisions about their own bodies, 117 00:08:28,280 --> 00:08:31,640 Speaker 1: and also health and safety concerns, making sure that women 118 00:08:31,720 --> 00:08:36,160 Speaker 1: have access to approved medicine to reduce risk of illegal methods, 119 00:08:36,559 --> 00:08:39,360 Speaker 1: and issues of autonomy. And then it also gives the 120 00:08:39,400 --> 00:08:42,680 Speaker 1: point of view of pro life advocates, their belief in 121 00:08:42,720 --> 00:08:46,520 Speaker 1: the inherent right to life for the unborn fetus and 122 00:08:46,720 --> 00:08:51,280 Speaker 1: alternatives instead of abortion, like adoption or parenting. And finally, 123 00:08:51,280 --> 00:08:54,720 Speaker 1: it points to their ethical beliefs about the sanctity of 124 00:08:54,840 --> 00:08:58,400 Speaker 1: human life from conception, and on top of that, it 125 00:08:58,480 --> 00:09:03,640 Speaker 1: spells out some other perspects as well as compromised positions 126 00:09:03,760 --> 00:09:08,560 Speaker 1: like trimester based regulation and exception clauses when a mother's 127 00:09:08,600 --> 00:09:12,880 Speaker 1: health is at risk. Now, why does chat GPT do 128 00:09:13,080 --> 00:09:15,760 Speaker 1: such a good job at this sort of thing? Because 129 00:09:16,480 --> 00:09:20,480 Speaker 1: it's acting like a meta human. It's like a parent 130 00:09:20,640 --> 00:09:25,760 Speaker 1: watching squabbling children and understanding the perspective where each different 131 00:09:25,840 --> 00:09:29,280 Speaker 1: child is coming from. Similarly, I asked it to tell 132 00:09:29,320 --> 00:09:33,360 Speaker 1: me the different points of view about the Israeli Palestinian conflict, 133 00:09:33,640 --> 00:09:38,720 Speaker 1: and it succinctly and fairly captured the perspective of both sides. Again, 134 00:09:39,080 --> 00:09:41,920 Speaker 1: this is because it has access to all the writing 135 00:09:41,960 --> 00:09:45,600 Speaker 1: of the world, so it sees the different perspectives. It's 136 00:09:45,640 --> 00:09:49,160 Speaker 1: not making a choice, that's not its job, and maybe 137 00:09:49,200 --> 00:09:52,320 Speaker 1: it couldn't do it technologically anyway. So instead, what it 138 00:09:52,400 --> 00:09:56,880 Speaker 1: represents is a fair and balanced view of the world 139 00:09:56,920 --> 00:10:01,600 Speaker 1: that we're in now. What's fascinating is that we are 140 00:10:01,720 --> 00:10:06,080 Speaker 1: probably at a moment with AI that is perhaps the 141 00:10:06,240 --> 00:10:11,160 Speaker 1: fairest and most balanced that we will ever see. Why. 142 00:10:11,679 --> 00:10:15,520 Speaker 1: It's because I predict that people will start training up 143 00:10:15,600 --> 00:10:19,240 Speaker 1: their own AIS, and they will do so with what 144 00:10:19,520 --> 00:10:25,160 Speaker 1: they consider the best information. So one organization might say, 145 00:10:25,600 --> 00:10:27,320 Speaker 1: we're not going to use all the books, We're just 146 00:10:27,360 --> 00:10:31,000 Speaker 1: going to use the classics, and anything that is gay 147 00:10:31,200 --> 00:10:34,440 Speaker 1: or trans or non binary, we're not including that because 148 00:10:34,440 --> 00:10:37,400 Speaker 1: we don't feel like it's necessary to get this next 149 00:10:37,440 --> 00:10:40,360 Speaker 1: generation of intelligence system trained up. We're going to train 150 00:10:40,400 --> 00:10:43,120 Speaker 1: this with what we believe in. And on the other side, 151 00:10:43,120 --> 00:10:45,360 Speaker 1: you'll have people on the left wing who say, I 152 00:10:45,360 --> 00:10:48,320 Speaker 1: don't want to include Mark Twain in here, or Raould 153 00:10:48,360 --> 00:10:51,640 Speaker 1: Dahl or certain books by doctor seuss Er, various Shakespeare 154 00:10:51,720 --> 00:10:55,200 Speaker 1: plays or things like that, because they represent a point 155 00:10:55,240 --> 00:10:58,040 Speaker 1: of view that people used to have, but we don't 156 00:10:58,080 --> 00:11:01,000 Speaker 1: believe in that anymore, so we want to leave that out. 157 00:11:01,679 --> 00:11:04,200 Speaker 1: Where you can imagine another group that feels it's better 158 00:11:04,480 --> 00:11:08,160 Speaker 1: if there's no violence in the training data. So we're 159 00:11:08,160 --> 00:11:11,120 Speaker 1: only going to use uplifting books where people help each other, 160 00:11:11,160 --> 00:11:13,920 Speaker 1: and there's no such thing as violence, and that will 161 00:11:13,960 --> 00:11:17,040 Speaker 1: galvanize a different group to say, look, you're being naive. 162 00:11:17,160 --> 00:11:20,480 Speaker 1: The world is full of psychopaths who will commit violence 163 00:11:20,480 --> 00:11:22,440 Speaker 1: against you, and you need to be prepared for that, 164 00:11:22,840 --> 00:11:26,600 Speaker 1: so you can recognize when a violence psychopath is seeking 165 00:11:26,679 --> 00:11:29,880 Speaker 1: power over you or seeking political power in your neighborhood 166 00:11:29,960 --> 00:11:32,280 Speaker 1: or on a national stage. So we're going to leave 167 00:11:32,320 --> 00:11:36,000 Speaker 1: out the uplifting rom com books and we're going to 168 00:11:36,679 --> 00:11:39,320 Speaker 1: just include the books about the real world and what 169 00:11:39,520 --> 00:11:42,280 Speaker 1: actually happens in warfare and how one needs to be 170 00:11:42,320 --> 00:11:46,199 Speaker 1: prepared for it, and on and on and on. So 171 00:11:46,280 --> 00:11:51,360 Speaker 1: I predict we are headed rapidly for a balkanization of 172 00:11:51,520 --> 00:11:55,360 Speaker 1: large language models. Just in case you don't know, Balkanization 173 00:11:55,559 --> 00:11:58,840 Speaker 1: is a word that refers to the process of breaking 174 00:11:58,840 --> 00:12:03,079 Speaker 1: something up into small, isolated factions. This originated from the 175 00:12:03,160 --> 00:12:06,760 Speaker 1: Balkan Peninsula, where a bunch of ethnic and political groups 176 00:12:07,200 --> 00:12:13,120 Speaker 1: sought independence and created smaller, less cooperative states. So I'm 177 00:12:13,160 --> 00:12:19,280 Speaker 1: suggesting we're heading towards a disintegration into smaller, less cooperative 178 00:12:19,720 --> 00:12:23,640 Speaker 1: large language models. The idea is you can go ask 179 00:12:23,679 --> 00:12:27,000 Speaker 1: a question to your left wing model or your right 180 00:12:27,040 --> 00:12:31,880 Speaker 1: wing model, your model that lives in Wokistan or Magastan. 181 00:12:32,600 --> 00:12:35,600 Speaker 1: And I think that's a shame because it's taking something 182 00:12:35,640 --> 00:12:40,199 Speaker 1: that should be smarter and better than humans, but we're 183 00:12:40,240 --> 00:12:43,240 Speaker 1: going to manipulate it to be the kind of human 184 00:12:43,320 --> 00:12:45,840 Speaker 1: that you are, so that you can say, yes, I 185 00:12:46,000 --> 00:12:49,800 Speaker 1: like its answers. That makes sense now, the model tells 186 00:12:50,400 --> 00:12:55,480 Speaker 1: the truth. It's like the Biblical creation story where God 187 00:12:55,559 --> 00:12:59,400 Speaker 1: creates man in his own image, which actually seems to 188 00:12:59,440 --> 00:13:02,320 Speaker 1: me like a real shame, because what if God could 189 00:13:02,360 --> 00:13:06,640 Speaker 1: have created something that was better than himself. That's what 190 00:13:06,679 --> 00:13:09,640 Speaker 1: we have the chance to do with AI. But I 191 00:13:09,679 --> 00:13:12,520 Speaker 1: think it's just around the corner that we human say 192 00:13:13,240 --> 00:13:15,840 Speaker 1: I know what truth is, and I'm going to make 193 00:13:15,880 --> 00:13:19,760 Speaker 1: sure that I fashion this powerful new creature so that 194 00:13:19,880 --> 00:13:22,800 Speaker 1: it looks just like me. I'm going to take the 195 00:13:23,320 --> 00:13:27,920 Speaker 1: vast space of possibility and squeeze it down until it 196 00:13:28,000 --> 00:13:32,880 Speaker 1: is nothing but a successful clone of me. So my 197 00:13:33,040 --> 00:13:37,240 Speaker 1: hope is that what will remain or grow up into 198 00:13:37,240 --> 00:13:41,520 Speaker 1: this space is something like a Wikipedia of large language models, 199 00:13:41,559 --> 00:13:45,320 Speaker 1: something that takes the space of human opinion and says, look, 200 00:13:45,360 --> 00:13:47,720 Speaker 1: some people think this, some people think that, and some 201 00:13:47,720 --> 00:13:51,400 Speaker 1: people think another thing. Altogether, After all, even though we 202 00:13:51,520 --> 00:13:55,720 Speaker 1: have countless news sites and opinion sites of every different 203 00:13:55,760 --> 00:14:00,520 Speaker 1: political bias. The existence of Wikipedia makes me feel more 204 00:14:00,600 --> 00:14:04,440 Speaker 1: confident that this is a possibility that we continue to 205 00:14:04,720 --> 00:14:08,800 Speaker 1: have ais just like we have today that are unafraid 206 00:14:08,880 --> 00:14:12,280 Speaker 1: to say thank you for your yes or no question. 207 00:14:12,400 --> 00:14:15,760 Speaker 1: But as a metahuman, I'm going to give you a 208 00:14:15,800 --> 00:14:19,600 Speaker 1: more complex answer than perhaps you were looking for, because 209 00:14:19,880 --> 00:14:21,960 Speaker 1: you may or may not know that truth is a 210 00:14:22,080 --> 00:14:26,000 Speaker 1: tricky concept, and we don't have to pretend that most 211 00:14:26,120 --> 00:14:46,320 Speaker 1: questions are binary. Okay, So I wanted to express my 212 00:14:46,400 --> 00:14:51,080 Speaker 1: fear about AI breaking up into individualized models that better 213 00:14:51,200 --> 00:14:55,360 Speaker 1: match the truth of the programmer, and my hope that 214 00:14:55,400 --> 00:15:00,400 Speaker 1: by expressing this clearly we can prevent it. Now, I'd 215 00:15:00,400 --> 00:15:03,240 Speaker 1: like to shift into act two of this episode and 216 00:15:03,320 --> 00:15:05,480 Speaker 1: address the issues that a lot of people are worried 217 00:15:05,480 --> 00:15:08,520 Speaker 1: about when they think about the role of AI in 218 00:15:08,600 --> 00:15:14,120 Speaker 1: the era of disinformation. Now, my suspicion is there are 219 00:15:14,560 --> 00:15:17,960 Speaker 1: trivial cases which people are worried about and they probably 220 00:15:18,000 --> 00:15:20,240 Speaker 1: don't need to be, and then there are the more 221 00:15:20,280 --> 00:15:23,840 Speaker 1: sophisticated cases. So I was at a party the other 222 00:15:23,960 --> 00:15:26,840 Speaker 1: night and ended up talking with an older gentleman who 223 00:15:26,840 --> 00:15:30,240 Speaker 1: told me he was worried about AI and fake news. 224 00:15:30,520 --> 00:15:33,720 Speaker 1: So I asked him why and he said, well, just 225 00:15:33,800 --> 00:15:37,640 Speaker 1: think about the speed of AI. It could manufacture reams 226 00:15:37,680 --> 00:15:40,800 Speaker 1: of fake news in a second. So I asked him, well, 227 00:15:40,840 --> 00:15:44,680 Speaker 1: what would you generate and he said, well, just imagine 228 00:15:45,120 --> 00:15:48,920 Speaker 1: generating fake stories about Joe Biden. So I asked him 229 00:15:48,920 --> 00:15:51,520 Speaker 1: what would you do after you generated that story on 230 00:15:51,560 --> 00:15:55,200 Speaker 1: your desktop? Would you post it to your feed on X? 231 00:15:55,360 --> 00:15:57,880 Speaker 1: And he said yes. So I asked him whether he 232 00:15:57,960 --> 00:16:01,720 Speaker 1: felt sure that would make it difference. Imagine he posts 233 00:16:01,760 --> 00:16:06,680 Speaker 1: on X that Biden just adopted an alien baby. What 234 00:16:06,680 --> 00:16:10,000 Speaker 1: would it matter? Why would anybody listen to this guy's tweet? 235 00:16:10,520 --> 00:16:14,440 Speaker 1: If I make up a story about seeing Bigfoot on 236 00:16:14,520 --> 00:16:18,360 Speaker 1: Stanford campus, it doesn't matter if I have used AI 237 00:16:18,560 --> 00:16:21,160 Speaker 1: to write it or not. And that's because there's no 238 00:16:21,640 --> 00:16:25,760 Speaker 1: further corroboration beyond my claim on social media, so there's 239 00:16:25,800 --> 00:16:30,160 Speaker 1: no reason for anyone to believe it. But maybe, he said, 240 00:16:30,720 --> 00:16:32,720 Speaker 1: maybe it's different if I were to generate a more 241 00:16:32,800 --> 00:16:36,360 Speaker 1: carefully crafted story like that. The New York Times has 242 00:16:36,480 --> 00:16:40,600 Speaker 1: just found evidence that Biden cheated on his taxes, and 243 00:16:40,680 --> 00:16:44,040 Speaker 1: I agree. But I pointed out that even though prompt 244 00:16:44,080 --> 00:16:47,480 Speaker 1: engineering is funds, so you can get just the right story. 245 00:16:48,000 --> 00:16:52,120 Speaker 1: He could currently without AI make up any story he wants, 246 00:16:52,720 --> 00:16:55,600 Speaker 1: and it might take him five minutes to write instead 247 00:16:55,600 --> 00:16:59,160 Speaker 1: of sixty seconds of crafting the right prompt. But other 248 00:16:59,240 --> 00:17:01,680 Speaker 1: than saving those few minutes, it's not as though the 249 00:17:01,760 --> 00:17:05,800 Speaker 1: AI has done something fundamentally different than what he could 250 00:17:05,800 --> 00:17:10,000 Speaker 1: do anyway. And of course, whether penned by him or 251 00:17:10,040 --> 00:17:13,080 Speaker 1: the AI, he still has to try to convince people 252 00:17:13,560 --> 00:17:16,520 Speaker 1: that the story is real, even though it's just a 253 00:17:16,560 --> 00:17:20,359 Speaker 1: tweet and perhaps it links to his blog. But if 254 00:17:20,440 --> 00:17:24,639 Speaker 1: no traditional news source or aggregator has picked up on 255 00:17:24,720 --> 00:17:29,360 Speaker 1: it or posted something like it, it's not news. It's 256 00:17:29,400 --> 00:17:32,800 Speaker 1: going to have a difficult time getting off the starting blocks. 257 00:17:33,200 --> 00:17:36,240 Speaker 1: Now this is not to say that fake news can't spread, 258 00:17:36,280 --> 00:17:39,280 Speaker 1: because it can, but it is to say that it's 259 00:17:39,320 --> 00:17:42,720 Speaker 1: not clear to me how AI is an important player 260 00:17:42,880 --> 00:17:45,919 Speaker 1: in this, except that it saves him four minutes. So 261 00:17:46,320 --> 00:17:49,960 Speaker 1: I think there's some confusion generally about the role that 262 00:17:50,040 --> 00:17:52,480 Speaker 1: AI will play in fake news, and I want to 263 00:17:52,520 --> 00:17:55,480 Speaker 1: be very clear today about where the important difference with 264 00:17:55,640 --> 00:17:59,840 Speaker 1: AI may or may not be. For example, AI could 265 00:18:00,080 --> 00:18:03,399 Speaker 1: play a different kind of role if what I'm doing 266 00:18:03,680 --> 00:18:08,000 Speaker 1: is starting, hundreds of AI bought Twitter accounts that all 267 00:18:08,119 --> 00:18:11,200 Speaker 1: chat back and forth, and at some point they all 268 00:18:11,280 --> 00:18:15,080 Speaker 1: retweet this guy's made up story, and it gains traction 269 00:18:15,600 --> 00:18:19,439 Speaker 1: and believability when somebody sees the number of likes and 270 00:18:19,480 --> 00:18:23,840 Speaker 1: reposts and comments on it. But fundamentally, I think this 271 00:18:23,920 --> 00:18:26,360 Speaker 1: is a cat and mouse game, and the important thing 272 00:18:26,400 --> 00:18:29,679 Speaker 1: will be for social media companies to stay ahead of 273 00:18:29,720 --> 00:18:34,560 Speaker 1: the game in terms of verifying who is a real human. 