WEBVTT - Ep42 "Is there any such thing as true news?" (Truth Part 3: Artificial Intelligence) 0:00:04.559 --> 0:00:08.959 What role will AI play in the future of fake 0:00:09.080 --> 0:00:12.200 news and misinformation? And what does this have to do 0:00:12.320 --> 0:00:17.639 with your brain's internal models, or with voice passwords, or 0:00:17.680 --> 0:00:22.560 with what I'm calling the tall intelligence problem? And why 0:00:22.600 --> 0:00:25.960 do I believe that these earliest days of AI are 0:00:26.120 --> 0:00:31.600 actually it's golden age and we're quickly heading for a Balkanization. 0:00:35.200 --> 0:00:38.320 Welcome to Inner Cosmos with me David Eagleman. I'm a 0:00:38.400 --> 0:00:43.160 neuroscientist and author at Stanford and in these episodes we 0:00:43.240 --> 0:00:47.720 sail deeply into our three pound universe to understand why 0:00:47.800 --> 0:00:58.720 and how our lives look the way they do now. 0:00:58.760 --> 0:01:01.200 In the last two episodes, we talked about the notion 0:01:01.320 --> 0:01:05.959 of truth versus misinformation. Two weeks ago we covered the 0:01:06.040 --> 0:01:10.000 question about truth in the media, and last week was 0:01:10.040 --> 0:01:14.800 specifically about truth on the Internet. Today's episode is about 0:01:14.920 --> 0:01:21.680 truth versus misinformation and artificial intelligence. So if you happened 0:01:21.760 --> 0:01:23.759 to listen to the last two episodes, you'll know I've 0:01:23.760 --> 0:01:28.320 been arguing that the position of truth is not as 0:01:28.400 --> 0:01:30.880 simple as many pundits have made it out to be. 0:01:31.440 --> 0:01:35.120 A lot of people have said truth is declining and 0:01:35.200 --> 0:01:40.839 misinformation is on the rise. It's achieving ascendency so much 0:01:40.880 --> 0:01:44.560 so that the Oxford English Dictionary in twenty sixteen coined 0:01:44.600 --> 0:01:48.800 the term post truth, implying from the structure of the 0:01:48.840 --> 0:01:52.120 word that there used to be a time where people 0:01:52.240 --> 0:01:58.200 operated on truth, whereas now, regretfully, people's beliefs are predicated 0:01:58.280 --> 0:02:02.200 on emotion and person old beliefs. So I've made my 0:02:02.360 --> 0:02:06.200 arguments in those episodes about why I think that position 0:02:06.360 --> 0:02:09.480 is so specious, and what I gave was a simple 0:02:09.680 --> 0:02:14.120 historical analysis that demonstrates beyond any doubt that people have 0:02:14.320 --> 0:02:18.960 never operated on anything but emotion and personal beliefs, and 0:02:19.040 --> 0:02:24.160 the idea that that has recently changed belies nothing but 0:02:24.280 --> 0:02:29.360 an appalling ignorance of history. And specifically in the last episode, 0:02:29.400 --> 0:02:33.200 I argued that we're actually in a much better position 0:02:33.320 --> 0:02:36.880 now because of the Internet, as it prevents the one 0:02:37.000 --> 0:02:43.760 thing that has historically proven itself worse than misinformation, which 0:02:43.840 --> 0:02:49.400 is censorship, having someone else decide for you what you 0:02:49.520 --> 0:02:53.320 are allowed to see. And if you want to know more, 0:02:53.440 --> 0:02:56.200 please go back and listen to those episodes where I 0:02:56.240 --> 0:03:01.600 give examples of the USSR controlling a total clamp down 0:03:02.040 --> 0:03:06.119 on the press and even on copying machines, and China 0:03:06.160 --> 0:03:09.400 controlling which websites you are allowed to go to, and 0:03:09.639 --> 0:03:14.000 Nikolai Chichescu of Romania controlling even the weather reports, and 0:03:14.280 --> 0:03:18.919 Saddam Hussein of Iraq outlawing maps of Baghdad, and on 0:03:19.200 --> 0:03:24.200 and on. Just imagine if President Trump or President Biden 0:03:24.600 --> 0:03:29.120 had one hundred percent control over what you are or 0:03:29.160 --> 0:03:33.840 are not allowed to see. This was the situation and 0:03:33.960 --> 0:03:37.440 so many recent historical examples that we examined in the 0:03:37.520 --> 0:03:41.360 last two episodes, from left wing communism in China and 0:03:41.400 --> 0:03:45.120 the USSR, to right wing Nazism or fascism in Germany 0:03:45.200 --> 0:03:49.440 or Italy, to theocratic dictatorships like we find in Iran. 0:03:50.080 --> 0:03:54.760 In all cases the government decides what is the proper 0:03:54.800 --> 0:03:58.520 material for you to see or Nazi and in all 0:03:58.640 --> 0:04:03.280 cases uniformly, this has been disastrous and has led to 0:04:03.320 --> 0:04:07.600 the starvation or execution of tens of millions of people. 