WEBVTT - How AI’s latest advancements are already reshaping creative professions 0:00:05.200 --> 0:00:05.600 Jiaoda. 0:00:05.680 --> 0:00:08.800 I'm Chelsea Daniels and this is the Front Page, a 0:00:08.880 --> 0:00:16.680 daily podcast presented by The New Zealand Herald. Artificial intelligence 0:00:16.680 --> 0:00:19.440 has been the hot topic of debate in the business 0:00:19.440 --> 0:00:23.400 world for the last few years, but increasingly it's an 0:00:23.440 --> 0:00:27.520 area that is encroaching on the creative industries. The latest 0:00:27.560 --> 0:00:31.440 open ai update is so advanced fans online have used 0:00:31.440 --> 0:00:35.080 it to eerily replicate the hand drawn art style of 0:00:35.200 --> 0:00:39.479 Japanese anime favorites Studio Ghibli. It's just the latest sign 0:00:39.560 --> 0:00:43.040 of AI coming for the arts, with recent headlines also 0:00:43.159 --> 0:00:48.239 highlighting concerns over entirely artificial models in ad campaigns and 0:00:48.560 --> 0:00:51.720 fake movie trailers that look close to the real thing. 0:00:52.280 --> 0:00:55.800 What protections are there in place for our creative sector 0:00:56.160 --> 0:00:59.120 or could they become one of the first industries to 0:00:59.280 --> 0:01:03.120 fall to our new AI overlords. Today on the Front Page, 0:01:03.240 --> 0:01:07.280 University of Sydney Business School Associate Professor Sandra Peter is 0:01:07.319 --> 0:01:09.800 with us to take us through the impact of. 0:01:09.720 --> 0:01:11.399 These emerging technologies. 0:01:16.800 --> 0:01:21.240 Sandra, you've recently written about open aiy's latest update. Can 0:01:21.280 --> 0:01:25.200 you explain what's happening here with these Studio Ghibli style images? 0:01:25.560 --> 0:01:29.640 So Open Eye's latest update to chat jipt was a 0:01:29.760 --> 0:01:35.520 significantly improved image generation capability. This means that it allowed 0:01:35.600 --> 0:01:40.280 users to create like really convincing images in the style 0:01:40.360 --> 0:01:45.040 off And everyone's been trying out the Ghibli style because 0:01:45.040 --> 0:01:49.200 it's been one of those very very clearly identifiable styles, 0:01:49.200 --> 0:01:51.600 and it was also used by some altmann to tell 0:01:51.640 --> 0:01:54.920 people about this thing, and it's been really enormously popular, 0:01:55.080 --> 0:01:58.960 so much so that basically their systems have crashed since 0:01:59.000 --> 0:02:01.480 they've released it. And I think it was about another 0:02:02.560 --> 0:02:06.280 ten million people that have joined the open aies efforts 0:02:06.560 --> 0:02:09.440 based on this new image generation model. 0:02:09.600 --> 0:02:11.119 And so how is this possible? 0:02:11.200 --> 0:02:14.800 Has openaiy just scraped all two dozen Ghibli movies in 0:02:14.919 --> 0:02:16.800 order to replicate this style. 0:02:17.639 --> 0:02:20.840 They haven't. They've scraped the Internet, and the internet has 0:02:20.880 --> 0:02:23.840 a lot of things on the internet, and we can 0:02:23.880 --> 0:02:27.440 talk about what goes into training some of these models, 0:02:27.600 --> 0:02:31.440 but the idea with Generative AI is that they've changed 0:02:31.440 --> 0:02:34.079 the way that they create these images. So they've kind 0:02:34.120 --> 0:02:38.720 of moved from the traditional what we call diffusion models, 0:02:38.760 --> 0:02:43.240 where we have models that gradually refine just noisy data 0:02:43.320 --> 0:02:48.440 into something called autoregressive algorithms. And no one wants to 0:02:48.480 --> 0:02:50.800 go into the details of that. All you need to 0:02:50.840 --> 0:02:55.120 know about is is that basically it treats images like language. 