WEBVTT 00:00:00.000 --> 00:00:09.680 Music. 00:00:09.527 --> 00:00:14.247 Hello and welcome to the edition where we cover the crossover of media and tech. 00:00:14.607 --> 00:00:18.887 And there's no better space for that to unpack than AI. 00:00:19.667 --> 00:00:23.307 I normally like to have authors on when they've got a new book out. 00:00:24.647 --> 00:00:29.507 Unfortunately, Chris Stoker-Walker keeps writing new books. So I have to keep having him on the show. 00:00:29.807 --> 00:00:35.127 Hello, Chris. How are you? Hello. I'm not too bad. Sorry, I do keep dominating your episode list. 00:00:35.647 --> 00:00:40.107 I mean, I'm more than happy for you to do so because you always have interesting things to discuss. 00:00:40.307 --> 00:00:42.667 And I don't have to talk about Elon Musk with you this time, 00:00:42.727 --> 00:00:45.487 which normally makes you cry on air. So we don't need to do that. 00:00:46.087 --> 00:00:49.527 But you do have a new book out. It's actually going to be out the day after 00:00:49.527 --> 00:00:52.527 this podcast comes out. It's called How AI Ate the World. 00:00:53.827 --> 00:00:59.527 It is. That's exciting. We'll dive into that in a minute because there's a huge 00:00:59.527 --> 00:01:03.867 amount going on in this space, which I guess is why you wanted to write the book. 00:01:05.247 --> 00:01:11.527 So let's start really with where do you see we are with AI as a general jumping off point? 00:01:11.687 --> 00:01:13.907 I know it's a broad question, but let's start broad and get narrow. 00:01:14.287 --> 00:01:18.447 No, it's a great one. And the book is itself simultaneously trying to be broad 00:01:18.447 --> 00:01:22.387 and also specific enough for people who have been following this space, 00:01:22.607 --> 00:01:26.307 people like the listeners, to actually be able to delve in and maybe learn something 00:01:26.307 --> 00:01:27.807 new or to think a little bit about it. 00:01:27.907 --> 00:01:32.207 And yeah, it's an interesting point in the kind of AI revolution, 00:01:32.387 --> 00:01:35.747 the generative AI revolution, at least in terms of we're around 00:01:35.747 --> 00:01:38.747 about 18 months since the release of chat gpt and 00:01:38.747 --> 00:01:44.507 i think a lot of the initial hype has gone away but the the kind of longer term 00:01:44.507 --> 00:01:49.327 ramifications and the i guess kind of we're in a to use a relationship metaphor 00:01:49.327 --> 00:01:54.707 kind of the the sort of you know fluttering heart sort of stomach turning exciting 00:01:54.707 --> 00:01:58.467 bit of the early stage of where everything is new and magnificent has 00:01:58.547 --> 00:02:02.247 kind of dissipated into this sort of question of, right, well, 00:02:02.407 --> 00:02:07.187 you know, we're together now, and I kind of think this is going to be a long-term thing. 00:02:07.347 --> 00:02:09.607 So how do we coexist peacefully? 00:02:09.807 --> 00:02:11.567 And I think that is the- Oh, you are such a romantic. 00:02:12.307 --> 00:02:15.667 Quite, exactly. I think this is the challenging question is like. 00:02:16.887 --> 00:02:22.107 Now that hype has gone, and also, you know, now that we're seeing kind of AI 00:02:22.107 --> 00:02:26.347 companies Companies may be struggling a little bit to convince us that this 00:02:26.347 --> 00:02:28.387 is the best thing since sliced bread. 00:02:28.567 --> 00:02:32.127 The scales have fallen a little bit from our eyes. I think that means that we 00:02:32.127 --> 00:02:36.087 can actually have some of the proper conversations that really we ought to have 00:02:36.087 --> 00:02:38.927 been having before it kind of got layered into everything. Yeah. 00:02:39.694 --> 00:02:43.894 Yeah, I mean, I was just before we spoke, I was just catching up on the Apple 00:02:43.894 --> 00:02:47.014 event yesterday and they tried to put, they didn't go too big on it, 00:02:47.074 --> 00:02:50.534 but they tried to put AI stuff in all the new iPads and keep talking about that. 00:02:51.014 --> 00:02:55.054 It seemed pretty clear to me a lot of the graphics that were generated by AI in that event. 00:02:56.014 --> 00:03:01.174 So it sort of is permeating things without us really having to acknowledge it. 00:03:01.234 --> 00:03:05.654 And obviously a lot of the discussion has been about regulation and the regulatory 00:03:05.654 --> 00:03:08.074 frameworks we're going to have to apply to this new technology. 00:03:08.074 --> 00:03:11.554 To me it's always felt a bit like people have 00:03:11.554 --> 00:03:14.794 gone oh we really messed up with web 2 and let social media run 00:03:14.794 --> 00:03:18.314 riot so we're going to have to really go big on regulation this 00:03:18.314 --> 00:03:23.294 time around is that your perception of things yeah it is i think i think politicians 00:03:23.294 --> 00:03:29.694 and the public both have been quite chastened by the last 20 25 years of particular 00:03:29.694 --> 00:03:37.414 social media you know just letting it run wild and buying the idea that these big tech platforms, 00:03:37.594 --> 00:03:42.674 companies and executives put forward that sunlight was the best disinfectant 00:03:42.674 --> 00:03:45.134 and that they knew best how to manage this. 00:03:45.394 --> 00:03:50.054 And actually, you know, if you had a light touch regime, things would work out in the end. 00:03:50.214 --> 00:03:53.314 And it turns out that maybe wasn't the case. 00:03:53.474 --> 00:03:57.114 And so it is interesting to see this kind of regulatory drive. 