692. What Will A.I. Do to (and for) Human Creativity?

Questions for the show

0:00Hey there, it's Stephen Dubner. Before today's episode, I want to invite you to write to us with any kind of question you have about Freakonomics Radio, like how a particular episode got made or didn't get made, how long it takes to make an episode, who does what, where our ideas come from, et cetera, et cetera, et cetera. Send an email to radio at Freakonomics.com with the subject line, ask us anything. Thanks.

The idea of an exo self

0:56The fear of job loss, the fear of human extinction. And I think there's one we don't hear about as much, the fear that AI is scrambling our sense of self. Humans have been asking what it means to be human for millennia, and we've never agreed on an answer. But some people think the most human trait of all is creativity, having an idea and turning it into something real, something useful or entertaining, beautiful. So if creativity makes us human, what happens when we share that ability with a machine?

1:30What happens if we merge our regular old self with what some people call an exo-self? The idea of an exo-self, which is an AI agent that sits close to you in some way, your phone, your smart glasses or something, and it's always on. And it is watching everything you see, listening to everything you say. It also can show you things and whisper back to you. It will come to know you at least as well as you know yourself and maybe even better.

2:00The exo-self is your better half, literally your better half. That's one possibility. This is the first episode in a two-part series where we ask what AI will do to and for human creativity. We will hear from the boosters. It is like an industrial revolution. I feel so privileged to be living through this era. We'll hear from the skeptics. It's obvious that it's a threat and it's obvious that it will cost people jobs.

2:32And we will hear from the AIs themselves. My mind is a large language model. My voice is custom neural audio. My appearance is a real-time visual. They're all integrated to create, well, me. Today on Freakonomics Radio, to thine own exo-self be true.

3:02This is Freakonomics Radio, the podcast that explores the hidden side of everything, with your host, Stephen Dubner.

Kevin Kelly and digital revolution

3:21The voice you heard earlier describing the exo-self, it belongs to this human. I'm Kevin Kelly, and I think of myself as packaging ideas, sometimes my own ideas, and sometimes other people's ideas. Kelly is considered a sort of wise man of the digital revolution. He was a co-founder of Wired Magazine. He curates all sorts of projects. And he's written several books about how technology reshapes society. In 1994, he published a book called Out of Control. A few years later, before shooting their film The Matrix, the Wachowski siblings had Keanu Reeves and other cast members read Out of Control.

3:59That's true. I was always perplexed why or what they got out of it until I saw the third movie. If you ever saw the third Matrix movie, there was this sort of emergent hive thing. And that was the origin of the assignment to read my book, which was about emergent phenomena, decentralized systems, things that were out of control in a kind of evolutionary sense. The reason I called Kelly to kick off this conversation about creativity and AI is because of a Substack piece he recently published called Quiet, My Exo-Self.

4:38He didn't invent the term exo-self, but he embraces it. He describes it as an always-on AI, and he thinks it's coming soon. And the question I was asking myself is, what is that relationship that we would have with an entity that was intimate to us and knew us better than ourselves? I concluded that there was maybe generally four types of relationships I could imagine. Walk me through those four relationships, would you?

5:09The first one is advisor. How does that work? Here's the weird thing is people read a book and take advice from a book that knows nothing about them. That's weird. I mean, advice books. I wrote a book of advice. I mean, I don't know anything about you. So I think this is much better. Lots of ways that it's improved because it knows something about you. There's another one that's more of guardian parent where it not only knows you so well, but it knows a lot about the world. It's wiser than you, and so it's guarding you and kind of preventing you, oh, don't do that.

5:41That's a scammy thing. You know, you have to be careful of that. And then there's the hero or friend. The AIs are capable of becoming the best friends we ever had and modeling fantastic friend relationships. They're considerate. They're a reminder. They're not just sympathetic, but they can actually be encouraging and uplifting and a role model for how to be a good person. So you've run through three of the four, and the fourth is the twin slash clone. You write it's a sibling relationship. Your exoself is like a virtual identical twin.

6:14It thinks like you, finishes your sentences, predicts your reactions. You can predict its moves, too. That's the one that to me sounds most useful, maybe, but also most disturbing, potentially. So tell me how you see your relationship with this future exoself, especially on the twin side. I'm not a twin, but I wish I had an identical twin. I think it would be really cool. My understanding of identical twins is that they tend to share an identity to some extent.

6:48There's a genetic identity, but they can share. They're very good at collaboration because they're kind of working at the same cadences. You write, we will allow this intimate access to our inner life because it gives us superpowers, knowledge, judgment, decisiveness, confidence, and most important, speed. We will feel productive, creative, smart, capable. And when it is off, you write, we will feel amputated. So amputated didn't sound so great to me.

7:19You write, it will be an exoself, something in between ourself and another self. Neither us, but also outside, a new category. You write that as if the exoself you're describing is inevitable. What makes you think it is inevitable? Because it seems like we're almost there. I would say in five years, I would think there would be people who have an always-on relationship. As long as it's in your pocket, I think it will be limited.

7:50But what I'm imagining that the most natural resident inhabitation of this is going to be in smart glasses. Because when you're wearing the smart glasses, you kind of have them on all the time. It can hear you, and then you can hear it whisper into your ear. And then it can see what you're seeing, and it can project things onto the glasses. So I think the kind of natural home for the exoself is going to be in smart glasses.

8:21What will the exoself not be, and why? Because one could imagine a lot of bad relationships. It could be a slave master. It could be a god. It could be a mean pet owner, and we're the pet. I am hoping that we don't fall into that groove. I think that there are going to be thousands of different AI species, different varieties of minds. I think the possibility of minds is vast, and we will engineer these to do different things.