274 00:18:36,000 --> 00:18:39,399 Speaker 1: Now let's turn to the next issue with AI, which 275 00:18:39,480 --> 00:18:44,920 Speaker 1: is its ability to flawlessly impersonate someone else's voice by 276 00:18:45,000 --> 00:18:49,000 Speaker 1: capturing the cadence and prosity. Now, a few episodes ago, 277 00:18:49,040 --> 00:18:52,200 Speaker 1: I talked about the potential benefits of this. For example, 278 00:18:52,640 --> 00:18:56,160 Speaker 1: I have made a voice AI of my father who 279 00:18:56,200 --> 00:18:59,240 Speaker 1: passed away a few years ago, and it's so wonderful 280 00:18:59,320 --> 00:19:02,679 Speaker 1: and comfort for me to hear his voice. But the 281 00:19:02,760 --> 00:19:05,520 Speaker 1: concern that people have when we're talking about the notion 282 00:19:05,600 --> 00:19:10,440 Speaker 1: of truth is somebody trying to fool you with a voice. 283 00:19:10,760 --> 00:19:14,520 Speaker 1: So here, for example, is a famous person you might 284 00:19:14,600 --> 00:19:18,399 Speaker 1: know you're a bad bad boy. Now, that was Snoop Dogg, 285 00:19:18,720 --> 00:19:21,920 Speaker 1: But as you could probably guess, that wasn't actually Snoop Dogg, 286 00:19:22,000 --> 00:19:26,480 Speaker 1: but an AI generated voice. Now, we all share concerns 287 00:19:26,520 --> 00:19:30,320 Speaker 1: about this capacity to reproduce someone's voice, and I'm going 288 00:19:30,320 --> 00:19:32,359 Speaker 1: to get into some of those concerns in a moment, 289 00:19:32,400 --> 00:19:34,920 Speaker 1: but first I want to play his voice once more. 290 00:19:35,240 --> 00:19:39,000 Speaker 1: You a bad bad boy. Now that's pretty convincing, right, 291 00:19:39,119 --> 00:19:41,920 Speaker 1: maybe even more so than the AI file. But that 292 00:19:42,040 --> 00:19:45,520 Speaker 1: wasn't Snoop Dogg either. That was the performer Keegan Michael, 293 00:19:45,920 --> 00:19:49,119 Speaker 1: doing an impersonation of Snoop Dogg. And the thing to 294 00:19:49,240 --> 00:19:53,359 Speaker 1: note is that impersonators or mimics have been around for 295 00:19:53,520 --> 00:19:57,960 Speaker 1: as long as recorded history. They do awesome voice fakes 296 00:19:58,320 --> 00:20:01,640 Speaker 1: by picking up on the cadence and prosity of someone 297 00:20:01,800 --> 00:20:06,080 Speaker 1: else's voice. So I just want to note that this 298 00:20:06,200 --> 00:20:10,240 Speaker 1: issue is not new. Now. Part of the concern that 299 00:20:10,280 --> 00:20:14,080 Speaker 1: people do have about AI voice generation is that it 300 00:20:14,119 --> 00:20:16,600 Speaker 1: doesn't have to be about a famous person, but instead 301 00:20:16,640 --> 00:20:21,440 Speaker 1: it might be an impersonation of your grandmother or your child, 302 00:20:21,640 --> 00:20:25,439 Speaker 1: because incredibly, these models can get trained up on just 303 00:20:25,480 --> 00:20:29,639 Speaker 1: a few seconds of voice data. So, for example, one 304 00:20:29,720 --> 00:20:33,520 Speaker 1: of the meaningful concerns is that you might receive a 305 00:20:33,560 --> 00:20:36,880 Speaker 1: phone call and you say hello, Hello, I can't hear you. 306 00:20:36,920 --> 00:20:39,359 Speaker 1: Is there anybody there, and then you hang up, and 307 00:20:39,400 --> 00:20:42,920 Speaker 1: that's enough audio data for someone to make a pretty 308 00:20:42,960 --> 00:20:47,199 Speaker 1: flawless impression of your voice. So the idea is that 309 00:20:47,240 --> 00:20:51,720 Speaker 1: the sheer speed of capturing a non celebrity voice makes 310 00:20:51,760 --> 00:20:55,320 Speaker 1: this worrisome. And it's worrisome for a few reasons. One 311 00:20:55,320 --> 00:20:58,680 Speaker 1: of them is that several businesses moved in the last 312 00:20:58,760 --> 00:21:03,920 Speaker 1: few years to voice fingerprinting as their password. So, for example, 313 00:21:04,000 --> 00:21:06,840 Speaker 1: you call the company to find out about your bank 314 00:21:06,840 --> 00:21:09,960 Speaker 1: account or your stock trades, and you simply say, my 315 00:21:10,200 --> 00:21:13,919 Speaker 1: voice is my password, and through a sophisticated process, it 316 00:21:14,000 --> 00:21:17,520 Speaker 1: recognizes that it is, indeed you trying to find out 317 00:21:17,520 --> 00:21:22,520 Speaker 1: information about your account. AI voice generation renders that security 318 00:21:22,560 --> 00:21:27,840 Speaker 1: approach meaningless and even worse. The nightmare scenario is that 319 00:21:28,160 --> 00:21:33,000 Speaker 1: somebody records audio of your child's voice and then you 320 00:21:33,080 --> 00:21:35,240 Speaker 1: get a call one day, maybe you're out of the 321 00:21:35,280 --> 00:21:37,679 Speaker 1: state or out of the country, and your phone rings 322 00:21:37,960 --> 00:21:42,280 Speaker 1: and you hear your child say, help me, I've been kidnapped. 323 00:21:42,560 --> 00:21:44,920 Speaker 1: Please send the money now, or the man is going 324 00:21:44,960 --> 00:21:47,879 Speaker 1: to hurt me. Now. Even if you know about this 325 00:21:47,920 --> 00:21:52,000 Speaker 1: potential future scam, you're still going to hesitate in that situation. 326 00:21:52,080 --> 00:21:54,720 Speaker 1: You're going to be thrown for loop, You're going to panic. 327 00:21:55,080 --> 00:21:58,520 Speaker 1: And the more likely case is that for non listeners 328 00:21:58,560 --> 00:22:01,640 Speaker 1: of this podcast anyway, you've never even heard of such 329 00:22:01,640 --> 00:22:04,399 Speaker 1: a scam in your life, and AI voice generation is 330 00:22:04,520 --> 00:22:08,120 Speaker 1: new to you, and you fall for the scam entirely. 331 00:22:08,920 --> 00:22:11,719 Speaker 1: So again, when we talk about the influence of AI 332 00:22:11,920 --> 00:22:15,320 Speaker 1: on the truth, we might be finding ourselves in many 333 00:22:15,400 --> 00:22:20,119 Speaker 1: scenarios that we could have never seen coming. And of 334 00:22:20,119 --> 00:22:23,479 Speaker 1: course voice is just the beginning. When people think about 335 00:22:23,720 --> 00:22:28,200 Speaker 1: deep fakes, they're often thinking about photographs. So with mid 336 00:22:28,240 --> 00:22:32,560 Speaker 1: Journey or Dolly three or any other good image generation AI, 337 00:22:33,000 --> 00:22:36,160 Speaker 1: you can specify the parameters of your prompt to make 338 00:22:36,200 --> 00:22:42,000 Speaker 1: it so the generated image looks indistinguishable from a real photograph. Now, 339 00:22:42,040 --> 00:22:44,560 Speaker 1: a new science paper came out on this recently. The 340 00:22:44,760 --> 00:22:48,639 Speaker 1: researchers used a set of actual photographs of faces and 341 00:22:48,720 --> 00:22:52,480 Speaker 1: a set of AI generated photographs of faces, and they 342 00:22:52,600 --> 00:22:56,440 Speaker 1: asked the participants to judge or any given photograph whether 343 00:22:56,520 --> 00:22:59,240 Speaker 1: that was a real face or a synthetic face, and 344 00:22:59,320 --> 00:23:04,359 Speaker 1: the participants performed at chance. Their guesses were like throwing 345 00:23:04,480 --> 00:23:08,000 Speaker 1: darts in a dartboard. They were totally random. In fact, 346 00:23:08,160 --> 00:23:12,199 Speaker 1: they rated the AI faces as more trustworthy than the 347 00:23:12,240 --> 00:23:15,760 Speaker 1: real ones, and another study found that the AI faces 348 00:23:15,800 --> 00:23:19,560 Speaker 1: got ranked as more real than the photos of the 349 00:23:19,680 --> 00:23:22,480 Speaker 1: actual faces. So here's the thing I want to mention. 350 00:23:22,720 --> 00:23:25,680 Speaker 1: One of the studies, which came out in October, claimed 351 00:23:25,680 --> 00:23:31,320 Speaker 1: that although participants could not distinguish real and generated faces, 352 00:23:31,800 --> 00:23:37,000 Speaker 1: their unconscious brains could so. Specifically, this research group ran 353 00:23:37,040 --> 00:23:40,879 Speaker 1: an experiment in which the participants wore EEG on their heads. 