00:03:57.114 --> 00:04:03.054 And, yeah, I think that's the notable difference to me between what we saw with 00:04:03.054 --> 00:04:05.854 the kind of Web2 era of social media and AI now, 00:04:06.034 --> 00:04:10.574 where these politicians are way more advanced in saying, whoa, 00:04:10.714 --> 00:04:14.314 whoa, hang on, let's not let you get too far ahead, 00:04:14.514 --> 00:04:17.234 while actually still letting them get quite far ahead. 00:04:17.674 --> 00:04:21.194 And, yeah, notable, obviously. Are you suggesting that politicians dealing with 00:04:21.194 --> 00:04:22.334 tech might be performative? 00:04:22.654 --> 00:04:26.914 Never. Never in a million years, Charlotte. I know. No, but it is interesting 00:04:26.914 --> 00:04:31.634 because obviously we're talking two weeks ahead of the second AI Safety Summit, 00:04:31.734 --> 00:04:36.634 which nobody has heard about because the UK government isn't trying to desperately 00:04:36.634 --> 00:04:39.314 come up with a reason for its existence. 00:04:39.814 --> 00:04:42.674 Where they were with the Bletchley Park Summit when they were saying this is 00:04:42.674 --> 00:04:46.594 going to be the global leadership of AI. And so it's going to be held in South Korea. 00:04:47.474 --> 00:04:51.574 Unfortunately, a lot of the organizations, you know, I think OpenAI and a few 00:04:51.574 --> 00:04:54.634 actually countries have kind of said, you know what, we're ducking out of this 00:04:54.634 --> 00:04:57.634 one. We're not going to bother attending, going to wait for the next one in 00:04:57.634 --> 00:05:00.654 France, which is happening early next year. 00:05:00.834 --> 00:05:06.034 And so, yeah, we are in this interesting phase where I think we recognize that needs to be regulation. 00:05:06.054 --> 00:05:10.814 And as you kind of joked about, actually, some of this is a little bit of performance 00:05:10.814 --> 00:05:15.154 here where they're saying we need to do stuff and it's important and they're pressing on this. 00:05:15.214 --> 00:05:19.474 But actually, if they thought that it was that important, then they would be 00:05:19.474 --> 00:05:21.674 getting on a plane to Seoul and talking about it. 00:05:22.905 --> 00:05:26.105 Yeah, it'll be interesting to watch what actually comes out of that summit. 00:05:26.325 --> 00:05:30.245 It sort of feels to me like, you know, the environmental summits, 00:05:30.245 --> 00:05:34.825 you know, the COP summit, where some of them have real significance and some 00:05:34.825 --> 00:05:39.605 of them don't. And it feels a bit like this is on that kind of wavelength as well. 00:05:39.885 --> 00:05:44.105 Yeah, and I would actually say that even, you know, and I was being a bit snarky 00:05:44.105 --> 00:05:47.245 about Rishi Sunak and the government, and that's not like me at all. 00:05:47.345 --> 00:05:53.785 Not at all. Yeah, I would say the Bletchley Declaration that people covered 00:05:53.785 --> 00:05:59.125 pretty breathlessly was really just an agreement to talk about, 00:05:59.185 --> 00:06:00.205 an agreement in the future. 00:06:00.245 --> 00:06:04.065 It actually didn't really do much. And if they're not even going to go to this 00:06:04.065 --> 00:06:07.305 soul wand, then yeah, it does seem quite disinteresting. 00:06:07.305 --> 00:06:11.625 But there is a really kind of important element of this, which is actually, 00:06:11.685 --> 00:06:18.585 you know, a lot of the running is being made at the minute through Western global North democracies. 00:06:18.625 --> 00:06:22.045 And it is important that this is kind of reconfiguring, I suppose, 00:06:22.125 --> 00:06:25.285 the compass and recalibrating a bit. And so it's a shame, actually, 00:06:25.385 --> 00:06:27.565 given quite how much Japan is doing. 00:06:27.665 --> 00:06:31.605 And I had a long history, which I explore in the book, of kind of making some 00:06:31.605 --> 00:06:35.625 speed on AI back in the 80s and the 70s. 00:06:36.025 --> 00:06:38.405 And China obviously doing so now. 00:06:39.305 --> 00:06:43.425 And kind of South Korea also backing things quite significantly. 00:06:44.645 --> 00:06:49.305 Yeah, it is a shame that it seems like a lot of people won't be having seats around the table. 00:06:49.865 --> 00:06:52.625 It does seem like a missed opportunity. opportunity uh 00:06:52.625 --> 00:06:55.645 if nothing else sounds lovely to be able to get a trip to seoul if anyone 00:06:55.645 --> 00:06:58.545 wants to pay for me to report on it on this summit 00:06:58.545 --> 00:07:02.285 i'd love to have a trip to seoul um but you've raised a number of important 00:07:02.285 --> 00:07:06.305 issues which i know are explored in the book first of all i think people and 00:07:06.305 --> 00:07:10.345 you can tell me why i'm wrong i think people have got almost over excited about 00:07:10.345 --> 00:07:16.005 ai i think people have gotten the fundamental point that at this moment at least 00:07:16.005 --> 00:07:18.125 and you can tell me if this might change. 00:07:18.685 --> 00:07:25.845 It can only tell us things that already exist somewhere, stuff that someone already knows, right? 00:07:26.005 --> 00:07:31.145 So please write me, tell me what happened in XY football match. 00:07:31.465 --> 00:07:34.145 That information is out there somewhere on the internet. 00:07:34.925 --> 00:07:41.065 They're not yet the Borg AI. It's not all thinking, all living, 00:07:41.145 --> 00:07:42.705 breathing technology itself. 00:07:42.985 --> 00:07:48.245 And I think we've slightly lost track of that, particularly in journalism, which I want to come to. 00:07:49.385 --> 00:07:54.045 So I think that's happening. And we're almost, you know, we're forgetting that 00:07:54.045 --> 00:07:57.665 actually AI needs human input still. 