0:02:55.280 --> 0:02:59.680 So Chagipiti now predicts words in a sentence, but it 0:02:59.720 --> 0:03:03.720 can also predict visual elements in an image. This means 0:03:03.720 --> 0:03:07.160 that you can basically use all the things that jipt 0:03:07.360 --> 0:03:11.520 has learned about Ghibli style, for instance, the fact that 0:03:11.560 --> 0:03:15.959 it's got things like soft pastels, it's a Japanese type animation, 0:03:16.120 --> 0:03:18.440 and so on, and it can use those to more 0:03:18.480 --> 0:03:22.960 accurately create these images from precise prompts that people give it. 0:03:23.200 --> 0:03:26.519 What are the copyright implications here of this happening. 0:03:26.639 --> 0:03:29.160 Oh, there's a lot of copyright images. And can I 0:03:29.160 --> 0:03:31.960 actual say when it comes to generative AI, this comes 0:03:31.960 --> 0:03:36.000 on top of many other controversies that we've had. But 0:03:36.280 --> 0:03:40.960 the ability to work with styles, and when I say styles, 0:03:41.000 --> 0:03:44.040 I don't just mean the Ghibli style. Everything becomes a 0:03:44.080 --> 0:03:49.600 style for generative AI. That means things like bananas or cats, 0:03:49.680 --> 0:03:54.000 or bloody corporate emails, they all become styles. The ability 0:03:54.040 --> 0:03:57.000 to work with these styles it becomes the heart of 0:03:57.000 --> 0:04:01.920 this controversy because for many artists, like our distinctive approaches 0:04:01.960 --> 0:04:05.200 to how we create art is not a style. That 0:04:05.280 --> 0:04:09.280 can be applied in a specific prompt. However, the traditionally 0:04:09.760 --> 0:04:14.160 the copyright law doesn't protect styles, only very specific expressions 0:04:14.240 --> 0:04:17.440 because we don't want to stifle creative expression. Right if 0:04:17.480 --> 0:04:21.080 you could copyright things like impressionism, that would limit what 0:04:21.120 --> 0:04:24.799 people can do. But there's a very clear difference between 0:04:24.800 --> 0:04:27.760 a general style and then the highly distinctive style that 0:04:27.839 --> 0:04:32.000 a person might have. If you remember a while back, 0:04:32.040 --> 0:04:35.520 there was a guy called Greg Rutowski. This was a 0:04:35.520 --> 0:04:38.520 Polish artist, and people were using his style over and 0:04:38.560 --> 0:04:42.160 over again to generate images on unstable diffusion. Now, if 0:04:42.560 --> 0:04:45.839 you try to do them in the style of Greg, 0:04:45.920 --> 0:04:51.760 this threatening sposed his livelihood and his craft. So creators 0:04:51.800 --> 0:04:54.840 have taken legal action against this on a number of fronts. 0:04:59.400 --> 0:05:02.400 The newest trend is to create studio I say create 0:05:02.640 --> 0:05:06.080 is to imitate Studio ghibli images truly just coming for 0:05:06.160 --> 0:05:10.360 one of our most beloved examples of human manual creativity. 0:05:10.480 --> 0:05:12.920 Everyone loves studio Ghibili because it's beautiful and also because 0:05:12.920 --> 0:05:17.000 it's painstakingly created by humans who love art. And now 0:05:17.040 --> 0:05:20.000 to actually bt is trying to imitate that and creating 0:05:20.040 --> 0:05:23.360 images that are just worse. They're just soul less versions 0:05:23.400 --> 0:05:25.760 of Studio Ghibli. It's also worth noting that Miyazaki, the 0:05:25.800 --> 0:05:28.920 mastermind behind Studio Ghibli, he once said of AI, I 0:05:28.960 --> 0:05:31.960 would never wish to incorporate this technology into my work 0:05:32.000 --> 0:05:34.360 at all, and once when he was presented with an 0:05:34.360 --> 0:05:36.640 example of how AI could be applied to his work, 0:05:36.720 --> 0:05:39.559 he saw the footage and said quote, I strongly feel 0:05:39.600 --> 0:05:41.960 that this is an insult to life itself. 