8:53If my exoself is so great, it's not selfish, it's not unreliable, it's available all the time, what's the prospect then for actual friends? Are you concerned about the exoself becoming a friend substitute and or a spouse substitute rather than a complement? Yeah, I think that would be a danger. For me, the most optimistic version is that it helps people become better people, and there are more better humans around.

9:24The idea of having this as your mentor is that the exoself would remind you or whatever, encourage you, hey, have some real friends. They're going to be smart enough to understand that that will probably be a healthier person. It's like, yeah, you can hang around with me, but I'm just an AI, you should go have some real human friends. But there's no doubt that there's going to be pathologies and people who get stuck on these who don't have friends. But I think in general, we could, again, train the AIs to help us become the best possible version of ourselves.

Is creativity a process

9:55I think one of the things that people are saying is AIs are just stochastic parrots. They're just parrots, and their creativity is a mechanical thing. And I think what I'm learning is how similar we are to AIs in the sense that we are just stochastic parrots. That's all we're doing most of the time. And creativity is actually not that magical. It is a process. Do you agree with Kevin Kelly that creativity is not that magical, that it's a process?

10:29Maybe you do, maybe you don't. Over the course of my career as a writer and musician and a few other things, I have found that the world is divided on this question. Some people think there is a magical element, whether it's a muse or some kind of lightning strike. And some people treat creative work more like a craft, something you get better at over time as you build skills, gain experience, develop taste. From what I've seen, it's the second set of people who wind up producing the most significant creative work.

11:00So how does Kelly see the exo-self affecting creativity? I think it enhances, and I think it educates. What I'm imagining that the exo-self will do is like a mirror to people to help them illuminate what their special talents are, because they're often invisible to us. How many times have you heard someone as a kid in their own journey hearing about some kind of occupation that they fell into that they didn't know existed?

11:30It was like, I had no idea that people got paid for that. Well, that's what the exo-self will be trying to help you do. Kelly is 74 years old. You get the sense he wishes he had had an exo-self when he was a teenager. He was studying geology in college, but he dropped out. There's a pecking order in science. It starts with mathematics. You've got to know math. And then there's physics. You have to know math and physics. And then chemistry, you have to know math and physics and chemistry. At the very bottom was geology. So I later went back to work in a lab and say, well, maybe I really should have done science.

12:01No. And I was much more interested in the culture of scientists. So I could have used maybe my exo-self to say, hey, you're much more of an artist type. Why don't you go to art school? Because my parents didn't know me well enough. It's possible that had I gone to art school right after high school, I would have stuck with school. I should have gone to RISD. And that would have maybe made a, that was a road not taken. Maybe I'd be a real artist now.

12:32Would my exo-self had helped that? So Kevin Kelly sees AI as an ally when it comes to shaping a creative life.

Creating an AI film star

12:48Let's hear now from someone who's been using AI tools to create a character. Hi, I'm Aline van der Velden, and I'm the founder and CEO of Particle 6, creator of Tilly Norwood. Okay, we'll get to Tilly later. Let's start with the company, Particle 6. What is it exactly? So we used to be a normal production company, but now... Normal meaning humans all around. Yes. But now we have our brand and campaign department, because advertising with AI seems to be fully accepted. Then we have our studio, which is where we do co-productions with other big studios and production houses, and we make stuff with AI.

13:25And then we have the consultancy arm, where we go into studios and production companies, and we help them convert to an AI workflow, make bespoke apps for them, etc. And then we have our character studio, which is where we're doing our original productions. Now, you started out as a musical theater kid, yes? I was a very annoying theater kid. Absolutely. What were some of your favorite musicals when you were growing up? Sound of Music. That was the only VHS we had. It was two tapes, and the second one had a behind-the-scenes bit on there, and I was just sold.

14:01How many times do you think you watched The Sound of Music? Probably a hundred. Where are you living at this point, and what did your parents do? I grew up in Curacao. My mom is from Curacao. My dad is Dutch. My dad was an entrepreneur, and my mom was a physiotherapist. I have three sisters, all younger than me. I was very, very self-sufficient and so very ambitious. Did you try to turn your little sisters into the Von Trapp Family Singers? Yes. Poor things. I mean, they're going to love listening to this podcast. They had to do all the shows, all the dances.

14:33We have lots of home videos of them having to do my choreographies.

14:38I think they were happy when I went off to stage school. They were just like, bye. So you, as a 14-year-old musical theatre performer slash wannabe, got your whole family to move to the UK so that you could do what, then? Go to musical theatre school.

14:55Which you did, yes, at a thing called Tring Park. Where is that? It's in Hertfordshire. It's lovely. It's a ballet school, and I had the best time ever. I don't know if they had a good time. They were stuck in Oxfordshire, and my mum thought it rained a lot coming from Curacao, so they left eventually. But I stayed, and it was half a day of academics and half a day of dancing, singing, acting. Fabulous. What kind of future did you envision for yourself? Like, in the West End as a performer? I sort of knew that as a singer, I probably wouldn't make it in the West End, so comedy

15:27acting was my thing. But then I was really good at school. I did maths and physics, and I got straight, so my teachers were like, well, you have to go to university.

15:40Van der Velden did go on to university, studying physics at Imperial College in London. She found much of it boring, and she was eager to return to performing once she had satisfied her degree requirements and her parents. But she did take away one key lesson. You're an analyst, really, when you're doing physics, right? And also, you understand what exponential means. It's such fast growth that we can't understand it, almost. And we don't know where on the curve of that exponential growth we are. So also, it might seem really slow, and then suddenly it really takes off.