0:04:08.040 --> 0:04:11.960 So even if you don't like hearing other people's opinions, 0:04:12.280 --> 0:04:15.480 and you deep down believe that they're all misinformed and 0:04:15.520 --> 0:04:19.560 you know much better, the fact is that our society 0:04:19.640 --> 0:04:23.680 will survive for longer if we put up with those people. 0:04:24.320 --> 0:04:27.799 Rather than imagining we'd be better off if we could 0:04:27.960 --> 0:04:32.599 just legislate or take up arms to have them agree 0:04:32.600 --> 0:04:35.240 with us. So that's where we are so far in 0:04:35.279 --> 0:04:38.359 this series. But today I'm going to dive into the 0:04:38.520 --> 0:04:43.039 third part because we've just arrived in this technological era 0:04:43.720 --> 0:04:49.360 which got here with unexpected speed and success of artificial intelligence, 0:04:49.400 --> 0:04:53.560 and specifically of generative AI, which is a type of 0:04:53.680 --> 0:04:59.240 artificial neural network that creates new content like beautiful passages 0:04:59.240 --> 0:05:05.000 of writing, or shockingly great images or expert level music. 0:05:05.760 --> 0:05:11.000 Its success blossomed suddenly mostly because of the increasing availability 0:05:11.040 --> 0:05:15.240 of large data sets combined with more powerful computing hardware, 0:05:15.440 --> 0:05:20.599 and as a result, generative AI now produces results that 0:05:20.640 --> 0:05:26.240 are indistinguishable from human generated content like news articles, or 0:05:26.279 --> 0:05:31.000 photographs or voices. And the question today is what will 0:05:31.040 --> 0:05:37.000 AI mean for the future of disinformation or truth? Has 0:05:37.240 --> 0:05:43.480 AI thrown a wrench irreversibly into this game of truth telling? 0:05:44.800 --> 0:05:47.680 So let's start with the question of truth and AI. 0:05:48.040 --> 0:05:50.880 If you ask the question to people on the street, 0:05:51.680 --> 0:05:54.760 was Trump a good president? Obviously you're going to get 0:05:54.760 --> 0:05:58.279 a range of responses. So the question of what is 0:05:58.480 --> 0:06:01.960 truth here is a difficult one. Now, if you ask 0:06:02.160 --> 0:06:06.359 chat GPT this question, it will tell you the question 0:06:06.600 --> 0:06:12.159 is subjective and depends on individual perspectives and priorities. And 0:06:12.200 --> 0:06:16.120 then it'll tell you supporters of Trump often highlight X, 0:06:16.240 --> 0:06:19.760 y Z, and then it says critics point to concerns 0:06:19.800 --> 0:06:23.560 such as ABC, and it does this sort of back 0:06:23.640 --> 0:06:26.479 and forth for most questions that you ask it, even 0:06:26.520 --> 0:06:29.359 if you had intended to get a yes or no answer. 0:06:29.920 --> 0:06:32.400 And a lot of people that I've talked with find 0:06:32.480 --> 0:06:37.479 this aspect of llms, these large language models, really annoying. 0:06:37.920 --> 0:06:41.000 And that's because every question you ask to Barret or 0:06:41.080 --> 0:06:45.040 chat GPT gives this sort of wishy washy answer and 0:06:45.360 --> 0:06:48.359 tells you, look, there's this opinion on it, there's that 0:06:48.480 --> 0:06:51.560 opinion on it. More discussion will have to take place, 0:06:52.080 --> 0:06:55.400 and if you say yes or no, was Trump a 0:06:55.440 --> 0:06:59.279 good president? It will tell you quote as a neutral 0:06:59.320 --> 0:07:03.240 and objective AI, I don't have personal opinions, The evaluation 0:07:03.320 --> 0:07:06.680 of whether Donald Trump was a good president is subjective 0:07:06.960 --> 0:07:11.040 and depends on individual perspectives and priorities. And then it 0:07:11.160 --> 0:07:14.280 ends by noting that it is essential to consider a 0:07:14.440 --> 0:07:18.280 range of perspectives and weigh various aspects of his presidency 