Q&A: There is No Algorithm for Truth - with Tom Scott

When you first started out, did you plan to have several platforms for your you know, YouTube, Twitter, Facebook, everything, or did you stick with one? I spent 15 years throwing everything I could at every platform that would happen. I did not know what would take off and what wouldn't. And I had those one-hit wonders on various platforms for a while. But yeah, it took 10-15 years before, luckily, extremely luckily, the one that worked was YouTube. I got lucky that it was the one that is able to pay my rent, and also about true. There is a version of me that got lucky with something else a year earlier, and has now spent five years honing his craft to stand up to me. I did try to do that once. Turns out I'm not. See? Great. Another question, please. I'm worried about deepfakes. So am I. And you talked about the gatekeepers. Is it not true that those deepfakes could become the gatekeepers of the future? Oh, I do hope not. For those who don't know, deepfakes are, again, artificial intelligence, machine learning systems that can essentially swap and generate faces. So if you've ever seen the video of the one from a few years ago of Barack Obama just swearing and doing some comedy lines, because they got someone to impersonate his voice and then just put his face on. I wish I knew how that's going to turn out. It is going to make some incredibly convincing fake news. But given that people on Facebook can currently be fooled with a headline and a domain name that no one's heard of, I don't know. I think the problem will be when that convinces someone at one of the gatekeepers' talks. When we get one of the scandals, I think it was during the Bush administration, when the slander of one American news organization, I can't remember which three letters it was, took a memo that had blatantly been printed in Microsoft Word as a proper typewritten memo from many years earlier about either Bush or Kerry's war record, something like that. That was a deepfake and it was done with Microsoft Word and no technical skill whatsoever and it convinced a news agency that wanted to build it. I think it would be possibly easier to convince people of that, but I don't know at which point that becomes a significant threat. I hope that answers the question. Okay, next question, please. With a lot of these recent recommendator systems and Google, unfortunately a lot of these platforms will have to make the choice of what's popular and what's to be shown. If someone wants to search for cats, they'll have to decide what kind of cat to show people. What are some of your thoughts on the things that these algorithms should optimize or ways that we can check on these algorithms, given the fact that you've said that existing algorithms can be where people will cheat them, or alternatively that these are also tied to, for example, companies' bottom lines? In an ideal world, then we would optimize for truth. In an ideal world, we would be able to look at something about homeopathic vaccines and go, well, yeah, they shouldn't pass that on. But how you do that without also giving that power to people with very different views, I don't know. Yesterday, actually I didn't have time to listen to the presentation. Yesterday, The Guardian leaked the moderation guidelines for a company called TikTok, which is short video clips that are generally not used by anyone over the age of 25. It baffles me. According to the article in The Guardian, their moderation rules for Turkey include blocking, and by that, they mean they're just not amplifying. You put it out, they just won't show it to anyone else. Anything to do with the rainbow flag or anything to do with LGBT stuff. That is not illegal in Turkey, but the Chinese company that runs TikTok has decided to do that anyway. So I am really wary of giving that recommendation. It's why I preface that with, in an ideal world. Because if you give that power to people, it can be used by the... I hate saying it, because it's not that polarized, and it's more complicated than that, but it can be used by the other side. I'm also wary of saying, well, it should be truly neutral, because algorithmic bias is there. If you can somehow remove algorithmic bias, yeah, absolutely, it should be okay. Do you have a follow-up? In some senses, the bias in these algorithms is actually what makes them useful, because if you want to search for cats, it is probably a good idea that the algorithm knows what kind of cats you like to show you that cat, which is what makes... Otherwise, there would be a plain database. You would just look up cats, and these are everything. One thing that I didn't go into, because I'm talking about more mass algorithmic stuff, is that every one of these recommendations is going to be personalized for you. The amount of data tracking that's involved in that is astonishing, and it's kind of outside the scope of what I'm doing. But yeah, the trouble is how much you try to convert people to your side. I certainly do think that the big tech companies have a moral imperative to not let the Nazis in, but where you draw that line, I don't know. Shall we take a question from the balcony, please? Hello up there, by the way. I was mostly looking down there. Sorry, you folks at home. It's absolutely blinding if I look up there. Yeah, this is somewhat of a follow-up to the previous question. I work in tech, and something that we struggle with is that we don't let our biases affect the platforms that we build. And the thing that I'm kind of worried about is the echo chamber effect. So as people in tech, mainly coming from the Bay Area and building these platforms, it will become more homogenized to what people in tech want to do. So I'm kind of wondering what your thoughts are on that, and also as a follow-up, where government intervention should come in for what should and should be regulated here. So this is a question that should be answered by a large and diverse group of people, because that's what you need for something like that. If it is a homogenous Bay Area group, then yeah, of course their biases are going to infect the platform, and I think that's probably the right word to use. But equally, TikTok is an app coming out of China. Their biases are going to infect the platform. The solution is, as I think you implied there, having a diverse hiring process, making sure you've got people from all sorts of social and economic and every background in there to try and make the product work. But God knows that's a talk in itself and an hour-long dissertation on how you make that work. How you do government intervention on that. You could require certain percentages of certain sex, ethnicity, gender, ethnicity, anything like that. I can't go the wrong way round twice in this presentation. It sucks. Sorry. You could require that. I do not have the data to know whether that's a good idea. I think the data is probably out there. I'm not familiar with it. Great. Let's take the next question from