0:05:46.080 --> 0:05:49.679 It emerged a few months ago that allegedly Meta CEO 0:05:49.880 --> 0:05:54.200 Mark Zuckerberg approved of his company using pirated versions of 0:05:54.320 --> 0:05:59.520 copyright protected books illegally uploaded to sites like libjen to 0:05:59.680 --> 0:06:03.560 try its AI software. If true, it seems like it 0:06:03.640 --> 0:06:05.919 might be breaching all sorts of laws, I suppose, But 0:06:06.080 --> 0:06:08.560 is there actually anything that can be done about it. 0:06:09.320 --> 0:06:11.880 This is one of those huge, if true, kind of things, 0:06:11.880 --> 0:06:15.560 and also something that is still being debated and contested. 0:06:15.680 --> 0:06:18.680 There is a good likelihood that things like libjan were 0:06:18.800 --> 0:06:21.520 ingested to train these models. We don't know at what 0:06:21.600 --> 0:06:25.800 stages and at what scale. It's also true that when 0:06:25.839 --> 0:06:29.279 we ingest what is commonly known as the Internet, we 0:06:29.320 --> 0:06:32.159 will end up ingesting things that have copyright that we 0:06:32.200 --> 0:06:35.440 don't necessarily think of being ingested. I think an easier 0:06:35.440 --> 0:06:38.719 way to think about that might be the controversy around 0:06:38.920 --> 0:06:41.800 generative AI models being able to write really really good 0:06:41.800 --> 0:06:46.360 dialogue for movies because they've ingested OpenSubtitles dot org, which 0:06:46.360 --> 0:06:50.360 has pirated subtitles for all the movies in all the languages, 0:06:50.640 --> 0:06:53.279 and even though it wasn't trained on actual movies and 0:06:53.320 --> 0:06:57.880 on movie script ingesting the pirated subtitles means that it's 0:06:57.880 --> 0:07:01.560 now very good at doing this. Yes, things can be done. 0:07:02.880 --> 0:07:03.800 The law will. 0:07:03.640 --> 0:07:07.960 Evolve much as the technology has, except that technology moves 0:07:08.000 --> 0:07:10.320 a lot faster than the law. But there's a lot 0:07:10.360 --> 0:07:13.600 of work on the way on new legislation to try 0:07:13.640 --> 0:07:16.040 to balance the fact that we need an enormous amount 0:07:16.040 --> 0:07:20.560 of data to create these models. We're protecting things like artists' 0:07:20.680 --> 0:07:24.400 identities and their creative work. And it's obviously not just 0:07:25.280 --> 0:07:30.400 studio Ghibili right. There's same concerns in written text in music. 0:07:30.480 --> 0:07:33.840 I've had Billie Eilish and Pearl Jam voicing concerns about 0:07:33.880 --> 0:07:36.840 generative AI in music. So this is very widespread and 0:07:36.920 --> 0:07:39.880 something that is being debated around the world now. The 0:07:39.960 --> 0:07:44.120 question of how we'll be able to kind of balance 0:07:44.160 --> 0:07:48.040 the fact that we do need a lot of data 0:07:48.120 --> 0:07:50.240 and there is scarcity in data, the fact that this 0:07:50.400 --> 0:07:54.800 legislation is has there are varied approaches around the globe. 0:07:54.880 --> 0:07:57.680 There have been caused in the US to allow companies 0:07:58.280 --> 0:08:01.000 to use copyrighted information to this go away for the 0:08:01.040 --> 0:08:04.920 company's training generative model. That we have an AI race, 0:08:05.640 --> 0:08:08.680 this is all this, this will all have to be 0:08:08.720 --> 0:08:10.360 worked out in the next couple of years. 0:08:20.960 --> 0:08:25.960 We've also recently seen the YouTube channel screen Culture demonetized. 0:08:26.160 --> 0:08:29.120 You may have seen some of their AI generated trailers 0:08:29.160 --> 0:08:30.960 that they make look like the real thing. 0:08:31.280 --> 0:08:33.559 How advanced is AI getting. 