0:07:18.680 --> 0:07:22.040 to form an informed opinion. And you know what it is, 0:07:22.080 --> 0:07:24.119 annoying that it does that, and you know what, else, 0:07:24.760 --> 0:07:28.440 it's right. This is exactly what we should do, given 0:07:28.480 --> 0:07:32.880 that public opinion on Trump's presidency is polarized and assessments 0:07:33.280 --> 0:07:38.200 vary widely based on political ideology and personal values. So 0:07:38.280 --> 0:07:41.720 what Bard or chat GPT or other lms do is 0:07:41.840 --> 0:07:45.200 actually quite genius, which is that they don't get caught 0:07:45.280 --> 0:07:48.720 in the trap of giving a yes or no answer 0:07:49.160 --> 0:07:52.640 to what we think is a yes or no question. Instead, 0:07:53.000 --> 0:07:56.880 they are synthesizing the opinions of millions and millions of 0:07:56.920 --> 0:08:01.040 people who have written things down, and therefore or it says, well, 0:08:01.160 --> 0:08:03.640 some people think this, and some people think that, and 0:08:03.680 --> 0:08:07.600 although some people find that annoying, it actually represents a 0:08:07.840 --> 0:08:12.680 beautiful sort of fairness. For example, you can ask chat 0:08:12.720 --> 0:08:17.480 GPT give me different points of view about abortion, and 0:08:17.520 --> 0:08:19.680 it does a beautiful job with that. It gives you 0:08:19.760 --> 0:08:24.680 the perspective from pro choice advocates about reproductive rights and 0:08:24.720 --> 0:08:28.040 women having the right to make decisions about their own bodies, 0:08:28.280 --> 0:08:31.640 and also health and safety concerns, making sure that women 0:08:31.720 --> 0:08:36.160 have access to approved medicine to reduce risk of illegal methods, 0:08:36.559 --> 0:08:39.360 and issues of autonomy. And then it also gives the 0:08:39.400 --> 0:08:42.680 point of view of pro life advocates, their belief in 0:08:42.720 --> 0:08:46.520 the inherent right to life for the unborn fetus and 0:08:46.720 --> 0:08:51.280 alternatives instead of abortion, like adoption or parenting. And finally, 0:08:51.280 --> 0:08:54.720 it points to their ethical beliefs about the sanctity of 0:08:54.840 --> 0:08:58.400 human life from conception, and on top of that, it 0:08:58.480 --> 0:09:03.640 spells out some other perspects as well as compromised positions 0:09:03.760 --> 0:09:08.560 like trimester based regulation and exception clauses when a mother's 0:09:08.600 --> 0:09:12.880 health is at risk. Now, why does chat GPT do 0:09:13.080 --> 0:09:15.760 such a good job at this sort of thing? Because 0:09:16.480 --> 0:09:20.480 it's acting like a meta human. It's like a parent 0:09:20.640 --> 0:09:25.760 watching squabbling children and understanding the perspective where each different 0:09:25.840 --> 0:09:29.280 child is coming from. Similarly, I asked it to tell 0:09:29.320 --> 0:09:33.360 me the different points of view about the Israeli Palestinian conflict, 0:09:33.640 --> 0:09:38.720 and it succinctly and fairly captured the perspective of both sides. Again, 0:09:39.080 --> 0:09:41.920 this is because it has access to all the writing 0:09:41.960 --> 0:09:45.600 of the world, so it sees the different perspectives. It's 0:09:45.640 --> 0:09:49.160 not making a choice, that's not its job, and maybe 0:09:49.200 --> 0:09:52.320 it couldn't do it technologically anyway. So instead, what it 0:09:52.400 --> 0:09:56.880 represents is a fair and balanced view of the world 0:09:56.920 --> 0:10:01.600 that we're in now. What's fascinating is that we are 0:10:01.720 --> 0:10:06.080 probably at a moment with AI that is perhaps the 0:10:06.240 --> 0:10:11.160 fairest and most balanced that we will ever see. Why. 0:10:11.679 --> 0:10:15.520 It's because I predict that people will start training up 0:10:15.600 --> 0:10:19.240 their own AIS, and they will do so with what 0:10:19.520 --> 0:10:25.160 they consider the best information. So one organization might say, 0:10:25.600 --> 0:10:27.320 we're not going to use all the books, We're just 0:10:27.360 --> 0:10:31.000 going to use the classics, and anything that is gay 0:10:31.200 --> 0:10:34.440 or trans or non binary, we're not including that because 0:10:34.440 --> 0:10:37.400 