00:07:58.770 --> 00:08:04.910 Yeah. And I think that that is, in some ways, actually the original sin of AI to call it AI. 00:08:05.090 --> 00:08:08.590 So back in the 50s, there was this kind of huge debate. 00:08:08.890 --> 00:08:13.670 Lots of people were putting effort and time into researching this thing and 00:08:13.670 --> 00:08:19.110 kind of deciding, well, we're all looking at this kind of big theory. 00:08:19.710 --> 00:08:22.550 What should we call it? And there were a bunch of different proposals, 00:08:22.790 --> 00:08:26.810 one of which was automata studies, which might have actually been a more fitting terminology. 00:08:27.670 --> 00:08:31.790 Turns out, you know, 1956, there was this big kind of, you know, 00:08:31.830 --> 00:08:35.710 conclave, and it came under the banner of artificial intelligence, 00:08:35.870 --> 00:08:39.370 which was great branding, because it meant that people got a lot of funding. 00:08:39.650 --> 00:08:43.410 But it was kind of an issue, because it has that word intelligence in it. 00:08:43.530 --> 00:08:48.270 And we're using those terminologies and those phrases ourselves, 00:08:48.550 --> 00:08:50.590 and actually maybe doing so incorrectly. 00:08:50.850 --> 00:08:54.430 I mean, you mentioned, Charlotte, that, you know, how AI 00:08:54.430 --> 00:08:57.670 quote-unquote thinks and I've done this myself and it's 00:08:57.670 --> 00:09:01.590 also an issue with journalism as you say which is we anthropomorphize 00:09:01.590 --> 00:09:08.650 this stuff in a way that encourages everyday users to imbue it with a belief 00:09:08.650 --> 00:09:12.910 that it can do more than it actually can so as you say pattern matching technology 00:09:12.910 --> 00:09:18.150 it's guessing the next most likely word if you say I want you to complete the sentence, 00:09:18.410 --> 00:09:23.330 the dog, it will basically just do a series of probability equations to say, 00:09:23.410 --> 00:09:27.130 you're probably going to want it to say the dog barked rather than the dog quoted 00:09:27.130 --> 00:09:28.790 Shakespeare or something like that. 00:09:29.090 --> 00:09:33.430 And yeah, it's not helped also by the big tech companies themselves. 00:09:33.630 --> 00:09:38.830 I mean, OpenAI has got shareholders and it's got multi-billion dollar valuations to upkeep. 00:09:38.930 --> 00:09:42.350 So they go, yeah, no, this is the magic thing that's going to transform us. 00:09:42.350 --> 00:09:45.390 But yeah you're right um we're not at 00:09:45.390 --> 00:09:48.390 that stage and arguably we may never actually be 00:09:48.390 --> 00:09:51.490 at that stage yeah i mean if i said please write a 00:09:51.490 --> 00:09:54.530 introduction to this podcast for chris stoker walker 00:09:54.530 --> 00:09:57.330 being a guest it would do that but it would draw on 00:09:57.330 --> 00:10:00.990 everything that already exists about you on the internet and it 00:10:00.990 --> 00:10:04.610 would be wrong and it might well be wrong absolutely um 00:10:04.610 --> 00:10:07.370 the other thing which you touched on in your talk about the 00:10:07.370 --> 00:10:10.230 summit and everything is and it comes 00:10:10.230 --> 00:10:14.530 from my point about it being based on information that already exists is 00:10:14.530 --> 00:10:18.470 that it really matters who is creating this information who is creating the 00:10:18.470 --> 00:10:23.850 technology and if it's a bunch of straight white blokes in silicon valley we 00:10:23.850 --> 00:10:29.030 will have inherent biases built into this technology which whatever we think 00:10:29.030 --> 00:10:32.110 of it is going to really have profound effects on a number of industries, 00:10:32.997 --> 00:10:38.177 It is. And the problem is, we have this issue with social media and with, 00:10:38.297 --> 00:10:42.617 frankly, every bit of big tech that we've had from smartphones that we use to 00:10:42.617 --> 00:10:46.377 Netflix and YouTube that we kind of consume content through. 00:10:47.397 --> 00:10:50.217 Quick plug, Chris has also written books about that. Yes, I have. 00:10:50.297 --> 00:10:54.257 And TikTok. But TikTok is slightly different because TikTok is technically kind 00:10:54.257 --> 00:10:56.177 of... Chris, when it gets banned, you're coming back on the show. 00:10:56.857 --> 00:11:00.917 Exactly. Good, good, good. I'm glad. Gosh, I've got that summer to look forward 00:11:00.917 --> 00:11:02.697 to of listening to court cases endlessly. 00:11:04.097 --> 00:11:10.897 But the thing is that we know that this is a problem and we've known that this 00:11:10.897 --> 00:11:11.737 is a problem for a long time. 00:11:11.857 --> 00:11:16.817 In fact, we've known this is a problem in a pre-tech era. The impact of US TV 00:11:16.817 --> 00:11:21.817 on the way that we talk and the way that we act and the things that we think about is enormous. 00:11:22.177 --> 00:11:28.757 But the problem with AI is it is just so ubiquitous and it's kind of touching 00:11:28.757 --> 00:11:33.837 spaces that weren't touched with social media and that weren't touched with tech. 00:11:33.837 --> 00:11:38.717 So the idea of if you layer in AI into job hiring decisions, 00:11:38.877 --> 00:11:44.517 filtering through CVs, that has a very acutely American phrasing to it. 00:11:44.597 --> 00:11:49.637 If you have employment policies being developed by law LLMs, 00:11:49.717 --> 00:11:53.657 which is happening in law firms, like I've spoken to them, they're using this 00:11:53.657 --> 00:11:55.837 to kind of generate stuff. Contracts being written. Yeah, precisely. 