0:08:33.720 --> 0:08:35.760 Are we going to be able to tell the reels 0:08:35.800 --> 0:08:38.360 from the fakes in a few years time, And what 0:08:38.400 --> 0:08:41.040 does that mean for film studios and the like. 0:08:41.200 --> 0:08:43.959 That's a that's a complex question. Let me let me, 0:08:44.080 --> 0:08:46.520 let me, let me take a step back and talk 0:08:46.559 --> 0:08:50.000 about the bigger picture. This is this is an arms race, 0:08:50.120 --> 0:08:53.240 and at the moment we're not really winning it. In 0:08:53.280 --> 0:08:57.240 that we are able to create very very good fakes, 0:08:57.320 --> 0:09:02.320 deep fakes, and any type of content that is indistinguishable 0:09:03.080 --> 0:09:05.319 from the real thing. We were already able to do 0:09:05.360 --> 0:09:08.559 this really quite well with images a few years ago. 0:09:08.920 --> 0:09:11.640 We are now increasingly good at doing it with a 0:09:11.960 --> 0:09:16.240 voice with audio, we are increasingly good at doing it 0:09:16.000 --> 0:09:19.280 with video. I think the thing to focus on here 0:09:19.559 --> 0:09:22.600 is first that most of the generators, that all of 0:09:22.600 --> 0:09:25.800 the generators that are out there online cannot reliably tell 0:09:25.840 --> 0:09:28.199 you if a work is AI generated or not. If 0:09:28.240 --> 0:09:31.760 you remember the Pope in a white Valenciaga jacket controversy, 0:09:32.280 --> 0:09:34.880 it's been an arms race since, and that most of 0:09:34.960 --> 0:09:38.599 us can create convincing content using over the counter, commercially 0:09:38.600 --> 0:09:41.440 available software. So if you were trying to create a 0:09:41.679 --> 0:09:44.200 deep fake of my voice, it would likely cost you 0:09:44.240 --> 0:09:47.160 two dollars a month with commercially available software, and it 0:09:47.200 --> 0:09:49.839 would be a fantastic clone of my voice that can 0:09:49.840 --> 0:09:52.600 fool my mom. If you were looking to create video, 0:09:52.679 --> 0:09:54.800 you'd probably need a minute and a half of video 0:09:54.840 --> 0:09:57.040 of me in the public domain to create deep fakes 0:09:57.040 --> 0:10:00.040 of me. So it is an arms race, and you 0:10:00.160 --> 0:10:02.320 know we're not good at telling these things apart. 0:10:02.679 --> 0:10:04.280 And it's not just movies. 0:10:04.679 --> 0:10:08.560 Recently I saw global fashion giant H and M announced 0:10:08.600 --> 0:10:12.520 plans to use AI by making digital twins of thirty 0:10:12.559 --> 0:10:14.880 of its models. Now the company has said that the 0:10:14.920 --> 0:10:18.439 models would own the rights to their likeness. But this 0:10:18.480 --> 0:10:22.200 is something that's been worrying the fashion industry insiders for 0:10:22.280 --> 0:10:25.360 some time. Does there need to be better regulations in 0:10:25.440 --> 0:10:29.040 place before companies start to innovate and use AI in 0:10:29.080 --> 0:10:29.520 this way? 0:10:30.120 --> 0:10:33.760 Obviously there needs to be better regulation in this space, 0:10:33.840 --> 0:10:37.080 but I think first it will be down to companies 0:10:37.120 --> 0:10:40.280 to figure out how they want to use these technologies ethically. 0:10:40.400 --> 0:10:44.440 Technology always evolves faster than the law, so the work 0:10:44.880 --> 0:10:46.960 in the legal space will take some time. So I 0:10:46.960 --> 0:10:50.440 think it's really really important for executives, for leaders in 0:10:50.440 --> 0:10:54.239 this space to upscale themselves around artificial intelligence, to understand 0:10:54.320 --> 0:10:58.679 what the ethical challenges are, what the practical challenges are, 0:10:58.679 --> 0:11:01.200 but also what some of the huge opportunities are in 0:11:01.240 --> 0:11:03.880 that space, and to make informed choices about what they 0:11:03.920 --> 0:11:08.040 do in their organizations. This has been a long time coming, right. 