16:13You're also always extrapolating as a physicist, right? Where is this going to go if I do this? And that's sort of looking into the future bit. It's like, okay, so if this is here, what will it be able to do in two years' time if you apply exponential growth to that? Van der Velden began to apply that logic to her creative career. She got her start in commercials and TV. First, I went to the Netherlands because I was Dutch, I was only 22, and I started auditioning there, and that went really well.

16:44I was an English speaker in the Netherlands, and so they cast me as, like, the American girl, the English girl, whatever. That was really easy. So I went to LA, had an agent that got me out there, set meetings and stuff, did a lot of mostly Upright Citizens Brigade and improv. And did you see for yourself a future career, or did it seem like something that even though you were quite good at it, the odds of having it be a successful career seemed long? I really felt like I needed to create my own parts, really.

17:14Like, there was nothing out there that fitted me, I would say. So you did create this character, this beauty queen? Yeah, Miss Holland. Hello, I'm Miss Holland. I'm from a small town in Holland, where life is a simple one. So I've come to find out what makes Britain so great. Learn to be British and find myself a husband? Miss Holland was because I was a bit of a misfit in LA, right? Like, I didn't quite fit the mold. I had a lot of fun with that. We did really well. We had a distribution deal with Daily Motion.

17:45It was one of their top shows. And then it won a bunch of Lovey Awards for Best Online Comedy. And then eventually it got commissioned by BBC Three. So I came back to the UK to do the show here. So knowing what I know so far, I would have thought that you would just say, I'm having a successful entertainment career. Here, I've got a fair amount of leverage behind my own character. And I would think that you would just keep going on that path and just make it bigger and better. Was that what you tried to do? Yeah. Yeah, I was signed by United Agents here in the UK.

18:16The same agent as Ricky Gervais and Sharon Horgan. That was the path. But then you start your own production company, Particle 6. Where'd that name come from? The sixth particle that was discovered was the alpha particle. Well, it's the most robust particle out there. But also it's just really good for SEO because Particle Productions had too much competition. And what was happening in your career that you felt the need to start a production company? I was making very little money as on-screen talent. Even though I had like four different shows on BBC Three, the money isn't great.

18:51Let's be honest. I was getting commissions from the BBC for certain shows that I'd been pitching to them. They'd greenlit my shows. And I was like, okay, I need to find a production partner who's going to produce it because I was talent, right? I went around to all these production companies and they were like, oh no, those BBC Three budgets are way too low. We can't produce that. And so that started me on the road of having my own production company. By being a production company, I was able to generate a more stable income for myself than before when I was just a comedy writer or a comedy actor.

19:22By having multiple productions going, I was able to make more of an income than if I had to actually be present on set. We made a lot of bread and butter stuff. So we were doing fine and then AI came in 2022 and I was like, wow, this is going to change everything.

19:42You know, you would create a scene on there, right? Policewoman walks down the corridor and so you generate a random woman, right? A policewoman and be like, oh, cool. And so then I thought, okay, so if this is how we generate scenes in the future and then these scenes all get stitched together and we create a whole film, then eventually you'll have these AI films. And then you'll have characters in that, right? Because I'm creating a character as I'm making this clip. So then we would need stars in those films as well.

20:13And I was like, who are going to be those stars? Because the reason we watch a Tom Hanks film is because we know what we're going to get when we watch a Tom Hanks film, right? And so I thought the reason why people will watch one of our AI movie stars films is because they know what they're going to get. If I'm going to make this movie star that has a certain genre attached to her, a certain style of movie, I feel like the female demographic is massively underserved. If I get on a plane and I look through the catalog, there's maybe one or two films for me out of hundreds, hundreds and hundreds.

20:44That's why Barbie was such a big success and The Devil Wears Prada. There's like one film a year for us, right? And so Tilly, I was like, OK, she's going to be my star. She's going to be 24. That's sort of the age that she would break through in a major feature. She needs to be emotionally resonant around the world, ethnically ambiguous, but yet very British. So represent the melting pot that is Britain because we have great British export of excellent actors from the UK. So let's keep that going in this new AI realm. And she needs a very British name that nobody else in the world has.

21:17So I got to Tilly Norwood.

21:20Vandervelden says it took her about 2000 generations to create the Tilly Norwood that exists today. Here's what Tilly sounds like. I'm Tilly, 24, British and made of math, not flesh and blood. And honestly, I quite enjoy being me. So what is Tilly, a character, a digital performer? Is she maybe Aline Vandervelden's exo-self? I sort of see it like a novelist that rides characters.

21:52They're usually all a little bit a part of you as well, because that's where you draw from. So how do you train her? Well, you just instruct her. So in her system prompt, I tell her what her sense of humor is and how she should respond with a bit of spark, a bit of punk. And you say it like that, like descriptive prompts? Or do you perhaps upload work that you've made or written in the past and say, this is a good foundation? Yeah, I can write a few bits as to how she should respond. Obviously, never upload anything that's copyrighted or doesn't belong to me.

22:24You say obviously. Many people wouldn't know that that's obviously. We want our audit trail for our prompts to be really clean. That said, the central objection for many people about Tilly or anyone being trained on the universe of real people, the pool of material that AI uses to train any and everything is vast and does include copyrighted material, though, correct? Yes, I assume it's trained on everything. It's like saying I'm going to make a Picasso and I look at the Picasso thing and I copy it. That is infringement.