00:11:56.177 --> 00:12:01.057 And it will be imbued with that kind of, not just an American sensibility, 00:12:01.117 --> 00:12:06.277 but also a kind of really curiously Silicon Valley libertarian sensibility, 00:12:06.397 --> 00:12:08.117 which some people will find great. 00:12:08.197 --> 00:12:13.337 And not to diminish or demean anybody's politics, but I don't think that's necessarily 00:12:13.337 --> 00:12:17.517 the kind of settled belief around the world. 00:12:17.517 --> 00:12:21.157 And just one really brief thing which I find is interesting and kind of like 00:12:21.157 --> 00:12:25.477 indicative of this story writ large and it helps kind of. 00:12:26.002 --> 00:12:28.762 Telling it through the numbers yeah we've known that the web 00:12:28.762 --> 00:12:32.122 has always been like a predominantly english language place 00:12:32.122 --> 00:12:34.982 um and that's where the majority of the training data 00:12:34.982 --> 00:12:38.802 for these generative ai tools comes from common rule which kind of yeah sorry 00:12:38.802 --> 00:12:42.822 to interrupt but when i'm saying that it's based ai is based on things that 00:12:42.822 --> 00:12:46.462 already exist training data is the phrase i really should be using yeah precisely 00:12:46.462 --> 00:12:50.682 so the training data that is you know on the web basically you know ai just 00:12:50.682 --> 00:12:54.142 kind of hoovers up all the content on the web which has its own issues because then you get 00:12:54.202 --> 00:12:57.662 like the bots you know elon's porn bots and you get kind of 00:12:57.662 --> 00:13:01.002 just the worst of reddit and all that stuff um and 00:13:01.002 --> 00:13:04.882 that suddenly becomes quote-unquote knowledge that this is based on but you 00:13:04.882 --> 00:13:10.042 know 50 or more of the web is english language actually english is spoken by 00:13:10.042 --> 00:13:15.322 like 15 or so of the world's population you look at the other end of the scale 00:13:15.322 --> 00:13:20.482 like um you know mandarin is spoken by like 16% of the world's population, 00:13:20.682 --> 00:13:23.182 but it's like 1% of the internet. 00:13:23.342 --> 00:13:26.022 And you can do the same sort of equation with Hindi, you can do it with French, 00:13:26.082 --> 00:13:30.942 and there's this disparity where English is way overrepresented on the web compared 00:13:30.942 --> 00:13:35.422 to its actual spoken use in real life. 00:13:35.802 --> 00:13:38.262 And then the opposite is the case for pretty much every other language, 00:13:38.342 --> 00:13:41.482 which means a really good example actually is Catalan. 00:13:42.202 --> 00:13:53.902 You know, the Catalan government controls the GTLD of .cat, which is... Explain that quickly. 00:13:54.142 --> 00:13:56.702 So basically, you have a website. 00:13:57.342 --> 00:13:59.682 It's a top-level domain, right? Yeah, top-level domain name. 00:13:59.722 --> 00:14:03.802 So basically a .com or a .net. There's a .cat. It's been around for ages. 00:14:04.502 --> 00:14:07.342 Lots of these are controlled by governments. the Catalan government requires 00:14:07.342 --> 00:14:09.442 anybody who registers a .cat domain 00:14:09.442 --> 00:14:13.302 name to have Catalan language on their website in some shape or form. 00:14:13.382 --> 00:14:17.422 So if you want to make a website about cats, people often use these top-level 00:14:17.422 --> 00:14:19.662 domains for different things, right? 00:14:20.457 --> 00:14:23.557 You would actually have to have some Catalan somewhere. Yeah. 00:14:23.597 --> 00:14:28.337 So you say you want to register ilove.cat and you have to... 00:14:28.337 --> 00:14:29.517 Just going to buy that now, continue. 00:14:29.837 --> 00:14:33.517 Yeah, exactly. It might be like thousands because domain name trading is amazing. 00:14:33.617 --> 00:14:35.097 It's a sort of subsection of society. 00:14:36.157 --> 00:14:39.357 You have to put some sort of Catalan on it. And you don't know Catalan probably. 00:14:39.537 --> 00:14:44.397 So what you do is in the 1990s or the 2000s, you go to like the best translation 00:14:44.397 --> 00:14:46.337 tool that you can find and it's rubbish. 00:14:46.337 --> 00:14:49.637 Bush and basically that then gets used as 00:14:49.637 --> 00:14:52.737 training data for official catalan language on 00:14:52.737 --> 00:14:55.517 chat gpt which is terrible and people are really 00:14:55.517 --> 00:14:58.757 worried about there's some great academic research by the center 00:14:58.757 --> 00:15:01.897 for democracy and technology that actually looks specifically at these issues that 00:15:01.897 --> 00:15:04.957 it's well worth reading it's fascinating 00:15:04.957 --> 00:15:07.637 there's so much to unpack i want to dive in and it 00:15:07.637 --> 00:15:10.977 links to your point about training data the issues 00:15:10.977 --> 00:15:15.217 around journalism is first 00:15:15.217 --> 00:15:18.237 of all let's start with the big question is ai going to take all our jobs in journalism 00:15:18.237 --> 00:15:20.857 no because they can't speak to people that's the 00:15:20.857 --> 00:15:23.877 bigger thing so i really briefly i teach journalism at 00:15:23.877 --> 00:15:26.937 a university as well i have encountered students who have 00:15:26.937 --> 00:15:32.797 tried to use chat gpt um for producing work it was immediately yeah well no 00:15:32.797 --> 00:15:37.037 didn't work it was because it immediately stood out because they basically did 00:15:37.037 --> 00:15:41.077 a story on a very small geographical area of newcastle which is where i know 00:15:41.077 --> 00:15:45.837 and they made up a school a school headmaster, 00:15:46.117 --> 00:15:48.357 a university, a university principal, 00:15:48.757 --> 00:15:52.017 all of which did not exist and all of which had fake names. 