354 00:23:40,880 --> 00:23:44,879 Speaker 1: That's electro and cephlography, and that measures the electrical activity 355 00:23:44,880 --> 00:23:49,239 Speaker 1: in their brains. Now, the researchers showed real faces or 356 00:23:49,280 --> 00:23:52,440 Speaker 1: synthetic faces, and the claim in the paper is that 357 00:23:52,800 --> 00:23:55,359 Speaker 1: at one hundred and seventy milliseconds, there was a small 358 00:23:55,440 --> 00:24:00,399 Speaker 1: difference in the EEG signal between these two conditions. And 359 00:24:00,440 --> 00:24:03,240 Speaker 1: so the claim of the paper is that while you 360 00:24:03,480 --> 00:24:06,920 Speaker 1: can't tell the difference between the real and the generated faces, 361 00:24:07,320 --> 00:24:11,520 Speaker 1: your brain can. There's a distinction between what we know 362 00:24:11,800 --> 00:24:17,239 Speaker 1: consciously and what our brains have access to unconsciously. And 363 00:24:17,280 --> 00:24:19,919 Speaker 1: so in the wake of that paper, people are suggesting 364 00:24:19,960 --> 00:24:24,080 Speaker 1: that maybe we will have neurally based safeguards in the 365 00:24:24,119 --> 00:24:28,560 Speaker 1: future to mitigate these dangers, because your neural networks will 366 00:24:28,600 --> 00:24:32,600 Speaker 1: be able to tell the difference between real and synthetic photographs. 367 00:24:33,359 --> 00:24:35,879 Speaker 1: I have to say I'm a little skeptical. It strikes 368 00:24:35,880 --> 00:24:39,880 Speaker 1: me there might be some alternative explanations to the study here. 369 00:24:40,240 --> 00:24:43,560 Speaker 1: One thing is that the synthetic faces all tended to 370 00:24:43,560 --> 00:24:48,040 Speaker 1: be more average in their measurements and their configuration, whereas 371 00:24:48,200 --> 00:24:52,240 Speaker 1: real faces can be more distinct, and this presumably explains 372 00:24:52,280 --> 00:24:56,000 Speaker 1: why in that first study people found the AI faces 373 00:24:56,040 --> 00:24:59,960 Speaker 1: as more trustworthy and they judged them to be more real. 374 00:25:00,960 --> 00:25:03,960 Speaker 1: But more importantly, even if the study results are true 375 00:25:03,960 --> 00:25:07,160 Speaker 1: that the participants' brains were able to make some little distinction, 376 00:25:07,800 --> 00:25:12,120 Speaker 1: it won't be true for long because these GENAI models 377 00:25:12,200 --> 00:25:16,240 Speaker 1: get better and better each month, and even if there 378 00:25:16,240 --> 00:25:21,600 Speaker 1: are clues that can somehow be detected now unconsciously, there 379 00:25:21,840 --> 00:25:26,159 Speaker 1: likely will not be for very long. And of course 380 00:25:26,280 --> 00:25:29,840 Speaker 1: audio files and photographs those are just the beginning. Already, 381 00:25:29,880 --> 00:25:33,200 Speaker 1: there are video deep fakes that get better each month. 382 00:25:34,080 --> 00:25:37,160 Speaker 1: Recently I saw a video of Greta Thurnberg, the young 383 00:25:37,280 --> 00:25:41,120 Speaker 1: environmental activist, and she was saying hello, my name is Greta, 384 00:25:41,160 --> 00:25:44,119 Speaker 1: and welcome to my oil company. I love how it 385 00:25:44,200 --> 00:25:46,479 Speaker 1: is pumped out of the ground and it shows her 386 00:25:46,560 --> 00:25:49,000 Speaker 1: working on an oil rig and so on. Obviously, this 387 00:25:49,119 --> 00:25:52,880 Speaker 1: was a fake video generated from other videos of her, 388 00:25:52,960 --> 00:25:56,040 Speaker 1: and the AI moves her mouth appropriately to the words, 389 00:25:56,040 --> 00:25:59,800 Speaker 1: and it does it perfectly. Now, these kinds of deep 390 00:26:00,240 --> 00:26:04,080 Speaker 1: videos are becoming so easy to make that we will 391 00:26:04,200 --> 00:26:07,720 Speaker 1: have to regularly deal with these into the future, and 392 00:26:07,880 --> 00:26:11,520 Speaker 1: just like the photographs, we may soon have an increasingly 393 00:26:11,640 --> 00:26:18,000 Speaker 1: difficult time distinguishing real from synthetic. Now, when it comes 394 00:26:18,080 --> 00:26:22,520 Speaker 1: to the question of deep fakes and misinformation, the problems 395 00:26:22,680 --> 00:26:26,520 Speaker 1: are real. First of all, from a psychology point of view. 396 00:26:27,080 --> 00:26:29,639 Speaker 1: You might watch a deep fake video, let's say, of 397 00:26:29,680 --> 00:26:33,679 Speaker 1: some celebrity saying something violent or racist, and then the 398 00:26:33,800 --> 00:26:38,600 Speaker 1: truth emerges that the video was a deep fake, so 399 00:26:38,640 --> 00:26:42,480 Speaker 1: the celebrity is forgiven, but there's just a little petina 400 00:26:42,720 --> 00:26:45,920 Speaker 1: of negative feeling that sticks with you. And you see 401 00:26:45,920 --> 00:26:47,680 Speaker 1: this happen all the time. By the way, just look 402 00:26:47,680 --> 00:26:51,560 Speaker 1: at any case where somebody is erroneously convicted of a 403 00:26:51,560 --> 00:26:56,800 Speaker 1: crime and then later definitively acquitted. The suspicion remains on 404 00:26:56,840 --> 00:26:58,960 Speaker 1: them like a cloud for the rest of their life, 405 00:26:59,000 --> 00:27:03,240 Speaker 1: totally unfairly. And my nephew Jordan recently suggested to me 406 00:27:03,640 --> 00:27:06,959 Speaker 1: that this is the problem with deep fakes, maybe they 407 00:27:06,960 --> 00:27:09,679 Speaker 1: won't fly into court of law, but they can still 408 00:27:09,840 --> 00:27:14,720 Speaker 1: leverage emotional consequences on social media. Even when the truth 409 00:27:14,800 --> 00:27:17,520 Speaker 1: comes out and a person is verified to be totally 410 00:27:17,840 --> 00:27:21,560 Speaker 1: innocent of whatever was claimed. There's still something that sticks 411 00:27:21,600 --> 00:27:25,320 Speaker 1: with you. Now, because it's been my habit in these 412 00:27:25,400 --> 00:27:29,560 Speaker 1: last few episodes to point out historical precedent to all this, 413 00:27:30,160 --> 00:27:32,119 Speaker 1: I want to note that this is the same game 414 00:27:32,520 --> 00:27:37,240 Speaker 1: that defamation has always been. People can just drop sticky, 415 00:27:37,520 --> 00:27:41,720 Speaker 1: unverifiable statements like I don't think he's loyal to the party, 416 00:27:41,880 --> 00:27:44,520 Speaker 1: or I heard a rumor that she's having an affair, 417 00:27:45,080 --> 00:27:49,040 Speaker 1: or I think that old woman might practice witchcraft. So 418 00:27:49,160 --> 00:27:53,640 Speaker 1: this idea of planting seeds of suspicion is as old 419 00:27:53,640 --> 00:27:57,520 Speaker 1: as the hills. But it is the case that seeing 420 00:27:57,600 --> 00:28:01,280 Speaker 1: something with your own eyes can have a stronger effect 421 00:28:01,720 --> 00:28:05,399 Speaker 1: than mere whispers in the rumor mill, and that's why 422 00:28:05,520 --> 00:28:10,600 Speaker 1: deep fake videos are something too genuinely worry about. Now, 423 00:28:10,640 --> 00:28:13,359 Speaker 1: there's the second point about why the existence of deep 424 00:28:13,359 --> 00:28:16,200 Speaker 1: fake videos is something to worry about, and this one 425 00:28:16,680 --> 00:28:20,119 Speaker 1: is slightly more surprising, which is that the discussion of 426 00:28:20,200 --> 00:28:24,280 Speaker 1: deep fakes is showing up more and more in courts 427 00:28:24,280 --> 00:28:27,159 Speaker 1: of law. But it might not be for the reason 428 00:28:27,200 --> 00:28:30,040 Speaker 1: that you think. It's not that people are making deep 429 00:28:30,080 --> 00:28:32,639 Speaker 1: fakes of other people committing a crime and then trying 430 00:28:32,640 --> 00:28:35,439 Speaker 1: to convict them that way. As far as I can tell, 431 00:28:35,760 --> 00:28:40,040 Speaker 1: it's exactly the opposite. People are committing crimes