0:11:08.080 --> 0:11:11.920 We've had digital humans like like Little Mikaela back in 0:11:12.000 --> 0:11:15.600 twenty sixteen. They were on Instagram, then they became you know, 0:11:15.800 --> 0:11:18.720 digital flesh and blood. They did ads for things like 0:11:19.320 --> 0:11:21.480 I think it was Prada back in the day and 0:11:21.960 --> 0:11:27.720 bal Main. We had them evolve music careers on Spotify. 0:11:27.840 --> 0:11:30.480 So this is not new, but it's up to companies 0:11:30.520 --> 0:11:33.000 in the first instance, to really figure out how they 0:11:33.000 --> 0:11:35.040 want to do this the right way, and also a 0:11:35.120 --> 0:11:38.000 huge opportunity for companies to lead in an ethical way 0:11:38.000 --> 0:11:38.600 in that space. 0:11:41.920 --> 0:11:45.680 One of the most iconic brands bringing in the holiday season, 0:11:46.800 --> 0:11:49.320 but take a closer look at the new Coca Cola 0:11:49.320 --> 0:11:52.320 commercial and you might notice that it was made with 0:11:52.559 --> 0:11:57.000 artificial intelligence. Social media is certainly caught with it shows 0:11:57.040 --> 0:12:01.120 how lifeless that Christmas commercial is. Book A Colloges put 0:12:01.120 --> 0:12:03.720 out an ad and ruin Christmas and their entire brand. 0:12:03.840 --> 0:12:09.040 It's less festive, more creepy holiday vibes. We's push for 0:12:09.120 --> 0:12:12.959 marketing efficiency, right, how do we create more with less? 0:12:13.200 --> 0:12:15.959 That's just a sort of business one on one. 0:12:17.400 --> 0:12:22.320 I've also seen, you know, entire ad campaigns generated by 0:12:22.440 --> 0:12:26.320 using AI. Campaigns that would usually require entire teams, days 0:12:26.320 --> 0:12:31.199 of shoots, makeup artists, stylists, lighting technicians, directors, photographers all 0:12:31.280 --> 0:12:34.480 replaced by just typing words into this kind of generator. 0:12:34.520 --> 0:12:36.280 There's something sad about that, isn't there. 0:12:36.440 --> 0:12:39.160 Well, there's something sad about the volume that we can 0:12:39.240 --> 0:12:44.200 make in this space. There again, opportunities in that space 0:12:44.240 --> 0:12:46.400 as well. If you think about the kind of a 0:12:46.520 --> 0:12:49.160 deep fake of me. We're not using it obviously to 0:12:49.200 --> 0:12:51.680 create videos of me, but if we need to do 0:12:51.760 --> 0:12:54.560 a little pick up, it might be easier to do 0:12:54.640 --> 0:12:57.760 it with a typing word in rather than booking the 0:12:57.760 --> 0:13:01.000 whole studio and everything else again and disrupt everybody's day. 0:13:01.160 --> 0:13:05.280 So there are opportunities in that space, but there's also 0:13:05.600 --> 0:13:10.640 a real danger that we might be just rehashing, recombining 0:13:10.679 --> 0:13:16.080 all things and not giving artists, creators, directors, producers, writers 0:13:16.160 --> 0:13:19.400 the opportunity to really bring our humanity to this. I 0:13:19.440 --> 0:13:22.880 think the useful way to think about AI is as 0:13:22.960 --> 0:13:26.200 an assistant. This is not a technology that should come 0:13:26.280 --> 0:13:29.440 to replace what we do. It's not there to take 0:13:29.480 --> 0:13:32.440 our jobs, but it's there to enhance us. So I 0:13:32.440 --> 0:13:34.640 would encourage people to think about what they can do 0:13:34.720 --> 0:13:37.560 with AI that we weren't able to do before, and 0:13:37.640 --> 0:13:40.720 think of it as adding to our capabilities rather than 0:13:41.120 --> 0:13:44.320 replacing us. There will be disruption, and I think all 0:13:44.400 --> 0:13:46.920 of us, all of us know that that will happen. 