00:15:52.197 --> 00:15:55.897 And yeah, basically it was just hallucinating. So you can immediately tell for 00:15:55.897 --> 00:15:59.037 that instance, they can't do the gumshoe reporting that we all know. 00:15:59.477 --> 00:16:04.077 It can't do that. And it also can't do, and I've said this so many times on 00:16:04.077 --> 00:16:05.797 the show, the kind of original reporting. 00:16:05.917 --> 00:16:10.757 It can't quote in brackets yet generate an idea. 00:16:10.897 --> 00:16:14.417 It can't speak speak to a source and that's really 00:16:14.417 --> 00:16:18.257 interesting and also it's not very good i think having a 00:16:18.257 --> 00:16:21.597 sense of what is newsworthy so i mean i've tried using this 00:16:21.597 --> 00:16:27.257 because it frankly would be a massive time hack um and you know if you feed 00:16:27.257 --> 00:16:31.837 an interview transcript into it and you say as you have to for prompts you are 00:16:31.837 --> 00:16:35.217 a world leading journalist you should read this interview you should digest 00:16:35.217 --> 00:16:38.777 the contents of this interview and then you should try and pick out the key 00:16:38.777 --> 00:16:40.717 quotes from this interview for you as if you were a journalist, 00:16:41.337 --> 00:16:46.237 it'll miss what I would class as the key, most newsworthy bits. 00:16:46.337 --> 00:16:48.897 Now, that's really actually an interesting defense, I think, 00:16:48.917 --> 00:16:52.917 of humanity, and most importantly, of the idea that journalism is an art, not a science. 00:16:53.077 --> 00:16:57.477 Because actually, maybe it would be newsworthier for you, or it might be newsworthy 00:16:57.477 --> 00:17:00.157 for someone else, because we would all approach an interview transcript, 00:17:00.337 --> 00:17:03.697 and we might have some of the same ideas of what's newsworthy from it, 00:17:03.757 --> 00:17:07.337 but maybe not all of them. So yeah, you can't really do that very well either. 00:17:08.737 --> 00:17:13.417 But surely there are going to be some newsroom tasks that it is going to replace. 00:17:14.037 --> 00:17:17.677 Yeah. And one of the things that I think is really interesting is Chris Moran 00:17:17.677 --> 00:17:22.757 at The Guardian has kind of highlighted a lot of this and done a lot of the work on this. 00:17:22.817 --> 00:17:25.797 I was at the International Journalism Festival in Perugia a few weeks ago. 00:17:26.397 --> 00:17:28.517 That's not a humble brag. That's the first time I've been there, 00:17:28.597 --> 00:17:29.617 but it was it was really fun. 00:17:29.897 --> 00:17:32.837 I don't get invited to these things, Chris. It's fine. You can brag. 00:17:33.017 --> 00:17:36.577 You can do a panel. We should do a panel on podcasting. and 00:17:36.577 --> 00:17:39.377 actually right see you there we'll do a 00:17:39.377 --> 00:17:42.177 panel we'll eat past it'll be delightful yeah it'd be brilliant it was 00:17:42.177 --> 00:17:45.117 tipping down with brain so actually it wasn't that great but the the talks 00:17:45.117 --> 00:17:48.017 were really interesting uh and you know he was talking about how they're 00:17:48.017 --> 00:17:50.457 trying to use it in some ways and there were other people who were using it 00:17:50.457 --> 00:17:55.237 to workshop headlines and things like that and a lot of the the automation will 00:17:55.237 --> 00:17:59.377 work i think where it doesn't work is where we're seeing it in kind of um a 00:17:59.377 --> 00:18:03.657 lot of local journalism of the idea of ai assisted reporters of you know They're 00:18:03.657 --> 00:18:05.277 basically just getting it to create something, 00:18:05.397 --> 00:18:08.557 having minimal oversight by a human and seeing if it's correct. 00:18:09.697 --> 00:18:16.437 But I do think, for example, things like what level did the stock market close at? 00:18:16.737 --> 00:18:20.417 It could write a small paragraph coherently on that. 00:18:21.117 --> 00:18:24.597 What were the football scores yesterday? You could easily do a small paragraph 00:18:24.597 --> 00:18:26.797 on that, which will save a reporter's time. 00:18:26.857 --> 00:18:33.237 I've said all this before. And I also think that I'm going to get on my soapbox 00:18:33.237 --> 00:18:35.597 actually a bit, if you don't mind, 00:18:35.657 --> 00:18:41.657 because I think we sort of have almost succumbed that AI is going to dominate 00:18:41.657 --> 00:18:44.797 journalism, it's going to take us over. 00:18:45.237 --> 00:18:48.257 And I think we don't have to accept that. 00:18:48.517 --> 00:18:51.557 I think it's fantastic that we can use it for a tool. 00:18:51.757 --> 00:18:57.157 You can use all sorts of, you've named some, testing headlines, 00:18:58.137 --> 00:19:01.017 scraping through data, we can use it for really fantastic things, 00:19:01.217 --> 00:19:06.677 but equally we don't have to accept it's going to write our jobs and actually 00:19:06.677 --> 00:19:09.597 human interest, local journalism, 00:19:10.197 --> 00:19:14.517 actually might really appeal to people in a world where AI is writing a load of rubbish. 