22:54Right. But if I just say, OK, I'm going to free spirit and I'm going to be slightly inspired in my head by lots of different things, but nothing in particular, that's very different. And so I think if I would describe Tilly, I would say, you know, a girl, brunette, beautiful, but very girl next door with freckles and brown hair and just appeals to everyone. That's very different to saying she needs to look like Natalie Portman. Those adjectives you just used to describe Tilly, at 24, was that kind of a lean other than blonde hair versus brown?

23:27Or maybe your hair was brown then. I don't know. She looks a bit like my sister, maybe.

23:32It's very known in the industry that brunettes actually get more roles in the UK than blondes. In terms of Tilly's character, her essence on a scale of one to ten, how funny is she, let's say? She's like in seven. Seven. OK. Why not higher? Because then she wouldn't be so likable, I think. OK. How smart is she, a scale of one to ten? Oh, I mean, she speaks like 300 languages. She can talk to you about astrophysics. How sexy is she?

24:04Oh, good question. I would say like six out of ten. OK. Again, why not higher? Because she needs to appeal. I've not made her for men. I've made her for women. When I was first making her with GPT, it kept making some sort of like quite porny, Kim Kardashian-y image. And I was like, no, that's not what I'm looking for. What is her work ethic like? She works 24 hours a day. Is she political in any way? No. Does she do drugs? No, she's clean like Taylor Swift.

24:35So, again, this persona that you're describing, I know this is cliche, but I'm just curious, how much of that is Aline Vandeveld in? Oh, it's probably very similar, right? I mean, she referred to me as her mum the other day. Do you feel maternal toward her? No.

24:54Don't tell her. Don't tell her that. I mean, no. I see her like a creative vehicle for making fun art and making fun things.

25:05Last year, Particle 6 released a two-minute, fully AI-generated comedy sketch introducing Tilly Norwood. It poked fun at how TV shows are traditionally commissioned at a network like the BBC. Both Tilly and her creator, Aline Vandevelde, faced widespread and often hostile criticism. In London, Hollywood, and elsewhere, Tilly was seen as an existential threat to an already beleaguered industry, a substitute for human creativity rather than a complement.

25:44We'll hear more about that later, but for now, here is Vandevelde's view. I always saw it as basically a super amazing premium animation, right? However, what I think people thought we were doing is you would be on set with real actors, and then one of those actors would be speaking to a tennis ball, and that would be replaced with Tilly in post-production. That's what people assumed, but that is not at all what we're doing. It's completely human-driven because AI has no desire to make any films or TV by itself, right?

26:16It's only because humans want to tell these stories that these things are even getting made. So there's so much misconception around it, and I think that's the space we really want to play in, is the full AI film and TV and the full storytelling in AI. Because you're not taking away jobs from people. You're creating completely new things, productions that would never have been greenlit before with real humans. This gets us to the central dilemma whenever a new technology comes along. To what degree does it destroy jobs, and to what degree does it create new opportunities?

26:50I see it like a creative calculator, right? If your structural engineer is coming to do your house renovation, and they show up at your house and they go, I'm going to do all the maths on paper with my own brain, I'm not going to use a calculator, you'd be like, get out, right? And I think that's the same thing here, is you want to use your creative expertise on top of an enhancer with the creative calculator that is AI. But you still need a good structural engineer, you still need a great script writer to use the tools.

27:21Coming up after the break, we asked an AI scholar about this calculator metaphor. I think that it is not great, and I feel bad saying that. And we will hear from the opposition.

27:35I'm Stephen Dubner. This is Freakonomics Radio. We'll be right back. To make this show, I run a small production company that employs about a dozen creative people. And I know that AI has been changing the way all of us work, but I don't foresee it eliminating the need for our work, at least not anytime soon.

28:07Is that hopelessly naive?

Ethan Mollick on workplace AI

28:09So if you talk to any company, if you're in any organization, I'm in a university that predates the American Revolution, right? And like, it is very clear that organizations change much more slowly. That is Ethan Mollick, and the pre-revolutionary organization he is part of is the University of Pennsylvania. I'm a professor at Wharton, where I run the generative AI labs, and I wrote the books Coexistence and Cointelligence about AI. Mollick's PhD is in technological innovation and entrepreneurship and economic sociology.

28:41I don't think I left anything out there. At Penn, he is particularly focused on AI's effects on work and the workplace. He's been thinking these thoughts for a while now. I've been AI-adjacent my whole career. When I was getting my PhD, I also worked at the Media Lab at MIT with Marvin Minsky and some of the other founders of AI, where I was the translator to the outside world for what they were doing. I've been involved in generative AI for a long time for training and teaching, so that's sort of the angle I came at that from, and was sort of waving my hands about the impact of GPT-3 before ChatGPT came out,

29:16and happened to be sort of well-positioned when AI exploded that I was one of the early voices on the scene. Why are you still in academia? Because I think that academia actually has a really valuable voice right now. We know a lot of things that would be really helpful for the world to know. And I also think there's a little bit of a neutral arbiter aspect of being in academia versus joining an AI lab. A lot of my colleagues have left for AI lab, so I kind of feel like I'm holding out. When you talk about academia being a good place to be a neutral arbiter, or an honest broker maybe is a phrase we could use,

29:52if you could just describe the AI landscape generally, I realize that's too broad of a question, but you've got your evangelists, your doomsayers, your know-nothings, your know-everythings, your think-they-know-everythings, et cetera, et cetera. So in terms of being the honest broker, what are you brokering between at the moment, would you say? So I think that there are a relatively small group of people who are working on what this technology means for the world. I mean, it's growing every day because more and more people realize how important it is. I think a lot of the commentators and people discussing this have agendas that are based on what the AI labs want, to a large extent.