here, please. Hi. I thought the talk was fantastic, by the way. Thank you. What do you think is more likely to lead to radicalisation? An echo chamber or a place where there's deregulations on free speech? So, an echo chamber. But if you've got that and the people are in it, you will find that radicalisation. I'm talking about the example of Reddit. The whole place was set up as anyone can say anything. But within the communities there, they could freely choose to turn themselves into echo chambers. So I think it's more likely to happen in one of that when someone gets there. But I also think that if you have this big open site like YouTube with a recommendation engine that can loosely connect communities all over the place, then that's more likely to create an attractor that leads to that. That is a horribly fudged and half-arsed answer. I apologise for that. I think if you're asking me if you've got a site where anyone can say anything, and you've got a site where it is just this small group, I would worry about the small group more, but I would worry about the big site turning into that. Thank you. Question. James. Thank you. Hi. I noticed you did not talk about coordinated inauthentic behaviour. And I'm curious, though there may not be an algorithm for truth, are you perhaps more optimistic about our battle against bots? And can we get to a point through biometrics or something like that in which we can actually verify whether or not this content is at least coming from a real person and thus take a big step in the war for protecting truth? Yeah, it's a very, very difficult choice. So coordinated inauthentic behaviour is colloquially known as the Russian bot farm. It is large groups of people working together to push a political point, to shout people down. Twitter is actually doing some reasonably good work tracking that. They have a transparency thing that lets you see how many people they have banned for coordinated inauthentic behaviour. But that's, of course, only the ones they've spotted. The solution to that is essentially the opposite of everything the web was built for. The web was built to be something that interpreted censorship as damage and rooted around it. And it was built to be that you could be anonymous, that you could be whoever you wanted to be. Without completely restructuring the web, I don't think that's a good thing. Because we've seen what happens on Facebook. For a while, Facebook required people to use their real name. I think they still do. They've stopped doing this quite as much now. But you could report people for having an inauthentic name. And you would be required to send an ID with that name on it. Which was obviously used to attack all sorts of groups. It was used to attack trans people who don't have their ID in the same name as the one they use. If you require everything to be posted by a real person, yes, that will silence coordinated inauthentic behaviour. But it will also silence people who are afraid of speaking out. I think there was a news article, and God knows, don't quote me on this, that said there was... Again, for anyone watching this on the internet, it's September 2019. That there was a Republican majority in the Senate for impeaching Donald Trump. In an anonymous vote. But all their votes are recorded. And so they have that problem on a small scale. This is a long, long way of saying, again, I don't know. That's two extremes. And you can't find that wonderful place in between. Because as with anything that works on that scale, you can't have both. How optimistic am I? Not very. Because, you know, it's been a long time since I've been on the other end of this. But, you know, eight, ten years ago I was setting up Twitter bots. Not for malicious purposes, but just for projects. And it's not difficult. It's designed not to be difficult. Because the more roadblocks you put in the way of someone joining the site, the less users sign up. I can put faith in some machine learning systems starting to detect them. But I think there's more likely to be an arms race of detection and evolution and detection and evolution. I'll take the last question from the gentleman in two rows in front of the previous one. Along with James. This gentleman in the blue. In the Superman t-shirt. Hi. I think your discourse was quite interesting. When we talk about Echo Chamber, the Nazi bomb and all the rest of it. One thing, probably for some reason you covered or did not need to cover, is about profiling. So the likes of the Dissolve company, Cambridge Analytica, they had narrowed it down to science. I don't remember. 82 or 86 profiles. And when we talk of something like that, that ends up promoting the whole idea of Echo Chambers. Which then creates a bubble. Which is essentially segmentation. Which goes away from the personal recommendation and creates groups of individuals who are programmed to do certain things. What's your view on that? I think there's a large and important distinction between the advertising targeting there. Which is very much like the segmentation that they have there. And the segmentation that any advertiser has is enormous. But I do think there's a distinction between that segmentation. Which is creepy. And the sort of automated recommendation engine. Which is profiling you based on the more anonymous things like what you watch. And I think there's a difference between the two. They're certainly creepy in very different ways. Trying to figure out what the difference between them is. I think it's the human element in there. Cambridge Analytica was using AI to track people and target that sort of thing. But there was someone going, this is what we want these people to believe. We're going to target people who are interested in animals and live near the coast. And we're going to tell them that Brexit could affect animal rights or ocean rights. We're going to target people who live in the cities and are on low incomes. And tell them that Brexit is going to help their jobs and help their income. I think it's very different for a human to make those decisions. And say, these are the beliefs we're going to try and force down people's throats. I think that's very different to an algorithm accidentally creating these rabbit holes. That's a horrible point to end on. Oh my god, that's awful. A different view. So Tom, what you have shown us this evening is the depth of thought. Addressing the question of there is no algorithm for truth. And I put a question mark on the end of that because I was questioning it. But you have shared with us this evening real insights as to A, how important this topic is. And B, just actually how complicated. And you're here at the Royal Institution, Tom, because actually you provide not only entertainment and inspiration. But frankly, a depth of thought that I personally really admire and I found extremely engaging. And I trust the audiences as well. And therefore, I would invite you all to thank Tom in the most usual way. Thank you. Thank you. Thank you very much, everybody.