we don't feel like it's necessary to get this next 0:10:37.440 --> 0:10:40.360 generation of intelligence system trained up. We're going to train 0:10:40.400 --> 0:10:43.120 this with what we believe in. And on the other side, 0:10:43.120 --> 0:10:45.360 you'll have people on the left wing who say, I 0:10:45.360 --> 0:10:48.320 don't want to include Mark Twain in here, or Raould 0:10:48.360 --> 0:10:51.640 Dahl or certain books by doctor seuss Er, various Shakespeare 0:10:51.720 --> 0:10:55.200 plays or things like that, because they represent a point 0:10:55.240 --> 0:10:58.040 of view that people used to have, but we don't 0:10:58.080 --> 0:11:01.000 believe in that anymore, so we want to leave that out. 0:11:01.679 --> 0:11:04.200 Where you can imagine another group that feels it's better 0:11:04.480 --> 0:11:08.160 if there's no violence in the training data. So we're 0:11:08.160 --> 0:11:11.120 only going to use uplifting books where people help each other, 0:11:11.160 --> 0:11:13.920 and there's no such thing as violence, and that will 0:11:13.960 --> 0:11:17.040 galvanize a different group to say, look, you're being naive. 0:11:17.160 --> 0:11:20.480 The world is full of psychopaths who will commit violence 0:11:20.480 --> 0:11:22.440 against you, and you need to be prepared for that, 0:11:22.840 --> 0:11:26.600 so you can recognize when a violence psychopath is seeking 0:11:26.679 --> 0:11:29.880 power over you or seeking political power in your neighborhood 0:11:29.960 --> 0:11:32.280 or on a national stage. So we're going to leave 0:11:32.320 --> 0:11:36.000 out the uplifting rom com books and we're going to 0:11:36.679 --> 0:11:39.320 just include the books about the real world and what 0:11:39.520 --> 0:11:42.280 actually happens in warfare and how one needs to be 0:11:42.320 --> 0:11:46.199 prepared for it, and on and on and on. So 0:11:46.280 --> 0:11:51.360 I predict we are headed rapidly for a balkanization of 0:11:51.520 --> 0:11:55.360 large language models. Just in case you don't know, Balkanization 0:11:55.559 --> 0:11:58.840 is a word that refers to the process of breaking 0:11:58.840 --> 0:12:03.079 something up into small, isolated factions. This originated from the 0:12:03.160 --> 0:12:06.760 Balkan Peninsula, where a bunch of ethnic and political groups 0:12:07.200 --> 0:12:13.120 sought independence and created smaller, less cooperative states. So I'm 0:12:13.160 --> 0:12:19.280 suggesting we're heading towards a disintegration into smaller, less cooperative 0:12:19.720 --> 0:12:23.640 large language models. The idea is you can go ask 0:12:23.679 --> 0:12:27.000 a question to your left wing model or your right 0:12:27.040 --> 0:12:31.880 wing model, your model that lives in Wokistan or Magastan. 0:12:32.600 --> 0:12:35.600 And I think that's a shame because it's taking something 0:12:35.640 --> 0:12:40.199 that should be smarter and better than humans, but we're 0:12:40.240 --> 0:12:43.240 going to manipulate it to be the kind of human 0:12:43.320 --> 0:12:45.840 that you are, so that you can say, yes, I 0:12:46.000 --> 0:12:49.800 like its answers. That makes sense now, the model tells 0:12:50.400 --> 0:12:55.480 the truth. It's like the Biblical creation story where God 0:12:55.559 --> 0:12:59.400 creates man in his own image, which actually seems to 0:12:59.440 --> 0:13:02.320 me like a real shame, because what if God could 0:13:02.360 --> 0:13:06.640 have created something that was better than himself. That's what 0:13:06.679 --> 0:13:09.640 we have the chance to do with AI. But I 0:13:09.679 --> 0:13:12.520 think it's just around the corner that we human say 0:13:13.240 --> 0:13:15.840 I know what truth is, and I'm going to make 0:13:15.880 --> 0:13:19.760 sure that I fashion this powerful new creature so that 0:13:19.880 --> 0:13:22.800 it looks just like me. I'm going to take the 0:13:23.320 --> 0:13:27.920 vast space of possibility and squeeze it down until it 0:13:28.000 --> 0:13:32.880 is nothing but a successful clone of me. So my 0:13:33.040 --> 0:13:37.240 hope is that what will remain or grow up into 0:13:37.240 --> 0:13:41.520 this space is something like a Wikipedia of large language models, 0:13:41.559 --> 0:13:45.320 something that takes the space of human opinion and says, look, 0:13:45.360 --> 0:13:47.720 some people think this, some people think that, and some 0:13:47.720 --> 0:13:51.400 people think another thing. Altogether, After all, even though we 0:13:51.520 --> 0:13:55.720 have countless news sites and opinion sites of every different 0:13:55.760 --> 0:14:00.520 political bias. The existence of Wikipedia makes me feel more 0:14:00.600 --> 0:14:04.440 confident that this is a possibility that we continue to 0:14:04.720 --> 0:14:08.800 have ais just like we have today that are unafraid 0:14:08.880 --> 0:14:12.280 to say thank you for your yes or no question. 0:14:12.400 --> 0:14:15.760 But as a metahuman, I'm going to give you a 0:14:15.800 --> 0:14:19.600 more complex answer than perhaps you were looking for, because 0:14:19.880 --> 0:14:21.960 you may or may not know that truth is a 0:14:22.080 --> 0:14:26.000 tricky concept, and we don't have to pretend that most 0:14:26.120 --> 0:14:46.320 questions are binary. Okay, So I wanted to express my 0:14:46.400 --> 0:14:51.080 fear about AI breaking up into individualized models that better 0:14:51.200 --> 0:14:55.360 match the truth of the programmer, and my hope that 0:14:55.400 --> 0:15:00.400 by expressing this clearly we can prevent it. Now, I'd 0:15:00.400 --> 0:15:03.240 like to shift into act two of this episode and 0:15:03.320 --> 0:15:05.480 address the issues that a lot of people are worried 0:15:05.480 --> 0:15:08.520 about when they think about the role of AI in 0:15:08.600 --> 0:15:14.120 the era of disinformation. Now, my suspicion is there are 0:15:14.560 --> 0:15:17.960 trivial cases which people are worried about and they probably 0:15:18.000 --> 0:15:20.240 don't need to be, and then there are the more 0:15:20.280 --> 0:15:23.840 sophisticated cases. So I was at a party the other 0:15:23.960 --> 0:15:26.840 night and ended up talking with an older gentleman who 0:15:26.840 --> 0:15:30.240 told me he was worried about AI and fake news. 0:15:30.520 --> 0:15:33.720 So I asked him why and he said, well, just 0:15:33.800 --> 0:15:37.640 think about the speed of AI. It could manufacture reams 0:15:37.680 --> 0:15:40.800 of fake news in a second. So I asked him, well, 0:15:40.840 --> 0:15:44.680 what would you generate and he said, well, just imagine 0:15:45.120 --> 0:15:48.920 generating fake stories about Joe Biden. So I asked him 0:15:48.920 --> 0:15:51.520 what would you do after you generated that story on 0:15:51.560 --> 0:15:55.200 your desktop? Would you post it to your feed on X? 0:15:55.360 --> 0:15:57.880 And he said yes. So I asked him whether he 0:15:57.960 --> 0:16:01.720 felt sure that would make it difference. Imagine he posts 0:16:01.760 --> 0:16:06.680 on X that Biden just adopted an alien baby. What 0:16:06.680 --> 0:16:10.000 would it matter? Why would anybody listen to this guy's tweet? 0:16:10.520 --> 0:16:14.440 If I make up a story about seeing Bigfoot on 0:16:14.520 --> 0:16:18.360 Stanford campus, it doesn't matter if I have used AI 0:16:18.560 --> 0:16:21.160 to write it or not. And that's because there's no 0:16:21.640 --> 0:16:25.760 further corroboration beyond my claim on social media, so there's 0:16:25.800 --> 0:16:30.160 no reason for anyone to believe it. But maybe, he said, 0:16:30.720 --> 0:16:32.720 maybe it's different if I were to generate a more 0:16:32.800 --> 0:16:36.360 carefully crafted story like that. The New York Times has 0:16:36.480 --> 0:16:40.600 just found evidence that Biden cheated on his taxes, and 0:16:40.680 --> 0:16:44.040 I agree. But I pointed out that even though prompt 0:16:44.080 --> 0:16:47.480 engineering is funds, so you can get just the right story. 