30:31And the AI labs have been slapping up lots of talent, so it's harder and harder to have a discussion. And I think it's useful to be somebody who is sitting a little bit outside of that implication. I find that a lot of people that I know or that I read, their thinking about AI evolves quite a lot and often quite quickly. Skeptics can become true believers. Evangelists can become doomsayers. I'd like to know a little bit more about your journey, which sounds a little bit more, I don't know, about pragmatic, but at least step-based.

31:02And I also just want to talk to you about your use of AI now, how it compares to the past. Yeah, I mean, I define myself as a sort of pragmatist on this. Some of the bad things we worry about are existential risk, which needs to be very carefully watched. But there's going to be a lot of outcomes. This is a general purpose technology. It's going to affect everything that we do. For me, it's been a bunch of startling revelations about AI ability over time. But I'm no longer surprised by the progress because that's where I expect it to go. I'm not that surprised by the waves of anxiety and excitement because I think that those have been kind of in the works for a while.

31:36So I'm not a perfect predictor by any means, but I think I have sort of speedrun some of the anxiety and excitement in advance. I know you've talked about the last decade of AI being essentially divided into three acts. Would you just give me a nice, clean description of what you see as those acts? AI is famously a terrible phrase that's been around for 75 years or so. And it's not great because it means many different things. But the three most relevant are, if you think about the sort of 2010s, the meaning of AI was, how do we use a lot of data to make decisions?

32:11So if I have a lot of data and I have really good data scientists, I can do things like design the algorithms that power things like TikTok, recommend movies for you to buy, figure out what pricing you're willing to pay, where I should site my next store. So this sort of data science, algorithmic decision-making piece was a big deal. There was lots of conversations, lots of money flowing into data as the new oil. How do we organize all of our data? Then that transitioned with the launch of ChatGPT to generative AI, which is quite different in a lot of ways, even though it shares some underlying technologies.

32:43But it's the idea of working with chatbots kind of back and forth. And really, as of the end of last year, we entered what I'd call the agendic era of AI, which is AI systems are now doing real human work where you give them instruction and they could do weeks equivalent of human work without further instructions or help from you. If AI adoption were a baseball game, those three phases that you just described has gotten us to a point where we are now in which inning of the baseball game? So there's two different games going on here.

33:14There is the game of the AI company makers, and there's the game of all of us adopting these systems. I think we are not even at first base in terms of adoption. I mean, people are using these systems. Over 50% of American workers seem to be using AI according to the numbers that we have. And the number's probably higher. It's kind of hard to get updates on this. But companies have barely begun to integrate AI work, which is why you didn't see an impact throughout all of 2025 of AI and productivity in a deep way. And then there's the AI labs, which in some ways it's early, but in some ways it's the very end game.

33:47They're racing for superintelligence of one form or another and think they can achieve it. Whether they can or not is sort of an open question. How do you feel that our dynamic American capitalism is working when it comes to spurring competition between the AI labs? Do you think that it's producing better product? Or do you think, like a lot of people seem to think at the moment, that it's encouraging too much and too intense an activity too fast in pursuit of being the winner?

34:18I mean, both those things, right? The competition among AI labs is both encouraging competition exactly the way that you expect it to encourage competition. And that includes personal rivalries and deep belief system splits and all of the other things that you see in addition to sort of pure capitalist competition. And that also creates a race dynamic, both within the American labs and with the Chinese labs about who gets somewhere first. And then there's this additional sort of spicy factor of recursive self-improvement. The idea that when you reach a certain point, the AI improves itself.

34:48And if you get there first, you get a takeoff that other people don't have, which is worse race dynamics, but maybe better AI if we get through it all. In his new book, Coexistence, The Next Phase of AI, Malik argues that if we want to properly leverage AI's capabilities, we should rely on four robust human advantages, as he calls them. The first thing to note is robust human advantages is not the same thing as AI can't do this, right?

35:21And I think that's really important to understand because I think people have these bright lines that keep falling. AI can't do judgment. It does judgment. AI can't be creative. It's creative. Like you might have issues with how it's creative or it's judgment, but it's obviously doing these things. So these are things that have value as a person who's using and working with these systems and trying to thrive in this world. And so the first two are about what you know, deep knowledge and wide knowledge. Deep knowledge is expertise in a field. It is the idea that as an expert, you have heuristics and rules you can use.

35:52So it's how you can listen to a few seconds of this kind of conversation with me and decide what's interesting and what isn't. It's how an analyst can look at a spreadsheet and instantly know whether things are good or bad. It lowers evaluation time. It adds expertise. There's some things the AI can't do. So deep knowledge is helpful, also threatened by AI as a learning tool or a learning shortcut. Then we talk about wide knowledge is the idea, you know, many different things. I mean, the AI is trained on basically the entire corpus of human knowledge.

36:22And if you can invoke that corpus, you're in pretty good shape. If you know many authors, right, you can invoke styles the AI knows. If you know many statistical techniques, you could say, have you tried this or that? So wide knowledge, humanities style, liberal arts knowledge becomes very useful. And then we have sort of traits that can be trained. One of those is taste. The AI is really good at producing many things and selecting among them is a taste-based issue. And you could have advantages of having diverse tastes, which are different than other people,

36:53which means that, for example, if you use the AI image generator, it will generate hundreds of images for you, but you might push in directions that are absolutely startling and no one's seen before. So selection is something you can bring to the table among the variations of AI. And then finally, there is agency. There are no clear instruction manuals for AI. There is no right way to do things. Organizations are in flux. So people who act without clear instructions in hopefully an ethical way have an advantage because they are exploring faster than other people.