0:13:47.600 --> 0:13:49.520 I always say it's not coming for your job, but 0:13:49.600 --> 0:13:52.839 it's definitely coming for your job description. But it's an 0:13:52.840 --> 0:13:58.880 important moment, I think for especially people who lead businesses 0:13:59.000 --> 0:14:03.360 lead creative things to upscale themselves on what this technology is. 0:14:04.200 --> 0:14:06.280 I would, of course say the University of Sydney had 0:14:06.280 --> 0:14:09.160 does the best work around the effluency anywhere in the world. 0:14:09.200 --> 0:14:12.120 So do come and do this with us, but find 0:14:12.360 --> 0:14:15.160 find ways to understand what this technology can do. It's 0:14:15.240 --> 0:14:18.160 really quite different to what we've been able to do before. 0:14:18.800 --> 0:14:22.680 Even the idea of style engines. It's not intuitive to 0:14:22.760 --> 0:14:25.360 us to think that, you know, things like bananas or 0:14:25.400 --> 0:14:29.280 corporate emos become styles. Catness is a style, right, So 0:14:29.840 --> 0:14:32.200 try to understand what the tech can do and then 0:14:32.320 --> 0:14:35.840 really be very mindful and very deliberate in how you 0:14:35.880 --> 0:14:38.520 implement it in your organization. I think it's up to 0:14:39.160 --> 0:14:42.040 people who lead, whether it's the creative industries or whether 0:14:42.040 --> 0:14:44.240 it's any business or government, to lead the way on 0:14:44.360 --> 0:14:47.320 how we think we ethically want to be doing this. 0:14:47.520 --> 0:14:50.800 It's not happening to us. We are making this future 0:14:50.840 --> 0:14:53.720 happen and the next two years will be crucial. 0:14:53.960 --> 0:14:56.480 Are there any protections that you'd like to see in 0:14:56.600 --> 0:15:00.280 place in order to better protect the creative industries? Because 0:15:00.320 --> 0:15:05.160 I think that there is, given how fast the technology 0:15:05.320 --> 0:15:08.880 is evolving, there is a fear that it could stifle 0:15:09.000 --> 0:15:12.240 that creative production and that creative process. 0:15:12.520 --> 0:15:15.160 I'm not a lawyer, so I live the I leave 0:15:15.240 --> 0:15:19.680 the lawmaking and the details of this to those who 0:15:19.800 --> 0:15:23.440 who know better. But I would want to see protections 0:15:23.560 --> 0:15:27.400 around artists, right. I do want to see protections around 0:15:27.520 --> 0:15:31.640 how their work is being used to train these models, 0:15:31.680 --> 0:15:33.680 because I think at the moment it's a bit of 0:15:33.680 --> 0:15:38.120 the wild West out there, and I am worried that 0:15:38.240 --> 0:15:41.600 AI generated content might end up in the long term 0:15:41.680 --> 0:15:46.760 severely diluting the earnings for original creators, which also means 0:15:46.760 --> 0:15:49.440 that fewer people will want to join the creative art 0:15:49.680 --> 0:15:54.120 So ultimately, I think owners consent in using any of 0:15:54.160 --> 0:15:56.040 this should be should be a requirement. 0:15:56.160 --> 0:15:58.400 Thanks for joining us, Sandra, Thanks for having me. 0:16:02.240 --> 0:16:05.320 That's it for this episode of The Front Page. You 0:16:05.360 --> 0:16:09.160 can read more about today's stories and extensive news coverage 0:16:09.200 --> 0:16:13.240 at enziherld dot co dot nz. The Front Page is 0:16:13.280 --> 0:16:17.000 produced by Ethan Sills and Richard Martin, who is also 0:16:17.160 --> 0:16:18.280 a sound engineer. 0:16:18.720 --> 0:16:20.240 I'm Chelsea Daniels. 0:16:20.840 --> 0:16:23.960 Subscribe to the Front Page on iHeartRadio or wherever you 0:16:24.000 --> 0:16:27.880 get your podcasts, and tune in on Monday for another 0:16:27.960 --> 0:16:29.600 look behind the headlines.