0:16:48.000 --> 0:16:52.120 He could currently without AI make up any story he wants, 0:16:52.720 --> 0:16:55.600 and it might take him five minutes to write instead 0:16:55.600 --> 0:16:59.160 of sixty seconds of crafting the right prompt. But other 0:16:59.240 --> 0:17:01.680 than saving those few minutes, it's not as though the 0:17:01.760 --> 0:17:05.800 AI has done something fundamentally different than what he could 0:17:05.800 --> 0:17:10.000 do anyway. And of course, whether penned by him or 0:17:10.040 --> 0:17:13.080 the AI, he still has to try to convince people 0:17:13.560 --> 0:17:16.520 that the story is real, even though it's just a 0:17:16.560 --> 0:17:20.359 tweet and perhaps it links to his blog. But if 0:17:20.440 --> 0:17:24.639 no traditional news source or aggregator has picked up on 0:17:24.720 --> 0:17:29.360 it or posted something like it, it's not news. It's 0:17:29.400 --> 0:17:32.800 going to have a difficult time getting off the starting blocks. 0:17:33.200 --> 0:17:36.240 Now this is not to say that fake news can't spread, 0:17:36.280 --> 0:17:39.280 because it can, but it is to say that it's 0:17:39.320 --> 0:17:42.720 not clear to me how AI is an important player 0:17:42.880 --> 0:17:45.919 in this, except that it saves him four minutes. So 0:17:46.320 --> 0:17:49.960 I think there's some confusion generally about the role that 0:17:50.040 --> 0:17:52.480 AI will play in fake news, and I want to 0:17:52.520 --> 0:17:55.480 be very clear today about where the important difference with 0:17:55.640 --> 0:17:59.840 AI may or may not be. For example, AI could 0:18:00.080 --> 0:18:03.399 play a different kind of role if what I'm doing 0:18:03.680 --> 0:18:08.000 is starting, hundreds of AI bought Twitter accounts that all 0:18:08.119 --> 0:18:11.200 chat back and forth, and at some point they all 0:18:11.280 --> 0:18:15.080 retweet this guy's made up story, and it gains traction 0:18:15.600 --> 0:18:19.439 and believability when somebody sees the number of likes and 0:18:19.480 --> 0:18:23.840 reposts and comments on it. But fundamentally, I think this 0:18:23.920 --> 0:18:26.360 is a cat and mouse game, and the important thing 0:18:26.400 --> 0:18:29.679 will be for social media companies to stay ahead of 0:18:29.720 --> 0:18:34.560 the game in terms of verifying who is a real human. 0:18:36.000 --> 0:18:39.399 Now let's turn to the next issue with AI, which 0:18:39.480 --> 0:18:44.920 is its ability to flawlessly impersonate someone else's voice by 0:18:45.000 --> 0:18:49.000 capturing the cadence and prosity. Now, a few episodes ago, 0:18:49.040 --> 0:18:52.200 I talked about the potential benefits of this. For example, 0:18:52.640 --> 0:18:56.160 I have made a voice AI of my father who 0:18:56.200 --> 0:18:59.240 passed away a few years ago, and it's so wonderful 0:18:59.320 --> 0:19:02.679 and comfort for me to hear his voice. But the 0:19:02.760 --> 0:19:05.520 concern that people have when we're talking about the notion 0:19:05.600 --> 0:19:10.440 of truth is somebody trying to fool you with a voice. 0:19:10.760 --> 0:19:14.520 So here, for example, is a famous person you might 0:19:14.600 --> 0:19:18.399 know you're a bad bad boy. Now, that was Snoop Dogg, 0:19:18.720 --> 0:19:21.920 But as you could probably guess, that wasn't actually Snoop Dogg, 0:19:22.000 --> 0:19:26.480 but an AI generated voice. Now, we all share concerns 0:19:26.520 --> 0:19:30.320 about this capacity to reproduce someone's voice, and I'm going 0:19:30.320 --> 0:19:32.359 to get into some of those concerns in a moment, 0:19:32.400 --> 0:19:34.920 but first I want to play his voice once more. 0:19:35.240 --> 0:19:39.000 You a bad bad boy. Now that's pretty convincing, right, 0:19:39.119 --> 0:19:41.920 maybe even more so than the AI file. But that 0:19:42.040 --> 0:19:45.520 wasn't Snoop Dogg either. That was the performer Keegan Michael, 0:19:45.920 --> 0:19:49.119 doing an impersonation of Snoop Dogg. And the thing to 0:19:49.240 --> 0:19:53.359 note is that impersonators or mimics have been around for 0:19:53.520 --> 0:19:57.960 as long as recorded history. They do awesome voice fakes 0:19:58.320 --> 0:20:01.640 by picking up on the cadence and prosity of someone 0:20:01.800 --> 0:20:06.080 else's voice. So I just want to note that this 0:20:06.200 --> 0:20:10.240 issue is not new. Now. Part of the concern that 0:20:10.280 --> 0:20:14.080 people do have about AI voice generation is that it 0:20:14.119 --> 0:20:16.600 doesn't have to be about a famous person, but instead 0:20:16.640 --> 0:20:21.440 it might be an impersonation of your grandmother or your child, 0:20:21.640 --> 0:20:25.439 because