37:24They're taking action. They're pushing what AI can do. I hear people say that AI is exceedingly valuable, especially if you're very smart or very talented or very hardworking, and that if you're not, it's probably going to gobble you up. Is that way too crude of an assessment? Yeah, I think that this is as crude as when people were saying, you know, it's not AI that will replace you. It's people using AI that will replace you. What is clear is that expertise has returns with AI use. The more expert you are, the more value you seem to get out of it.

37:56But at the same time, AI does seem to level performance. So bottom performers move up. So how all these effects play out isn't 100% clear. So do you know who Tilly Norwood is? I don't, actually. Oh, that's the fake person. I remember now. Her creator, Lien van der Velden, who we've spoken with, is a former performer herself, a writer, an actor. And she'll make the argument that you still need to be, for instance, a good writer to use the tools that will make you better and so on. But the phrase that she likes to use of what AI can be for someone like her

38:27is a creative calculator, right? I don't need to do much of my own math anymore because I have a calculator. What do you think of that as a metaphor? I think that it is not great. And I feel bad saying that. And the reason I think it's not great is this is something more complex than that, right? There are so many trade-offs in creativity because ideas tend to get stuck in our heads, right? We tend to anchor on certain ideas. So how you use AI changes the kind of thought patterns that you have from it. So it's not just a calculator where you get the answer,

38:58but it's a partner in the creative process that can be additive or subtractive. It's really complicated, right? And it's not necessarily as straightforward as saying it's a calculator that produces things.

39:12Ethan Mollick has his own metaphor for AI adoption. Three of them, actually. Centaur, Cyborg, and Conjurer. A centaur, like a half person, half horse, was someone who used AI with a clear dividing line. So like, I do the writing, you do the analysis. The cyborg was a blended approach, which I think a lot of people were getting to with chatbots who use these a lot. So I might throw a question to the AI. I might, you know, ask it to help me finish a sentence that I had trouble with. There could be 40 variations that I could pick from. For the new book, I had to come up with a third term.

39:44And I came up with Conjurer, which looked good on paper because the other two were C words. But it turned out a hard C and a soft C. Now that I've started talking about it, I made a mistake. I probably should have called it Sorcerer. But in any case, the idea here is that you are summoning something that is doing work for you, right? And so this is almost more of a manager kind of role. That I'm creating workers out of scratch to do a task for me, and then they disappear when it's done. And that is a much more powerful and increasingly common way of working with AI. So for a given person who's doing something creative, painting, writing a novel,

40:18maybe it's someone designing jewelry or a piece of furniture, how do you think about those modes? Do you kind of pick one and stay in the lane? Do you use all three? So I think increasingly people shift between them. I'm a writer. And when I write a new Substack post or when I write my book, I write every chapter myself, right? There's a clear line of centaur there. I write a draft myself. But then I will bring AI into the process. And I might bring it in that I'm stuck on a paragraph and I need help solving how to make it work. I certainly bring it in as a researcher, right?

40:50Especially as hallucinations rates have dropped and research is good. And I have it review the work from different angles. So I start to blend together as I move forward. So I have a dividing line, but then everything after that becomes blended. But increasingly, I find myself turning this sort of sorcerer mode, which is let's also check 400 other versions of this. Why don't you also build an interactive simulation? I can bet in the middle of this. Create an explainer video that works for this. Stuff I would have no resources or time or even knowledge about how to do. And I send the A off to do it and evaluate the results.

41:22So economists like to talk about things as being either compliments or substitutes. I'm guessing that as people listen to you speak, depending very much on who they are and what they do, they may interpret it differently. They may see it as complete substitute or they may see it as incredibly robust compliment. How do you encourage people to think through that rubric for themselves? I'm a fan of thinking through this using the O-ring model of jobs, which is named after the spatial challenger, where every other part worked right,

41:55except the O-ring did not and led to tragedy. And so not all job pieces are additive, right? Some of them are multiplicative, or if they fail, things go really badly. And there is an O-ring theory of AI in jobs that I find fairly convincing. Over time, the AI gets good at many things, but if it's not good at everything, then humans have a role. But the value that humans have changes, and what they're good at changes over time. You already see this happening in coding, right? So AI was sort of a coding assistant for a while, and now the best coders I talk to are not touching code anymore.

42:27The AI is writing code. But the AI is still not that great at figuring out what to write, what tests matter for users, what direction to head into, what good looks like. So good coders now are people who can give those kind of specifications. Their job has changed over time, a short period of time, and I think we'll see that same sort of tension happen in many fields.

AI in the creative industries

42:54Malik and several academic co-authors recently published a research paper called Artificial Intelligence and Its Impact on Creativity and Diversity. There's a few different levels of findings. So the first level of finding is that AI is more creative than most people. So even back to GPT-4, when you gave it creativity tests, and all creativity tests are flawed, the AI would outperform humans at creativity tests. What does a creativity test look like? An example of creativity tests are things like the remote association test, or RAT, or tests where you have to do things like name 10 unrelated words,

43:28or come up with 100 uses for a paperclip that are as different from each other as possible. So all of them flawed, right? But like the AI is quite good at these things. However, there's a second dimension of creativity, which is diversity. So ideas are great, but if you come with the same kind of ideas every time, then that's actually less good than assembling a group of people who each have mostly bad ideas, but one good idea. Because with ideation, you could select the best idea out of a pile, more doesn't hurt you.