00:19:16.659 --> 00:19:20.639 I mean, I've always thought this anyway. And so we've got that, 00:19:20.679 --> 00:19:21.839 and I'm going to get on my soapbox a bit. 00:19:21.959 --> 00:19:26.659 And, of course, the other thing that's a huge issue in AI, and I have no doubt, 00:19:26.719 --> 00:19:28.899 I haven't had a chance to read the book yet, but I have no doubt you explore 00:19:28.899 --> 00:19:33.539 it, is the issues of copyright, where this training data has come from. 00:19:33.759 --> 00:19:38.599 Now, your books, presumably, and my book, and the articles we've written, 00:19:38.799 --> 00:19:43.379 are being fed somewhere into an LLM, are being used as this training data. 00:19:43.839 --> 00:19:47.879 We haven't consented to that. I doubt the publishers we've worked for have consented 00:19:47.879 --> 00:19:49.399 to that. It's a huge issue. 00:19:49.919 --> 00:19:57.819 It is, and it's one that I think OpenAI has started literally this week to try and head off. 00:19:57.819 --> 00:20:02.539 It kind of announced that in 2025, which is, A, a really long time away, 00:20:02.699 --> 00:20:08.299 you'll be able to ask for your data to not be included as trailing data. 00:20:08.499 --> 00:20:12.479 Which is also too late, and bolting the doors after the horse is gone because 00:20:12.479 --> 00:20:19.459 ultimately it's already been subsumed into this and the knowledge that is imparted 00:20:19.459 --> 00:20:21.819 from that is already in those models. 00:20:21.939 --> 00:20:26.359 So yeah, this is a huge problem. It's one actually that I speak to Ed Newton-Rex 00:20:26.359 --> 00:20:31.579 who used to work at Stability AI, the company behind Stable Diffusion about in the book. 00:20:31.619 --> 00:20:37.219 He left the company back in November over 00:20:37.219 --> 00:20:40.739 a kind of difference of opinion because these AI companies are using 00:20:40.739 --> 00:20:43.579 essentially fair use doctrine as their defense so the idea 00:20:43.579 --> 00:20:47.259 in copyright law that you know you can use 00:20:47.259 --> 00:20:50.559 it for kind of transformative purposes source data and you 00:20:50.559 --> 00:20:54.959 can get away with it provided that you know you're not relying or 00:20:54.959 --> 00:20:59.119 recreating wholesale this content and 00:20:59.119 --> 00:21:02.279 profiting off it and that was 00:21:02.279 --> 00:21:05.479 something that state stability AI Newton Rex 00:21:05.479 --> 00:21:08.699 claimed was using as its defense in a US Copyright 00:21:08.699 --> 00:21:11.479 office um consultation about what to 00:21:11.479 --> 00:21:14.339 do about generative ai and he basically said i didn't feel comfortable with 00:21:14.339 --> 00:21:18.019 that he's actually become a bit of a crusader in the months since and when i 00:21:18.019 --> 00:21:21.319 spoke to him he was a little bit demure about talking out about this and he's 00:21:21.319 --> 00:21:25.899 become quite radical now um on it which is good to see but yeah you're right 00:21:25.899 --> 00:21:29.359 and and you know i think that they know these ai companies that something is 00:21:29.359 --> 00:21:30.959 up and that they need to kind of sort it because. 00:21:31.499 --> 00:21:34.119 Otherwise we wouldn't be seeing them making all these deals with news publishers 00:21:34.119 --> 00:21:37.979 to try and get good quality training data, because that's what they need, right? 00:21:38.139 --> 00:21:43.979 We have estimates that useful training data for generative AI tools will run 00:21:43.979 --> 00:21:45.639 out in the next year or two. 00:21:45.999 --> 00:21:48.399 One estimate was 2026. I've seen another one for 2025. 00:21:49.779 --> 00:21:57.199 These are hungry beasts. And so trying to feed them new high-quality data is quite challenging. 00:21:57.279 --> 00:22:01.279 And yeah, that's, I think, why journalism is so valuable to them. 00:22:02.625 --> 00:22:08.625 I also worry, this is a slight about turn, I also worry the effects of various 00:22:08.625 --> 00:22:10.545 industries, but in particular journalism. 00:22:10.825 --> 00:22:15.945 Now, you and I have both done the back in the day cub reporter journalism of 00:22:15.945 --> 00:22:21.685 can you not flip a press release, but can you just write this bit of information? 00:22:21.905 --> 00:22:24.845 Again, it could be the football scores, it could be the stock market close, 00:22:24.945 --> 00:22:26.805 it could be, you know, a very basic thing. 00:22:26.805 --> 00:22:30.025 Can you write this in an inverse pyramid style i need 00:22:30.025 --> 00:22:33.465 350 words in half an hour and i 00:22:33.465 --> 00:22:37.425 think publishers are going to be really tempted to give 00:22:37.425 --> 00:22:41.905 that to ai because it's simple and it can probably or we'll get to a point where 00:22:41.905 --> 00:22:45.965 it can get that pretty right but then you are not going to be training younger 00:22:45.965 --> 00:22:51.405 journalists you know the kids you lecture what are they going to do for their 00:22:51.405 --> 00:22:56.965 first job if that bit of first job is taken by chat GPT? 00:22:57.285 --> 00:23:02.105 And that is the fundamental question I think we're starting to wrestle with. 00:23:02.285 --> 00:23:07.905 And some people have been kind of beating that drum very loudly since November 00:23:07.905 --> 00:23:11.985 2022, seeing this as kind of a hugely transformative technology. 00:23:12.105 --> 00:23:17.485 It's something that we talk about a lot in the book. of the IMF came out in 00:23:17.485 --> 00:23:23.125 January this year and said 40% of jobs will be touched in some way by generative AI. 00:23:23.305 --> 00:23:26.425 And they didn't go into detail what that meant. It didn't mean necessarily that'll 00:23:26.425 --> 00:23:30.805 be wholly replaced, but it means that could be changed. And there are these kind of. 00:23:31.756 --> 00:23:35.936 Longer term issues that I think the tech world in particular, 00:23:36.036 --> 00:23:39.856 which isn't very well focused on thinking about the long-term ramifications 00:23:39.856 --> 00:23:43.196 of these things, it's always disruptive and what is the new thing? 