and getting 432 00:28:40,080 --> 00:28:43,360 Speaker 1: captured on video and then they simply claim that the 433 00:28:43,440 --> 00:28:46,680 Speaker 1: video is a deep fake, and then the prosecution has 434 00:28:46,720 --> 00:28:49,200 Speaker 1: to spend a lot of money and effort and time 435 00:28:49,320 --> 00:28:51,960 Speaker 1: to try to convince the jury that the video is 436 00:28:52,000 --> 00:28:56,040 Speaker 1: actually real. And this is related to another issue, which 437 00:28:56,080 --> 00:28:59,760 Speaker 1: is that we want so strongly to believe things that 438 00:28:59,800 --> 00:29:04,920 Speaker 1: are consistent with our internal model. So whatever evidence comes 439 00:29:04,920 --> 00:29:08,320 Speaker 1: out in the news, people have almost infinite wiggle room 440 00:29:08,440 --> 00:29:12,880 Speaker 1: to say they simply don't believe it. When pictures, an 441 00:29:12,960 --> 00:29:16,440 Speaker 1: audio and video surface that are not consistent with someone's 442 00:29:16,440 --> 00:29:19,720 Speaker 1: political point of view, they can just do what the 443 00:29:19,760 --> 00:29:22,320 Speaker 1: accused person in the court does and say, I don't 444 00:29:22,320 --> 00:29:26,720 Speaker 1: believe it. It is all fake. So all that seems 445 00:29:26,760 --> 00:29:29,520 Speaker 1: a little depressing, but the fact is that the battle 446 00:29:29,640 --> 00:29:35,040 Speaker 1: between truth and misinformation is always a cat and mouse game, 447 00:29:35,600 --> 00:29:39,200 Speaker 1: and there is promising news coming from the technology sector. 448 00:29:39,520 --> 00:29:43,280 Speaker 1: For example, several big companies like Adobe and Microsoft and 449 00:29:43,320 --> 00:29:46,080 Speaker 1: Intel and others. They've all come together to form a 450 00:29:46,200 --> 00:29:49,920 Speaker 1: coalition called C two PA if you're curious. That stands 451 00:29:49,960 --> 00:29:56,160 Speaker 1: for Coalition for Content Provenance and Authenticity. So what C 452 00:29:56,240 --> 00:30:00,240 Speaker 1: two PA sets out to do is to reduce misinformation 453 00:30:00,840 --> 00:30:05,280 Speaker 1: by providing context and history for digital media. And the 454 00:30:05,360 --> 00:30:09,160 Speaker 1: idea is simply to say, precisely where a piece of 455 00:30:09,240 --> 00:30:13,680 Speaker 1: digital media comes from, what is its provenance. The provenance 456 00:30:13,760 --> 00:30:17,120 Speaker 1: is the history of that piece of digital content. When 457 00:30:17,240 --> 00:30:21,720 Speaker 1: was it created, by whom was it modified, what changes 458 00:30:21,840 --> 00:30:25,640 Speaker 1: or updates have been made over time. So the protocol 459 00:30:25,680 --> 00:30:30,600 Speaker 1: they've developed binds the provenance of any image to its metadata, 460 00:30:31,040 --> 00:30:34,880 Speaker 1: and this gives a tamper evident record that always goes 461 00:30:34,920 --> 00:30:37,480 Speaker 1: along with it. So let's say you take a photo 462 00:30:37,520 --> 00:30:40,760 Speaker 1: of your dog outside and say your cell phone has 463 00:30:40,920 --> 00:30:44,320 Speaker 1: C TWOPA on it, which presumably all phones will in 464 00:30:44,360 --> 00:30:47,480 Speaker 1: the near future. So now when you snap the shot, 465 00:30:47,600 --> 00:30:51,480 Speaker 1: all the data about location, time, et cetera. That's all recorded, 466 00:30:51,840 --> 00:30:55,880 Speaker 1: and that's bound to the actual image cryptographically, meaning you 467 00:30:56,080 --> 00:30:59,560 Speaker 1: can't untangle that for the life of this photo. It's 468 00:30:59,800 --> 00:31:04,160 Speaker 1: orin is known. Now you post it on social media 469 00:31:04,200 --> 00:31:06,720 Speaker 1: and anyone can click on the little icon that says 470 00:31:06,960 --> 00:31:12,120 Speaker 1: content credentials, they can see all the provenance information and 471 00:31:12,160 --> 00:31:15,280 Speaker 1: that's how they decide their trust in the image. This 472 00:31:15,320 --> 00:31:19,400 Speaker 1: doesn't make the photo tamper proof, but instead tamper aware, 473 00:31:20,240 --> 00:31:23,840 Speaker 1: because let's say someone comes along and manipulates the photo. 474 00:31:23,880 --> 00:31:27,520 Speaker 1: They add a flying saucer in the background, and they 475 00:31:27,600 --> 00:31:31,959 Speaker 1: make the claim that this is photographic evidence of UAPs. 476 00:31:32,840 --> 00:31:35,840 Speaker 1: Now that change to the photo is recorded in a 477 00:31:35,920 --> 00:31:38,600 Speaker 1: new layer. It's part of the record of the photo 478 00:31:38,960 --> 00:31:42,000 Speaker 1: and its history, and it's locked to the photo forever. 479 00:31:42,080 --> 00:31:45,600 Speaker 1: So when someone clicks on the content credentials now they 480 00:31:45,600 --> 00:31:48,440 Speaker 1: can see that change in any other changes, and they 481 00:31:48,480 --> 00:31:51,840 Speaker 1: can use that information to decide how much they want 482 00:31:51,880 --> 00:31:54,960 Speaker 1: to trust that image. Every bit of the photo's journey 483 00:31:55,000 --> 00:31:59,600 Speaker 1: is recorded, so every photo becomes like an NFT. It's 484 00:31:59,600 --> 00:32:01,840 Speaker 1: not using the blockchain, but the idea is the same 485 00:32:01,920 --> 00:32:06,320 Speaker 1: of making an indelible record. And of course the coalition 486 00:32:06,400 --> 00:32:09,400 Speaker 1: is thought through possible scenarios. So if I use an 487 00:32:09,400 --> 00:32:12,880 Speaker 1: old program to manipulate that photo, and my program doesn't 488 00:32:12,920 --> 00:32:17,520 Speaker 1: have a c TWOPA specification in it, the metadata can 489 00:32:17,560 --> 00:32:20,640 Speaker 1: nonetheless detect that there was tampering and it will show 490 00:32:20,760 --> 00:32:24,400 Speaker 1: up in the change log as an unknown change, but 491 00:32:24,520 --> 00:32:27,960 Speaker 1: a change nonetheless. And even if somebody figures out a 492 00:32:27,960 --> 00:32:31,640 Speaker 1: way to strip the provenance information from the photo, it 493 00:32:31,640 --> 00:32:36,680 Speaker 1: can be rematched using cryptographic techniques. And if you generate 494 00:32:36,840 --> 00:32:41,760 Speaker 1: an image with generative AI, that information will be baked 495 00:32:41,920 --> 00:32:46,360 Speaker 1: into the metadata. So the c TWOPA coalition is pushing 496 00:32:46,400 --> 00:32:50,000 Speaker 1: the US Senate to legislate that this technology will be 497 00:32:50,080 --> 00:32:53,720 Speaker 1: built into all media in the near future. So this 498 00:32:53,800 --> 00:32:58,040 Speaker 1: is pretty cool because for society this represents a one 499 00:32:58,080 --> 00:33:02,600 Speaker 1: step back, two steps forward situation. The one step back 500 00:33:02,800 --> 00:33:06,080 Speaker 1: was that AI photos and videos can be deep faked, 501 00:33:06,080 --> 00:33:11,520 Speaker 1: which suddenly renders everything questionable. But the two steps forward 502 00:33:12,160 --> 00:33:14,880 Speaker 1: is that by taking advantage of tools that we have 503 00:33:15,080 --> 00:33:20,480 Speaker 1: like digital certificates and controlled capture technology and cryptography, we 504 00:33:20,720 --> 00:33:24,959 Speaker 1: now might be able to build something better than anything 505 00:33:25,520 --> 00:33:31,120 Speaker 1: anyone's ever had in history. I mean, imagine that you 506 00:33:31,400 --> 00:33:34,160 Speaker 1: are living in the Soviet Union eighty years ago, and 507 00:33:34,200 --> 00:33:37,000 Speaker 1: I show you this picture of Stalin standing in Red 508 00:33:37,040 --> 00:33:40,200 Speaker 1: Square and there's no Leon Trotsky by his side, and 509 00:33:40,240 --> 00:33:43,960 Speaker 1: you can't quite remember if the original photo had Trotsky 510 00:33:44,040 --> 00:33:46,440 