Pod Insights The Vergecast AI can be your writing buddy, your blogger, or your Dungeon Master About this episode For the next four Tuesdays, Verge senior reporter Ashley Carman will explore how artificial intelligence and machine learning are shaping the future of a variety of industries.
In this episode, Ashley explores the wild world of AI writing and storytelling to find out if her job is in jeopardy. Is AI ready to replace seasoned writers or experienced reporters?
Guests include Nick Walton (co-founder and CEO of Latitude), Samanyou Garg (founder of Writesonic), and artist K Allado-McDowell.
This podcast was made by producer Liam James, senior audio director Andru Marino, senior reporter James Vincent, and senior reporter Ashley Carman.
Learn more about your ad choices. Visit podcastchoices.com/adchoices
Transcript 6,986 words · 10 speakers
Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Speaker 00 This episode is brought to you by On Investing, an original podcast from Charles Schwab. Each week, hosts Lizanne Saunders, Schwab's Chief Investment Strategist, and Colin Martin, Head of Fixed Income Research and Strategy for the Schwab Center for Financial Research, bring you fresh insights on what's happening in the markets and why, and what the implications might be for your portfolio. Join Colin and Lizanne as they explore questions like, how do you evaluate corporate bonds that look interesting? And what sectors are on the move right now? Download the latest episode and subscribe at schwab.com/oninvesting, or wherever you get your podcasts.
0:45 Nilay Patel? Hey, Vergecast listeners, it's Neil. For the next few Tuesdays in the Vergecast feed, we're running a mini-series we made about the different uses of artificial intelligence and machine learning all over the place, all hosted by Verge senior reporter Ashley Carmen, who is here now. Hey, Ashley.
0:59 Speaker 02 What's up?
1:00 Nilay Patel? It's episode three, the penultimate episode. This is like the Game of Thrones style big finish. What's going on in this one?
1:06 Speaker 02 So we've talked about audio, voice clones, we've talked about video. Now we are talking about text generation.
1:12 Nilay Patel? This is the one that's going to replace you and me and our jobs as writers.
1:16 Speaker 02 Exactly. Very near and dear to both of our hearts.
1:19 Nilay Patel? I don't think any robot can replace me, or you, or anyone. But what's going on in this episode?
1:24 Speaker 02 So we're talking about a technology called GPT-3, and some of the use cases for it, and how it might also be used in the future.
1:31 Nilay Patel? Give me an example.
1:32 Speaker 02 So like authors who are actually using AI to help them write a book, like publish a book that's co-written by both a human and AI.
1:40 Nilay Patel? Like a writing buddy, like a robot writing buddy?
1:42 Speaker 02 Yeah, kind of, kind of like a robot writing buddy.
1:45 Nilay Patel? This is my dream. All right, let's listen to it. Episode three of the Vergecast AI series. Roll the tape.
1:53 Speaker 02 We've always been interested in talking to computers, to sit down and have a conversation with the machines that do so much in our lives. Unlike our previous episodes on speech synthesis, this doesn't necessarily mean audio, but being able to chat with humans is a hallmark of artificial intelligence. The most famous test of AI capability, the Turing Test, is one in which a computer tries to convince someone that it too is a human based on nothing more complex than a conversation. Despite this, teaching computers to converse is extremely difficult. One of the earliest examples, a text-only AI therapist designed in the 1960s called Eliza, worked by asking the same basic questions over and over.
2:35 Speaker 03 Eliza is a computer program that anyone can converse with via the keyboard, and it'll reply on the screen. We've added human speech to make the conversation more clear.
2:46 Speaker 04 Men are all alike.
2:47 Speaker 03 In what way?
2:48 Speaker 04 They're always bugging us about something or other.
2:51 Speaker 03 Can you think of a specific example?
2:54 Speaker 02 In recent years, though, artificial intelligence has made huge improvements in our ability to generate text. You've likely already seen some of these new capabilities in Google's suggested email replies or iMessages autocomplete. But this is just the beginning, and not only is AI able to hold more convincing conversations, it's helping to write stories, articles, and even play games with us.