00:23:43.496 --> 00:23:48.636 Disruptive in their favorite terminology has a second meaning in pre-tech English, 00:23:48.756 --> 00:23:50.636 which means not always a good thing. 00:23:50.936 --> 00:23:54.436 Or yeah, what happens if you lose that ability? 00:23:55.656 --> 00:24:00.196 And I guess it means that core training for journalists might change. 00:24:00.256 --> 00:24:05.176 It means that maybe Maybe interview skills become even better taught in terms 00:24:05.176 --> 00:24:08.776 of actually, if we know that it can automate some of the reporting or the writing 00:24:08.776 --> 00:24:14.056 process, how do we get the best quality training data of sorts for that individual story? 00:24:14.336 --> 00:24:19.576 Well, by getting better interviews, by getting more killer quotes from people. 00:24:19.736 --> 00:24:24.576 So maybe that's the way to do it. And we're in a really difficult phase where 00:24:24.576 --> 00:24:28.816 we don't yet know quite how much impact this is going to have. 00:24:28.816 --> 00:24:34.116 And we have to try and train people to have two simultaneous thoughts in their mind. 00:24:34.216 --> 00:24:38.136 One of the world where their job is completely transformed, and one of the world 00:24:38.136 --> 00:24:41.576 where their job is transformed, but maybe not as much as they fear right now. 00:24:41.876 --> 00:24:45.576 And that it can be useful. I really don't want to be a Luddite about this. 00:24:45.676 --> 00:24:48.716 I don't think this is the apocalypse and it's going to have no value. 00:24:48.976 --> 00:24:54.216 I think we have to learn the lessons where we handed over everything to social platforms. 00:24:54.436 --> 00:24:59.876 I think we have to think about, like, you know, we have to think about, 00:24:59.996 --> 00:25:05.016 okay, so there's senior journalists who may be able to use AI effectively. 00:25:05.276 --> 00:25:08.236 They've done the work, they've been trained over the years, they know what to 00:25:08.236 --> 00:25:12.336 do. They can use the tools effectively, they can use basic journalism skills effectively. 00:25:12.556 --> 00:25:17.476 But you have to train the next lot of journalists coming up who are not at that level yet. 00:25:18.196 --> 00:25:23.016 And this applies to all sorts of industries. You were We're talking about legal firms using it. 00:25:23.156 --> 00:25:26.596 You know, lawyers have to learn the basics first. 00:25:26.796 --> 00:25:30.016 You know, you have to work your way up. Computer programmers in tech, 00:25:30.176 --> 00:25:34.916 there's lots of talk about how AI is going to help program quicker and write code quicker. 00:25:35.096 --> 00:25:39.136 But you've still got to learn the basics and do those basic things first and 00:25:39.136 --> 00:25:40.456 make the basic mistakes first. 00:25:41.356 --> 00:25:48.776 So there is that. And I really do agree with you that AI can't create human creativity. 00:25:48.896 --> 00:25:50.696 And I'm skeptical it ever will be able to. 00:25:51.476 --> 00:25:54.496 Yeah i don't think it will i think i think the thing 00:25:54.496 --> 00:25:57.876 that it's good at is mimicking is 00:25:57.876 --> 00:26:00.576 kind of copying and so it can and it can 00:26:00.576 --> 00:26:06.856 maybe create kind of a facsimile of what it thinks again thinks it doesn't think 00:26:06.856 --> 00:26:11.216 but what yeah what you're and what but you're you're talking about it like you're 00:26:11.216 --> 00:26:16.296 anothermorphotizing about now aren't you yeah exactly that's that's the challenge 00:26:16.296 --> 00:26:18.776 is you know It's so easy to kind of slip into this. 00:26:18.816 --> 00:26:21.736 I mean, I called it a hungry beast as well. 00:26:22.656 --> 00:26:27.076 That's all anthropomorphizing it. But like, yeah, it can kind of. 00:26:28.228 --> 00:26:33.188 It can get what it thinks is the kind of base level knowledge of this. 00:26:33.248 --> 00:26:37.068 And it touches on loads of different industries. 00:26:37.468 --> 00:26:39.908 You mentioned kind of the impact that it'll have on some of them. 00:26:39.968 --> 00:26:44.868 One thing that I find quite scary is some academic researchers think that it 00:26:44.868 --> 00:26:47.408 can be used to kind of replace human participants in studies. 00:26:47.648 --> 00:26:52.188 And the argument they make is actually it's better for cases where maybe you're 00:26:52.188 --> 00:26:56.308 asking people to relive horrible moments in their lives, which is good. 00:26:56.308 --> 00:27:02.188 But then also the problem is the things that make us human are the kind of foibles 00:27:02.188 --> 00:27:08.528 and the quirks that separate me from you and separate both of us from listeners and makes us unique. 00:27:08.528 --> 00:27:13.928 And what kind of generative AI does is it acts as a steamroller, 00:27:13.988 --> 00:27:17.928 kind of smushes us all down into a single layer, homogenous mass, 00:27:18.248 --> 00:27:25.088 builds us back up into kind of a sort of claymation version of us as humans 00:27:25.088 --> 00:27:26.868 and says, look, here is a human. 00:27:26.968 --> 00:27:30.888 I made it with this person's eyebrows and this person's head and this person's 00:27:30.888 --> 00:27:32.288 heart and this person's soul. 00:27:33.288 --> 00:27:36.508 And it's just kind of a law of averages of all of us. And actually, 00:27:36.608 --> 00:27:40.648 that's not what makes us interesting. It's not what's kept us around for so long. 00:27:41.428 --> 00:27:46.668 I'm also deeply skeptical when I hear people, very senior in the tech industry, 00:27:46.948 --> 00:27:49.108 talking about the new jobs AI will create. 