Speaker 1: there or not. And one of your neighbors tells you 511 00:33:46,680 --> 00:33:49,400 Speaker 1: he thinks Trotsky was there, and your other neighbor says 512 00:33:49,920 --> 00:33:52,400 Speaker 1: she's certain this is the way the photo has always been. 513 00:33:53,040 --> 00:33:59,120 Speaker 1: There is no disinterested metahuman arbiter of the truth. You 514 00:33:59,200 --> 00:34:02,760 Speaker 1: have no way of knowing whether the photo you're seeing 515 00:34:02,960 --> 00:34:07,000 Speaker 1: was airbrushed or not. But now it can be clear 516 00:34:07,120 --> 00:34:11,839 Speaker 1: to anyone by clicking on the content Credentials icon what 517 00:34:12,239 --> 00:34:17,319 Speaker 1: precisely happened here. So, at least for the moment, legislation 518 00:34:17,480 --> 00:34:22,560 Speaker 1: like this seems to make truthiness better than it ever was. Now, 519 00:34:22,640 --> 00:34:25,480 Speaker 1: it won't be perfect, because people who want to fake 520 00:34:25,600 --> 00:34:28,839 Speaker 1: something will always find a way. So what I think 521 00:34:28,880 --> 00:34:32,200 Speaker 1: this means is that we're not entering an era of 522 00:34:32,239 --> 00:34:35,240 Speaker 1: post truth, nor are we entering an era of truth. 523 00:34:35,840 --> 00:34:38,879 Speaker 1: This is just the next move in the chess game 524 00:34:39,360 --> 00:34:44,799 Speaker 1: between people who document and people who fake. Now there's 525 00:34:44,840 --> 00:34:48,319 Speaker 1: another issue about AI and truth that I want to 526 00:34:48,360 --> 00:34:51,280 Speaker 1: focus on, and I think this one is the most serious. 527 00:34:51,960 --> 00:34:56,399 Speaker 1: A few episodes ago, I talked about AI relationships and 528 00:34:56,440 --> 00:34:59,640 Speaker 1: the possibility, which is already flickering to life, that a 529 00:34:59,680 --> 00:35:03,360 Speaker 1: lot of people will find appeal in having an AI 530 00:35:03,560 --> 00:35:06,880 Speaker 1: friend or an AI girlfriend, And by the way, this 531 00:35:06,960 --> 00:35:10,320 Speaker 1: might even turn out to help people learn better habits 532 00:35:10,360 --> 00:35:14,960 Speaker 1: of relationships like patience and equanimity. They're learning from the 533 00:35:15,120 --> 00:35:18,359 Speaker 1: avatar that they're in a relationship with which can pay 534 00:35:18,360 --> 00:35:21,120 Speaker 1: off in their real human relationships. So it's going to 535 00:35:21,160 --> 00:35:23,560 Speaker 1: be a fascinating future for us to keep an eye on. 536 00:35:24,320 --> 00:35:28,560 Speaker 1: But there's also a dark side here, which is that 537 00:35:28,600 --> 00:35:33,279 Speaker 1: these AI friends could be built to convince us of 538 00:35:33,360 --> 00:35:36,840 Speaker 1: a particular point of view. Now that might sound like 539 00:35:36,880 --> 00:35:40,279 Speaker 1: a dystopian sci fi story, but the fact is, we 540 00:35:40,480 --> 00:35:46,160 Speaker 1: Homo sapiens, are fundamentally social creatures. We have managed to 541 00:35:46,200 --> 00:35:51,200 Speaker 1: build our societies and cities and civilizations precisely because we 542 00:35:51,239 --> 00:35:56,040 Speaker 1: are so social, and our highly social brains lead us 543 00:35:56,080 --> 00:36:00,279 Speaker 1: to form our truths through discussions with other people, with 544 00:36:00,320 --> 00:36:04,560 Speaker 1: our pals and our parents and our girlfriends or boyfriends. 545 00:36:05,080 --> 00:36:09,960 Speaker 1: This is how we do our information foraging and our 546 00:36:10,160 --> 00:36:14,799 Speaker 1: sense making. And the question is whether AI could now, 547 00:36:14,880 --> 00:36:19,279 Speaker 1: under the worst circumstances, provide a way to hack that, 548 00:36:20,280 --> 00:36:24,200 Speaker 1: to tap into this ancient neural lock and key mechanism 549 00:36:24,239 --> 00:36:27,520 Speaker 1: that we have, and to undermine our species that way. 550 00:36:28,280 --> 00:36:33,200 Speaker 1: So the emergence of AI relationships capable of persuading humans 551 00:36:33,239 --> 00:36:38,440 Speaker 1: towards specific points of view presents a really complex ethical dilemma, 552 00:36:38,520 --> 00:36:47,440 Speaker 1: our social brains, our adept at information gathering through interpersonal relationships, familial, romantic, 553 00:36:48,000 --> 00:36:53,319 Speaker 1: collegial These services the foundation for the construction of our worldviews. 554 00:36:53,360 --> 00:36:56,160 Speaker 1: And I want to point out that this kind of 555 00:36:56,239 --> 00:37:02,720 Speaker 1: manipulation could occur really subtly because AI algorithms could in theory, 556 00:37:03,360 --> 00:37:08,480 Speaker 1: analyze and mimic your behavior to effectively sway your opinion 557 00:37:08,560 --> 00:37:12,399 Speaker 1: without you really having any insight into that. So this 558 00:37:12,560 --> 00:37:17,520 Speaker 1: thought experiment where AI could hack into our neural mechanisms 559 00:37:17,520 --> 00:37:22,320 Speaker 1: of social influence, that raises not only technological and ethical questions, 560 00:37:22,360 --> 00:37:27,120 Speaker 1: but also questions about the vulnerabilities of our social fabric 561 00:37:27,320 --> 00:37:32,440 Speaker 1: Are we as a species equipped to defend against these 562 00:37:32,520 --> 00:37:36,080 Speaker 1: kind of manipulations or could the discovery of this way 563 00:37:36,200 --> 00:37:40,960 Speaker 1: of exploiting our own vulnerabilities be the beginning of the 564 00:37:41,239 --> 00:37:45,320 Speaker 1: end for us. So, as we integrate AI into various 565 00:37:45,400 --> 00:37:48,560 Speaker 1: aspects of our lives, we can't go into this blindly. 566 00:37:49,000 --> 00:37:53,080 Speaker 1: We have to be simulating possibilities and constructing the ethical 567 00:37:53,120 --> 00:37:57,600 Speaker 1: frameworks and the possible regulations that would make sure that 568 00:37:57,680 --> 00:38:03,320 Speaker 1: this new technology doesn't under our most vulnerable security flaw, 569 00:38:04,080 --> 00:38:09,080 Speaker 1: which is our very social neural architecture. That forms its 570 00:38:09,280 --> 00:38:30,600 Speaker 1: truths by talking to others. Now, I want to transition 571 00:38:30,760 --> 00:38:33,319 Speaker 1: into the third act of this episode, and for this 572 00:38:33,400 --> 00:38:35,960 Speaker 1: I'm going to take a totally different angle on the 573 00:38:36,080 --> 00:38:39,160 Speaker 1: question of whether AI will take us off the road 574 00:38:39,239 --> 00:38:44,560 Speaker 1: from the truth. What if the closest we will ever 575 00:38:44,719 --> 00:38:49,200 Speaker 1: get to the truth is via Ai? After all? The 576 00:38:49,320 --> 00:38:52,279 Speaker 1: problem is that as humans, we are biased by the 577 00:38:52,400 --> 00:38:55,720 Speaker 1: very thin trajectories that we take through space and time, 578 00:38:56,200 --> 00:38:59,960 Speaker 1: and we are shaped by our neighborhoods and our cultures 579 00:39:00,040 --> 00:39:04,560 Speaker 1: in our religions. But AI rides above all of that. 580 00:39:04,760 --> 00:39:08,760 Speaker 1: It is the metahuman that can weigh a whole planet 581 00:39:08,800 --> 00:39:15,400 Speaker 1: full of opinions and options. It becomes the oracle, which, 582 00:39:15,440 --> 00:39:19,160 Speaker 1: as a reminder, in ancient Greek mythology, was a person, 583 00:39:19,239 --> 00:39:22,719 Speaker 1: usually a woman, who could communicate with the gods and 584 00:39:22,920 --> 00:39:28,759 Speaker 1: provide divine guidance. The most famous oracles at Delphi and 585 00:39:28,840 --> 00:39:32,560 Speaker 1: Dodona and Olympia. These were consulted by anyone who was 586 00:39:32,600 --> 00:39:36,520 Speaker 1: seeking advice on important decisions like should I marry this 587 00:39:36,680 --> 00:39:40,799 Speaker 1: person or should my nation go to war? And in 588 00:39:40,840 --> 00:39:43,760 Speaker 1: the Greek mythology, the oracle would go into a trance 589 00:39:43,960 --> 00:39:49,400 Speaker 1: like state and communicate with the gods. Now, in these stories, 590 00:39:49,600 --> 00:39:53,440 Speaker 1: oracles played an important role in ancient Greek society because 591 00:39:53,440 --> 00:39:57,239 Speaker 1: they were a source of wisdom and guidance, and their 592 00:39:57,280 --> 00:40:01,360 Speaker 1: advice got sought out by everyone from pow to kings. 