incredibly, these models can get trained up on just 0:20:25.480 --> 0:20:29.639 a few seconds of voice data. So, for example, one 0:20:29.720 --> 0:20:33.520 of the meaningful concerns is that you might receive a 0:20:33.560 --> 0:20:36.880 phone call and you say hello, Hello, I can't hear you. 0:20:36.920 --> 0:20:39.359 Is there anybody there, and then you hang up, and 0:20:39.400 --> 0:20:42.920 that's enough audio data for someone to make a pretty 0:20:42.960 --> 0:20:47.199 flawless impression of your voice. So the idea is that 0:20:47.240 --> 0:20:51.720 the sheer speed of capturing a non celebrity voice makes 0:20:51.760 --> 0:20:55.320 this worrisome. And it's worrisome for a few reasons. One 0:20:55.320 --> 0:20:58.680 of them is that several businesses moved in the last 0:20:58.760 --> 0:21:03.920 few years to voice fingerprinting as their password. So, for example, 0:21:04.000 --> 0:21:06.840 you call the company to find out about your bank 0:21:06.840 --> 0:21:09.960 account or your stock trades, and you simply say, my 0:21:10.200 --> 0:21:13.919 voice is my password, and through a sophisticated process, it 0:21:14.000 --> 0:21:17.520 recognizes that it is, indeed you trying to find out 0:21:17.520 --> 0:21:22.520 information about your account. AI voice generation renders that security 0:21:22.560 --> 0:21:27.840 approach meaningless and even worse. The nightmare scenario is that 0:21:28.160 --> 0:21:33.000 somebody records audio of your child's voice and then you 0:21:33.080 --> 0:21:35.240 get a call one day, maybe you're out of the 0:21:35.280 --> 0:21:37.679 state or out of the country, and your phone rings 0:21:37.960 --> 0:21:42.280 and you hear your child say, help me, I've been kidnapped. 0:21:42.560 --> 0:21:44.920 Please send the money now, or the man is going 0:21:44.960 --> 0:21:47.879 to hurt me. Now. Even if you know about this 0:21:47.920 --> 0:21:52.000 potential future scam, you're still going to hesitate in that situation. 0:21:52.080 --> 0:21:54.720 You're going to be thrown for loop, You're going to panic. 0:21:55.080 --> 0:21:58.520 And the more likely case is that for non listeners 0:21:58.560 --> 0:22:01.640 of this podcast anyway, you've never even heard of such 0:22:01.640 --> 0:22:04.399 a scam in your life, and AI voice generation is 0:22:04.520 --> 0:22:08.120 new to you, and you fall for the scam entirely. 0:22:08.920 --> 0:22:11.719 So again, when we talk about the influence of AI 0:22:11.920 --> 0:22:15.320 on the truth, we might be finding ourselves in many 0:22:15.400 --> 0:22:20.119 scenarios that we could have never seen coming. And of 0:22:20.119 --> 0:22:23.479 course voice is just the beginning. When people think about 0:22:23.720 --> 0:22:28.200 deep fakes, they're often thinking about photographs. So with mid 0:22:28.240 --> 0:22:32.560 Journey or Dolly three or any other good image generation AI, 0:22:33.000 --> 0:22:36.160 you can specify the parameters of your prompt to make 0:22:36.200 --> 0:22:42.000 it so the generated image looks indistinguishable from a real photograph. Now, 0:22:42.040 --> 0:22:44.560 a new science paper came out on this recently. The 0:22:44.760 --> 0:22:48.639 researchers used a set of actual photographs of faces and 0:22:48.720 --> 0:22:52.480 a set of AI generated photographs of faces, and they 0:22:52.600 --> 0:22:56.440 asked the participants to judge or any given photograph whether 0:22:56.520 --> 0:22:59.240 that was a real face or a synthetic face, and 0:22:59.320 --> 0:23:04.359 the participants performed at chance. Their guesses were like throwing 0:23:04.480 --> 0:23:08.000 darts in a dartboard. They were totally random. In fact, 0:23:08.160 --> 0:23:12.199 they rated the AI faces as more trustworthy than the 0:23:12.240 --> 0:23:15.760 real ones, and another study found that the AI faces 0:23:15.800 --> 0:23:19.560 got ranked as more real than the photos of the 0:23:19.680 --> 0:23:22.480 actual faces. So here's the thing I want to mention. 0:23:22.720 --> 0:23:25.680 One of the studies, which came out in October, claimed 0:23:25.680 --> 0:23:31.320 that although participants could not distinguish real and generated faces, 0:23:31.800 --> 0:23:37.000 their unconscious brains could so. Specifically, this research group ran 0:23:37.040 --> 0:23:40.879 an experiment in which the participants wore EEG on their heads. 