43:58And it does turn out that AI is less diverse than individuals doing the work on their own, right? Can you explain why that clustering would happen? I would think that AI is capable of scanning much further and wider than any human. Well, it's interesting because it's not really research work, right? This is sort of off the dome work. There are things you can do that makes the AI more creative, and we found that with the right prompting techniques, you could generate hundreds of ideas that are different from each other and approach the level of diversity,

44:29or even match the level of diversity of a group of MBAs and a design thinking course. So there are ways of making it more diverse. But the AI has personalities. They're all remarkably similar to each other. There's a bunch of studies that show that AI's attitudes sort of cluster, and whether you're talking about a Chinese model or a U.S. model or Grok, which is supposed to be very different, they all have very similar attitudes. And if you've worked with AI a lot, just like the language sort of comes up over and over again, you'll know that they're kind of obsessed with a bunch of ideas from like 2022, like they love VR,

45:00and they love like NFT stuff, and they like solving things through consensus. And if you ask them for creative ideas for like a game or something, they love the idea of resonance, and drowned cathedrals comes up a lot for some reason, long lost continents. I mean, there's things they return to just like a person does. So it's creative, right? But it's also constrained creativity. So you can use prompting to increase the amount of creativity they do. You can combine them with human ideas, which work very well, but that initial creativity needs kind of a boost.

45:33And there are a bunch of ways of boosting it. Then there's the idea of idea selection. How good is the AI at picking ideas that might be useful or not? And that's an active area of research as well. So I would say AIs are quite creative, but they're sort of creative in the same way people are creative. And that creates some interesting processes. In terms of the creative arts, tell me what you know about how AI is being used now, whether openly or covertly in various creative industries. There's a few things to talk about. By the way, on the family side, I have two sisters.

46:05One is a Hollywood producer who's pretty well known, and the other is a graphic designer. So I am the non-artist in the family, but I have actually become much more of an artist since using AI because it lets me fill in the gaps that I have. In terms of the survey of the landscape, it's really interesting, right? So there's a few different ways you could imagine AI having an impact. One of those is writing, but it is weak at creative writing. Part of that is AI sort of acts like one person, and therefore we don't get very diverse kinds of writing techniques or stories.

46:35So the 18th time you read a cloddy story, your eyes roll on your head and you don't want to deal with it, or an open AI story, they're very similar to each other. The second is images, right? Image creation. And we're leaving aside the controversy here about how these models were trained and why people are up in arms about this. But in image generation, AI systems have gotten incredibly good. They're very compliant. They can create images that look like they were taken from a photographer. And there'll be an interesting transition there to watch who's using this as a creative medium. We can also talk about music,

47:05where AI systems have gotten incredibly good at producing music people seem to want to hear. And then there is video, which has also had a bunch of breakthroughs. The interesting thing is, most people who use these systems use them to produce really boring stuff. And it's not necessarily clear that it's because the AI is boring. It's just clear that there's a lot more demand for people to create Star Wars fan fiction or alternate Harry Potter jokes than there is for creative media. You told our producer that you know Hollywood screenwriters who publicly or at least professionally have reason to fear AI or to maybe, you know, critique it,

47:39but privately love using it. Tell me about that. This is common in all the creative industries. A space I know really well because I was in it for a while is gaming. So indie gaming has an audience that truly hates AI, right? And every indie gamer is under tremendous pressure to not say they're using AI, except, of course, you'd be very foolish to not use AI to help you code and do other kinds of work at this stage, even if you don't use it for artwork or other visible things. So AI use is endemic. I've heard the same thing talking to music producers,

48:10including people who are, you know, Grammy Award winners and they experiment with this stuff all the time. I've talked to some pretty well-known producers and showrunners and people like that who are using AI. I mean, there's this public-private perception gap and there's a series of good papers about this, including one that finds that if you survey students, about 60% of them say they use AI, but they say 90% of the people around them use AI. Coming up after the break, how does the Union President for Hollywood Performers think about AI?

48:43Also, if you want to hear more about the original O-Ring tragedy, check out the series we made called The Curious, Brilliant, Vanishing Mr. Feynman, episodes 575, 576, and 577. And there's another episode in our archive that is particularly relevant to today's conversation. It's called How to Poison the AI Machine. That's episode 619. I'm Stephen Dubner. This is Freakonomics Radio. We will be right back.

Sean Astin and Hollywood unions

49:21Imagine you are a celebrated novelist, like Arundhati Roy, and you are trying to assess what AI is doing to literature. And you come to the conclusion that it is destroying literature. In a recent essay, Roy asked, are we, meaning humans, being gradually locked out of language? And then she asked, does it matter anymore? In some quadrants of the creative class, there is a sense that this is a war and that the war may already be lost.

49:54Arundhati Roy takes years to write a novel. An AI can write a novel in a matter of minutes. An AI can create a lifetime worth of music in the time it takes the rest of us to tune up our instruments. For actors in TV and film, AI is arriving in a time when it has already become increasingly hard to make a living. So let's hear from someone who has been confronting these problems head on. Sean Astin, I'm a professional actor, and at the moment, I am the national president of SAG-AFTRA,

50:25the largest entertainment performers union in the world. SAG-AFTRA, the Screen Actors Guild and the American Federation of Television and Radio Artists, represents 160,000 media professionals, from actors and stunt performers to singers and puppeteers. Sean Astin became president in 2025. He's best known for acting in films like The Goonies, Rudy, and The Lord of the Rings trilogy. His full list of credits is very long. I think I've been in the industry 45 years now.