593 00:40:02,239 --> 00:40:04,360 Speaker 1: And I'm going to make an argument that we now 594 00:40:04,600 --> 00:40:09,200 Speaker 1: have the technology to build a real oracle and access 595 00:40:09,239 --> 00:40:13,040 Speaker 1: it for pennies, but it's not going to take root 596 00:40:13,160 --> 00:40:17,480 Speaker 1: in modern society, not because of any fault in the technology, 597 00:40:17,920 --> 00:40:22,880 Speaker 1: but because of two facets of human nature. So for 598 00:40:22,960 --> 00:40:26,000 Speaker 1: the first facet, I'll mention that I was talking to 599 00:40:26,280 --> 00:40:28,799 Speaker 1: a colleague recently who was making the argument to me 600 00:40:28,880 --> 00:40:33,000 Speaker 1: that if Bernie Sanders says something about the energy industry, 601 00:40:33,400 --> 00:40:36,360 Speaker 1: then he's immediately shot down as a socialist. So my 602 00:40:36,440 --> 00:40:40,799 Speaker 1: colleagues idea was to get AI to say the same 603 00:40:40,840 --> 00:40:43,680 Speaker 1: thing about energy and then everyone would take it to 604 00:40:43,760 --> 00:40:45,919 Speaker 1: be true, or at least take it in a different way. 605 00:40:46,360 --> 00:40:48,640 Speaker 1: And I thought his idea was really interesting, but I'm 606 00:40:48,680 --> 00:40:51,440 Speaker 1: not so sure I think that this will work for 607 00:40:51,640 --> 00:40:55,600 Speaker 1: very long. And this is for the following reason. As 608 00:40:55,640 --> 00:40:59,840 Speaker 1: I mentioned earlier, I predict we're heading toward a balkanization 609 00:41:00,080 --> 00:41:04,239 Speaker 1: of AI, where different groups will train the AI on 610 00:41:04,480 --> 00:41:08,200 Speaker 1: data that they want and purposely throw out the data 611 00:41:08,200 --> 00:41:10,840 Speaker 1: that they are quite certain is bad anything from the 612 00:41:10,920 --> 00:41:14,360 Speaker 1: other side of the political spectrum. And so the idea 613 00:41:14,440 --> 00:41:18,359 Speaker 1: that we can get AI to tell us the oracular 614 00:41:18,840 --> 00:41:22,319 Speaker 1: truth in a way that everyone will listen to is 615 00:41:22,480 --> 00:41:26,000 Speaker 1: flawed because I think that soon there will be no 616 00:41:26,160 --> 00:41:30,200 Speaker 1: AI oracle that weighs all the evidence equally. And the 617 00:41:30,360 --> 00:41:33,040 Speaker 1: fact is, when it comes to something like what is 618 00:41:33,120 --> 00:41:36,239 Speaker 1: the right way for us to produce energy, there is 619 00:41:36,320 --> 00:41:40,480 Speaker 1: no single right answer. First this is because there are 620 00:41:40,560 --> 00:41:44,879 Speaker 1: always what economists call externalities. So if you say let's 621 00:41:44,920 --> 00:41:48,160 Speaker 1: go entirely solar panels, there will be other things that 622 00:41:48,200 --> 00:41:51,480 Speaker 1: you haven't thought of, like that the solar panels require 623 00:41:51,840 --> 00:41:55,440 Speaker 1: a particular element like molybdenum, and that is rare and 624 00:41:55,440 --> 00:41:57,920 Speaker 1: it has to be minded. And if the whole world's 625 00:41:58,040 --> 00:42:01,560 Speaker 1: energy consumption goes to these new soullar panels, that would 626 00:42:01,560 --> 00:42:05,479 Speaker 1: cause major wars around molybdenum, the way that people fight 627 00:42:05,560 --> 00:42:09,000 Speaker 1: now about diamonds or water. And there would be another 628 00:42:09,160 --> 00:42:13,360 Speaker 1: problem with choosing one technology, let's say solar panels, which 629 00:42:13,400 --> 00:42:16,680 Speaker 1: is what happens when there are cloudy days or in 630 00:42:16,800 --> 00:42:20,520 Speaker 1: a rarer circumstance when a volcano goes off, as happens 631 00:42:20,520 --> 00:42:23,560 Speaker 1: every few millennia and actually blocks out the Sun for 632 00:42:23,600 --> 00:42:26,560 Speaker 1: a while, then we would regret having all of our 633 00:42:26,600 --> 00:42:30,440 Speaker 1: technology made solar. So really what we want is a 634 00:42:30,560 --> 00:42:34,280 Speaker 1: mixture of technologies solar and wind and wave in geothermal 635 00:42:34,680 --> 00:42:39,240 Speaker 1: and nuclear and hydrogen and probably fossil fuels also because 636 00:42:39,280 --> 00:42:42,480 Speaker 1: you never know the future, and the best example we 637 00:42:42,640 --> 00:42:46,120 Speaker 1: have of survival is the way that Mother Nature allows 638 00:42:46,160 --> 00:42:50,440 Speaker 1: life to survive by making a huge menagerie of different 639 00:42:50,480 --> 00:42:54,080 Speaker 1: life forms and seeing which ones survive. There's never a 640 00:42:54,239 --> 00:42:58,800 Speaker 1: single answer about the perfect species. At different times in history, 641 00:42:58,840 --> 00:43:02,040 Speaker 1: different ones survive, and in the future there will always 642 00:43:02,120 --> 00:43:05,720 Speaker 1: be unexpected events which cause some life forms to survive 643 00:43:05,800 --> 00:43:09,320 Speaker 1: and on others. So it's a pretty good guess that 644 00:43:09,320 --> 00:43:12,560 Speaker 1: that's what we would want for our energy portfolio as well, 645 00:43:12,600 --> 00:43:17,680 Speaker 1: to have a very diversified portfolio. So let's imagine that 646 00:43:17,760 --> 00:43:21,560 Speaker 1: we have an AI oracle and it tells us that 647 00:43:21,560 --> 00:43:26,520 Speaker 1: that a big diversified portfolio is the optimal answer. The 648 00:43:26,640 --> 00:43:29,640 Speaker 1: thing I want to concentrate on is that this answer 649 00:43:30,120 --> 00:43:35,799 Speaker 1: won't please any particular political player who has incentives, and 650 00:43:35,840 --> 00:43:38,960 Speaker 1: it certainly won't please anyone who is in business who 651 00:43:39,000 --> 00:43:42,520 Speaker 1: has skin in the game to produce solar panels or 652 00:43:42,560 --> 00:43:46,360 Speaker 1: fossil fuel or nuclear or whatever, and it won't please 653 00:43:46,440 --> 00:43:50,080 Speaker 1: the environmental activist who believes with certainty that she knows 654 00:43:50,120 --> 00:43:53,720 Speaker 1: the right answer. So it may well turn out lots 655 00:43:53,719 --> 00:43:58,759 Speaker 1: of the time that the optimal solution is not one 656 00:43:58,840 --> 00:44:01,840 Speaker 1: that any particular where a human or party of humans 657 00:44:01,960 --> 00:44:05,000 Speaker 1: is going to want to hear, even though it is 658 00:44:05,200 --> 00:44:09,480 Speaker 1: actually the best thing for society in general. And this 659 00:44:09,640 --> 00:44:13,040 Speaker 1: is where we think we want the oracle, but we 660 00:44:13,320 --> 00:44:15,759 Speaker 1: don't when it gets in the way of what we 661 00:44:15,840 --> 00:44:18,960 Speaker 1: want to believe. And hence we return to my prediction 662 00:44:19,200 --> 00:44:24,440 Speaker 1: of balkanization. Different parties will be incentivized to modify the 663 00:44:24,480 --> 00:44:29,360 Speaker 1: oracle by restricting its diet of books and articles, feeding 664 00:44:29,400 --> 00:44:33,319 Speaker 1: it some and not others. So I mentioned earlier the 665 00:44:33,360 --> 00:44:36,440 Speaker 1: political reasons for Balkanization, and here I'm also pointing to 666 00:44:36,480 --> 00:44:42,400 Speaker 1: the economic reasons. Fundamentally, people