00:27:49.168 --> 00:27:51.648 What's the phrase? They use sort of prompt engineer. 00:27:52.328 --> 00:27:58.048 And those kind of rather, I think, Orwellian phrases that basically someone 00:27:58.048 --> 00:28:00.608 that punches stuff into chat GPT all day. 00:28:01.528 --> 00:28:04.868 I am unconvinced. Look, there'll be some people that are good at that, 00:28:04.868 --> 00:28:11.448 But I'm unconvinced that will be a replacement for jobs in various industries. 00:28:12.008 --> 00:28:17.368 No, it won't. I mean, you're right. It's like horse whisperers or people who 00:28:17.368 --> 00:28:18.508 think they can talk to animals. 00:28:18.628 --> 00:28:22.648 They have the ability to do this thing, but it's ultimately going to be niche 00:28:22.648 --> 00:28:26.568 and it's not going to be as ubiquitous as other jobs. 00:28:26.868 --> 00:28:31.848 And the really interesting thing to me is that this kind of job displacement, 00:28:31.848 --> 00:28:37.008 job replacement thing is more likely to affect white collar workers and blue 00:28:37.008 --> 00:28:40.828 collar workers it's kind of actually the the flip side of a lot of the issues 00:28:40.828 --> 00:28:41.628 that we've worried about, 00:28:42.248 --> 00:28:45.188 in the last few decades um because this 00:28:45.188 --> 00:28:49.428 is a and let's be blunt that's kind of why people are freaking out because a 00:28:49.428 --> 00:28:52.408 lot of the people have never been affected by tech disruption or have benefited 00:28:52.408 --> 00:28:56.848 from it are now finding a problem yeah precisely and i think that's why people 00:28:56.848 --> 00:29:02.608 are being heard more and why we're getting maybe more action starting to be thought about is, 00:29:02.708 --> 00:29:05.348 yeah, because if you can, 00:29:05.508 --> 00:29:07.828 it's also the politicians, right? 00:29:07.948 --> 00:29:12.828 One of the smart use cases for an LLM, a large language model of the type that 00:29:12.828 --> 00:29:15.128 ChatGPT has an element of, 00:29:15.208 --> 00:29:19.648 it's got so much more now than just an LLM, but one smart use case of it actually 00:29:19.648 --> 00:29:25.388 would be, in the best case, to improve government continuity of you feed... 00:29:26.328 --> 00:29:31.248 You know, policy ideas, proposals over the last century or whatever into an 00:29:31.248 --> 00:29:34.468 LLM and the government can call upon that and say, you know, 00:29:34.468 --> 00:29:39.088 what have we already thought about for ideas to try and help improve things? 00:29:39.448 --> 00:29:44.108 What worked and what didn't? And can you suggest something better that might 00:29:44.108 --> 00:29:45.408 actually have a chance of succeeding? 00:29:46.168 --> 00:29:50.768 The issue with that is obviously, it could potentially gap and this is years 00:29:50.768 --> 00:29:53.468 and years and decades and decades down the line. 00:29:53.528 --> 00:29:56.648 And if everything goes right and if it's a fair worth a day and if there's a 00:29:56.648 --> 00:30:01.988 why in the day of the month or whatever like it could become a replacement and 00:30:01.988 --> 00:30:03.728 that's why people are concerned because as you say, 00:30:04.368 --> 00:30:09.628 it's the people in power who now have that kind of looming over them and are really worried about it. 00:30:10.768 --> 00:30:14.968 Uh it's all going to be fascinating i am no 00:30:14.968 --> 00:30:18.188 doubt going to have to talk to you again about all of this the book uh hey 00:30:18.188 --> 00:30:21.608 how ai ate the world will probably be out by 00:30:21.608 --> 00:30:24.428 the time you listen to this show uh chris after a 00:30:24.428 --> 00:30:27.148 bunch of other books out youtube is tiktok boom and the history of 00:30:27.148 --> 00:30:29.908 the internet in bite-sized chunks which you can get he writes for 00:30:29.908 --> 00:30:32.568 basically everyone so he's making the rest 00:30:32.568 --> 00:30:35.568 of us look bad um thanks so much for 00:30:35.568 --> 00:30:38.468 for being here chris is there anything else you want to plug or 00:30:38.468 --> 00:30:41.328 where can people keep up with you yeah they can follow me 00:30:41.328 --> 00:30:44.088 still on x even though and we said we wouldn't talk about elon 00:30:44.088 --> 00:30:47.048 musk even though elon musk is still ruining the platform once 00:30:47.048 --> 00:30:49.968 known known as twitter that's my main hangout although 00:30:49.968 --> 00:30:53.168 i am with book promo more on linkedin nowadays 00:30:53.168 --> 00:30:55.928 as well which is interesting i've seen you pop up on 00:30:55.928 --> 00:30:58.988 my uh tiktok fyp as well oh yeah 00:30:58.988 --> 00:31:02.248 i do that only when it's book promo because i have a face for radio and podcasts 00:31:02.248 --> 00:31:07.728 rather than a face for tiktok um i'm at charlotte a henry across social media 00:31:07.728 --> 00:31:13.268 obviously the most wonderful thing you could do for me is head over to newsletter.theedition.net 00:31:13.268 --> 00:31:17.728 and take out a pay subscription that's keeps the The whole show on the road. 00:31:18.348 --> 00:31:22.788 There's also blogs and stuff at theedition.net. Thank you so much once again, 00:31:22.808 --> 00:31:25.648 Chris, for joining the show. And I'll see you all next week. 00:31:26.160 --> 00:31:34.970 Music.