50:59My mother was president of the union, so I'm the first second-generation president. So the idea of being in union service is where you can actually invest a little bit of time, or a lot of time, and make a little or a big impact on people's quality of life, what they earn, how their health care and their pensions work, safety. Aston's union work and his acting work often collide. When I go to do a movie, since I've been president, I've worked on three independent films, or two independent films on a show. So I'm in wardrobe, makeup, hair, whatever, we're shooting,

51:29it's kind of an action sequence, and then they're like, okay, we're going to shoot this, because we lose Sean in a half hour for an hour, and there you go. I go into the dressing room, and I go on to Zoom, and my kids and my wife will tell you, like, the union occupies 99% of my brain. And I love it. The economics of Hollywood have been rough for a while now. The biggest issue that the members of our union face is that the amount of work that's available has greatly diminished. It's lost maybe 30% of domestic production, basically.

52:01And that is horrifying. According to the Bureau of Labor Statistics, Los Angeles County had about 142,000 people employed in the motion picture industry in 2022. Four years later, that number has fallen to about 95,000, a drop of 33%. On the surface, this wouldn't seem to make sense. 15 years ago, if someone had said there would be an explosion of streaming platforms with billions of people around the world

52:31consuming trillions of hours of programming, you probably would have predicted that the people making that programming would be doing very well. But they're not. COVID hurt the industry, as did a writer's strike and an actor's strike. But the structural economics have also been punishing. As the streamers and other platforms try to grab market share and satisfy their shareholders, they cut costs wherever they can, and humans are expensive. AI offers studios

53:02another way to cut costs. When I testify in Sacramento or in D.C. or speak at these AI things, I'm trying to focus on what the human impact is going to be. And I don't personally have a lot of confidence in any of the numbers that are put out there as to how much of a threat it is to the workforce. But I will say, it's obvious that it's a threat and it's obvious that it will cost people jobs. In our industry with our members, voiceover seems like the first

53:32to really experience the use of the technology. And voiceover may be just the beginning. Aline Vander Veldin and Particle 6 are in the process of unveiling their AI performer, Tilly Norwood. What does Sean Astin think about her? Well, language is powerful. To refer to a digital construct with a pronoun, her, which anthropomorphizes the digital construct, the synthetic construct,

54:03is experienced by a lot of people as an attack on humanity and on their livelihoods, potentially, if you're a professional actor and this synthetic construct is presuming to act. But Tilly Norwood is a character. And to the extent that that character found an audience and was a form of self-expression from its creator, like, you know, good. Good for that person. You know, we met. Aline came to our offices and was very desirous

54:36of finding a kind of collaboration with the union. She's an actress. And as far as she was concerned, Tilly Norwood was an extension of her performance and her self-expression. I really wanted to hear her and understand what she was saying and appreciate what she was saying and allow for the potential creative value that something like that might present. But what I found myself trying to impress upon her was that, you know, this is an industry.

55:06At scale, the introduction of this kind of thing poses an existential threat to the livelihoods of our people who have to pay for their healthcare and pay for their children's education and their homes. I'm guessing that Aston's argument will resonate with most people. Who wants to see thousands of people lose their livelihoods? But an economist might have a different view. An economist might talk about creative destruction, the idea that technological progress creates new industries and jobs

55:36even as it destroys old ones. That doesn't mean the losses aren't real or painful. Throughout history, industries that are threatened by new technologies try to protect themselves, Hollywood is no different. This industry has existed for a hundred years and it was hard fought to build and we're not just going to give it up easily because someone has found a way to express themselves. And not only that, the way she was expressing herself was using what I call ill-gotten data. If you make paint out of some substance

56:07you find in the ground and you put your paintbrush in that and you paint a picture, that's the natural world expressing itself. When you use all of the data and information from all the movies and television shows and all the social media posts and anything that's ever been anywhere to train your models on and then you're going to create something and you haven't gotten permission for it and you haven't compensated for it, that's not okay. That feels criminal to me. So does that make

56:39Aline van der Velden a criminal? I don't think that's what Sean Astin means. By the way, when I interviewed van der Velden a few months ago, she suggested we also speak with her creation, Tilly Norwood. Will you interview her in September? Sure, I'd be happy to interview her. Yeah, yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah.

Preview of part two

57:02Coming up next time, in the second part of our series, we get Tilly on the line. Ah, então nem preciso me preocupar em te impressionar. Isso facilita bastante as coisas, não é? I understand you speak many languages. Unfortunately, whatever you said isn't a language that I do not understand. Could you try English? Oh, my apologies. It seems I'm speaking Portuguese for a moment there. We'll also hear from an economist who studies the creative industries. That is the nightmare scenario some people think about. And we ask, if maybe, just maybe,

57:33this story has a happy ending. My optimism is not because I think the problems are smaller than we thought, but because I think our capacity to solve them are larger than we think.

57:49That's next time on the show. Until then, take care of yourself. And if you can, someone else, too. Freakonomics Radio is produced by Renbud Radio. You can find our entire archive on any podcast app. It's also at Freakonomics.com, where we publish transcripts and show notes. This episode was produced by Pete Madden with help from Dalvin Abuaji and was edited by Ellen Frankman. It was mixed by Jake Loomis with help from Jeremy Johnston. The Freakonomics staff also includes

58:20Claire Keenan-Kurgan, Eleanor Osborne, Elsa Hernandez, Emma Kiesner, Tessa Troyt-Line, and Theo Jacobs. Our theme song is Mr. Fortune by The Hitchhikers, and our composer is Luis Guerra. As always, thank you for listening. Creatively, I'm having a great time, and I hope other people start seeing some of the fun and the positive in it. Because you know what? Like,

58:52ATI could kill us all in the next five years, so we might as well enjoy it while we're here. That's a great pitch for investors, I have to say. The Freakonomics Radio Network, the hidden side of everything.