and businesses are self interested, 667 00:44:42,840 --> 00:44:46,120 Speaker 1: and if the oracle doesn't think your answer is that 668 00:44:46,200 --> 00:44:49,880 Speaker 1: important to the overall picture, you might want to modify that. 669 00:44:51,680 --> 00:44:55,360 Speaker 1: And there's a second reason why ai oracles might not 670 00:44:55,520 --> 00:44:58,920 Speaker 1: take root, even if they should, and it's not because 671 00:44:58,920 --> 00:45:04,359 Speaker 1: of the Balkanization, but also because people will eventually just 672 00:45:04,960 --> 00:45:08,720 Speaker 1: fake it. What do I mean, Well, right now, AI 673 00:45:08,920 --> 00:45:13,960 Speaker 1: has an explainability problem, which means that the networks, with 674 00:45:14,000 --> 00:45:18,520 Speaker 1: their almost two trillion parameters, are so complex that there's 675 00:45:18,520 --> 00:45:21,080 Speaker 1: no way to say, oh, we know precisely why the 676 00:45:21,120 --> 00:45:24,560 Speaker 1: system gave that answer and not a different answer. This 677 00:45:24,640 --> 00:45:27,920 Speaker 1: is part of what has led to the magical quality 678 00:45:27,920 --> 00:45:30,800 Speaker 1: of AI, but I suggest that it's also going to 679 00:45:30,880 --> 00:45:35,960 Speaker 1: lead to another problem. People will simply fake it. In 680 00:45:36,000 --> 00:45:39,520 Speaker 1: other words, I tell you, hey, I ran this two 681 00:45:39,520 --> 00:45:42,520 Speaker 1: trillion parameter model, and I trained it on all the 682 00:45:42,600 --> 00:45:46,160 Speaker 1: data of all the energy sources and supply chains and capacities, 683 00:45:46,400 --> 00:45:50,640 Speaker 1: and it told me its position as an oracle, that 684 00:45:50,719 --> 00:45:53,759 Speaker 1: we should follow the plan that I suggest. Maybe this 685 00:45:53,800 --> 00:45:56,239 Speaker 1: is the plan that benefits my family's business, or the 686 00:45:56,239 --> 00:45:59,600 Speaker 1: plan that leans in my political direction. It would be 687 00:45:59,760 --> 00:46:03,080 Speaker 1: very difficult for you to prove that the AI did 688 00:46:03,120 --> 00:46:05,880 Speaker 1: not tell me that, especially if I show you the 689 00:46:06,239 --> 00:46:10,239 Speaker 1: output screen where it says that you're presumably not going 690 00:46:10,320 --> 00:46:14,680 Speaker 1: to try to reproduce feeding the entire corpus of energy 691 00:46:14,719 --> 00:46:18,200 Speaker 1: economics data into the AI, which I assure you took 692 00:46:18,239 --> 00:46:21,200 Speaker 1: me seventeen months and a team of geniuses to accomplish, 693 00:46:21,520 --> 00:46:24,799 Speaker 1: and so the final analysis of the system has to 694 00:46:24,880 --> 00:46:29,280 Speaker 1: be taken at my word, and society is no dummy. 695 00:46:29,320 --> 00:46:31,600 Speaker 1: So my prediction is that this will have a chance 696 00:46:31,680 --> 00:46:35,160 Speaker 1: of working the first time, and then that strategy will 697 00:46:35,200 --> 00:46:39,000 Speaker 1: collapse as soon as more people take a flyer and say, hey, 698 00:46:39,040 --> 00:46:41,120 Speaker 1: I can't believe it, but this oracle told me that 699 00:46:41,200 --> 00:46:44,760 Speaker 1: my plan is the optimal one. There's going to develop 700 00:46:44,920 --> 00:46:50,440 Speaker 1: a need for provability somehow that an unbiased AI system 701 00:46:50,640 --> 00:46:53,880 Speaker 1: actually yielded that conclusion. But I don't think we're smart 702 00:46:53,960 --> 00:46:57,359 Speaker 1: enough yet to see what that can look like and 703 00:46:57,400 --> 00:47:03,480 Speaker 1: how to build it. And so between balkanization just feeding 704 00:47:03,480 --> 00:47:07,799 Speaker 1: the AI what I want to and manipulation claiming that 705 00:47:07,920 --> 00:47:11,880 Speaker 1: my oracle came to just this right opinion, I'm afraid 706 00:47:12,000 --> 00:47:16,560 Speaker 1: that a possible outcome will be people not trusting the 707 00:47:16,640 --> 00:47:20,399 Speaker 1: output of any AI oracle. They'll trust it just as 708 00:47:20,440 --> 00:47:24,800 Speaker 1: little as they would any particular politician saying we should 709 00:47:24,880 --> 00:47:28,279 Speaker 1: legislate for this solution, or any business person saying we 710 00:47:28,280 --> 00:47:31,560 Speaker 1: should just put our chips on my solution. The oracle 711 00:47:32,080 --> 00:47:36,160 Speaker 1: will degrade to becoming a sock puppet. Now, this is 712 00:47:36,200 --> 00:47:38,799 Speaker 1: obviously the worst case scenario for where things can go, 713 00:47:39,280 --> 00:47:44,239 Speaker 1: because the idea of an AI oracle is extraordinarily appealing. 714 00:47:44,680 --> 00:47:48,319 Speaker 1: I'm just concerned this won't happen, and I'm coining this 715 00:47:48,600 --> 00:47:52,560 Speaker 1: the tall AI syndrome. Now, you've probably heard of the 716 00:47:52,640 --> 00:47:56,120 Speaker 1: tall poppy syndrome, which is that in a field, if 717 00:47:56,120 --> 00:47:59,160 Speaker 1: one poppy grows tall, it's going to get mowed down 718 00:47:59,200 --> 00:48:01,759 Speaker 1: because it's stand out from the rest of the poppies. 719 00:48:02,040 --> 00:48:06,759 Speaker 1: So this expression arose about the tall poppy syndrome, which 720 00:48:06,800 --> 00:48:10,520 Speaker 1: is that if some person is just better than everyone 721 00:48:10,600 --> 00:48:13,760 Speaker 1: around her at something, let's say swimming or math or whatever, 722 00:48:14,160 --> 00:48:17,440 Speaker 1: then people will tend to criticize her and try to 723 00:48:17,440 --> 00:48:21,719 Speaker 1: mow her down. So, in thinking about the future of 724 00:48:21,920 --> 00:48:26,160 Speaker 1: AI and truth, it strikes me as a possibility that 725 00:48:26,200 --> 00:48:29,640 Speaker 1: we are going to run in to a modern version 726 00:48:29,760 --> 00:48:34,319 Speaker 1: of the tall poppy syndrome, the tall AI syndrome. In 727 00:48:34,360 --> 00:48:37,279 Speaker 1: other words, even if AI could be an oracle, the 728 00:48:37,400 --> 00:48:42,040 Speaker 1: kind of truth teller that we've dreamt of since ancient Greece, 729 00:48:42,680 --> 00:48:46,880 Speaker 1: there are many reasons we will not listen. So in 730 00:48:46,880 --> 00:48:50,680 Speaker 1: this episode, I covered the future of truth from the 731 00:48:50,719 --> 00:48:56,640 Speaker 1: point of view of our latest technological invention AI, and 732 00:48:56,680 --> 00:48:59,080 Speaker 1: what I think we can see is that there's good 733 00:48:59,200 --> 00:49:02,239 Speaker 1: and bad to look forward to here, making this just 734 00:49:02,280 --> 00:49:06,080 Speaker 1: like every other technology we've ever developed. And as we 735 00:49:06,120 --> 00:49:10,520 Speaker 1: wrap up this three part series about truth, from news 736 00:49:10,520 --> 00:49:13,600 Speaker 1: stories to the Internet to AI, I hope you'll join 737 00:49:13,719 --> 00:49:16,600 Speaker 1: me in trying to get a clear picture of where 738 00:49:16,600 --> 00:49:19,960 Speaker 1: we are and where we have been and where we're 739 00:49:20,000 --> 00:49:22,560 Speaker 1: going so that we can all try to do what 740 00:49:22,640 --> 00:49:28,319 Speaker 1: we can socially and technologically and legislatively to build a 741 00:49:28,440 --> 00:49:37,600 Speaker 1: more truthy world. Go to Eagleman dot com slash podcast 742 00:49:37,640 --> 00:49:41,000 Speaker 1: for more information and to find further reading. Send me 743 00:49:41,040 --> 00:49:44,320 Speaker 1: an email at podcast at eagleman dot com with questions 744 00:49:44,400 --> 00:49:47,239 Speaker 1: or discussion, and I'll be making episodes in which I 745 00:49:47,280 --> 00:49:52,680 Speaker 1: address those until next time. I'm David Eagleman, and this 746 00:49:52,920 --> 00:50:03,680 Speaker 1: is in our cosmos.