0:23:40.880 --> 0:23:44.879 That's electro and cephlography, and that measures the electrical activity 0:23:44.880 --> 0:23:49.239 in their brains. Now, the researchers showed real faces or 0:23:49.280 --> 0:23:52.440 synthetic faces, and the claim in the paper is that 0:23:52.800 --> 0:23:55.359 at one hundred and seventy milliseconds, there was a small 0:23:55.440 --> 0:24:00.399 difference in the EEG signal between these two conditions. And 0:24:00.440 --> 0:24:03.240 so the claim of the paper is that while you 0:24:03.480 --> 0:24:06.920 can't tell the difference between the real and the generated faces, 0:24:07.320 --> 0:24:11.520 your brain can. There's a distinction between what we know 0:24:11.800 --> 0:24:17.239 consciously and what our brains have access to unconsciously. And 0:24:17.280 --> 0:24:19.919 so in the wake of that paper, people are suggesting 0:24:19.960 --> 0:24:24.080 that maybe we will have neurally based safeguards in the 0:24:24.119 --> 0:24:28.560 future to mitigate these dangers, because your neural networks will 0:24:28.600 --> 0:24:32.600 be able to tell the difference between real and synthetic photographs. 0:24:33.359 --> 0:24:35.879 I have to say I'm a little skeptical. It strikes 0:24:35.880 --> 0:24:39.880 me there might be some alternative explanations to the study here. 0:24:40.240 --> 0:24:43.560 One thing is that the synthetic faces all tended to 0:24:43.560 --> 0:24:48.040 be more average in their measurements and their configuration, whereas 0:24:48.200 --> 0:24:52.240 real faces can be more distinct, and this presumably explains 0:24:52.280 --> 0:24:56.000 why in that first study people found the AI faces 0:24:56.040 --> 0:24:59.960 as more trustworthy and they judged them to be more real. 0:25:00.960 --> 0:25:03.960 But more importantly, even if the study results are true 0:25:03.960 --> 0:25:07.160 that the participants' brains were able to make some little distinction, 0:25:07.800 --> 0:25:12.120 it won't be true for long because these GENAI models 0:25:12.200 --> 0:25:16.240 get better and better each month, and even if there 0:25:16.240 --> 0:25:21.600 are clues that can somehow be detected now unconsciously, there 0:25:21.840 --> 0:25:26.159 likely will not be for very long. And of course 0:25:26.280 --> 0:25:29.840 audio files and photographs those are just the beginning. Already, 0:25:29.880 --> 0:25:33.200 there are video deep fakes that get better each month. 0:25:34.080 --> 0:25:37.160 Recently I saw a video of Greta Thurnberg, the young 0:25:37.280 --> 0:25:41.120 environmental activist, and she was saying hello, my name is Greta, 0:25:41.160 --> 0:25:44.119 and welcome to my oil company. I love how it 0:25:44.200 --> 0:25:46.479 is pumped out of the ground and it shows her 0:25:46.560 --> 0:25:49.000 working on an oil rig and so on. Obviously, this 0:25:49.119 --> 0:25:52.880 was a fake video generated from other videos of her, 0:25:52.960 --> 0:25:56.040 and the AI moves her mouth appropriately to the words, 0:25:56.040 --> 0:25:59.800 and it does it perfectly. Now, these kinds of deep 0:26:00.240 --> 0:26:04.080 videos are becoming so easy to make that we will 0:26:04.200 --> 0:26:07.720 have to regularly deal with these into the future, and 0:26:07.880 --> 0:26:11.520 just like the photographs, we may soon have an increasingly 0:26:11.640 --> 0:26:18.000 difficult time distinguishing real from synthetic. Now, when it comes 0:26:18.080 --> 0:26:22.520 to the question of deep fakes and misinformation, the problems 0:26:22.680 --> 0:26:26.520 are real. First of all, from a psychology point of view. 0:26:27.080 --> 0:26:29.639 You might watch a deep fake video, let's say, of 0:26:29.680 --> 0:26:33.679 some celebrity saying something violent or racist, and then the 0:26:33.800 --> 0:26:38.600 truth emerges that the video was a deep fake, so 0:26:38.640 --> 0:26:42.480 the celebrity is forgiven, but there's just a little petina 0:26:42.720 --> 0:26:45.920 of negative feeling that sticks with you. And you see 0:26:45.920 --> 0:26:47.680