3:17 Speaker 05 I'm Nick Walton, co-founder and CEO of Latitude, and we're the creators of AI Dungeon.
3:23 Speaker 02 AI Dungeon shows what these new AI text generation systems are capable of. Think of it like an old-school choose-your-own-adventure book. But instead of picking between a small number of predetermined actions, you can do pretty much anything you want. Just type in what your character does next, and AI Dungeon will continue writing the story.
3:42 Speaker 05 It's dynamically made on the fly, adapting to whatever actions you choose. So you could meet a dragon, and instead of fighting it, decide to give it a cupcake and become best friends and open a bakery shop, or any number of infinite possibilities.
3:55 Speaker 02 AI Dungeon works in a number of settings, from classic swords and sorcery to detective stories and cyberpunk adventures. The system is not perfect at storytelling, though. It can't sustain longer, complicated plots, and will forget the name of characters. It's like an improv partner with a limited attention span. It works so long as you keep the story going in new directions.
4:17 Speaker 05 One was early on when I was playing and I met this band of goblins and I was like, "Oh no, they're gonna attack me." And so what I decided to do instead was tell them I'd teach them to play instruments. And so they let me come into their band and I taught them how to play instruments. And we eventually became this traveling musical group that we would tour around the kingdom at different festivals, playing music. And we eventually became like activists for human orc equality and like got invited to play for the king and like I eventually retired and like wrote books. And I was like, "This is something that would never happen in any other game, this kind of flexibility."
4:51 Speaker 02 This kind of text generation leads to adventures in the game that could not easily be replicated with your typical tabletop RPG. Nick shares with us a time where he was on a quest for a magical book in AI Dungeon.
5:03 Speaker 05 And then I started to read the magical book and like the AI could generate an entire magical book and system just through my play. Like I could read it and it was generating the pages of the magical book with like a very coherent magic system. And I was like, "This depth that it can go to and the extent to which it can make the world feel alive just kind of blew me away."
5:23 Speaker 02 Nick says he thinks programs like AI Dungeon will become much more common in the future, used by writers of all sorts to create story ideas and dialogue. And though we worry about software taking our jobs, he thinks this partnership will be mutually beneficial, a collaboration between human and machine.
5:40 Speaker 05 A script writer might be able to give a bunch of ideas for what a script might look like to an AI. The AI can be like, "Hey, what if we did this or what if we did that?" And the writer takes that and refines it and changes it. And so there's this collaborative partnership where they work together to make things that aren't really possible on their own.
5:58 Speaker 02 Where might we start to see this technology making an impact first? Maybe in video games where open world environments contain hundreds of individual characters with thousands of lines of dialogue reacting to the world around them. AI could help generate this script on the fly, fleshing out these worlds so they feel even more real.
6:15 Speaker 05 The amount of detail you have to write becomes so overwhelming that there's just a limit to how much humans can do that. But you can imagine having, you know, a few humans who are kind of the creative directors who say, "This is the world I want" and really are shaping that. But then it lets an AI fill in all the details and make the world feel alive and scale that content so there's a level of content that never really runs out essentially, that the player can continue to find and be new every time.
6:45 Speaker 02 What's powering these programs, though? And why have they gotten so good in recent years? The answer lies in a new sort of machine learning architecture that's particularly adept at understanding the relationship between words. These systems are known as large language models, and all the big tech companies are building them, including Google, Microsoft, and Facebook. They work by analyzing massive data sets of writing, which they study to discover what words tend to appear next to one another. Based on the strength of these relationships between words, they can then generate their own text. For example, if you're writing a list that starts, "I need to buy eggs and flour", then the next word is more likely to be "milk" than "bricks".
7:27 Speaker 02 Or if you begin a sentence with the quote, "a long time ago", then the words next are quite likely to be "in a galaxy far, far away". Here's Nick again, who uses a large language model called GPT-3, licensed from Research Lab OpenAI, to power AI engine.
7:44 Speaker 05 The best way I can describe it is it's like those word predictors on your phone. So your phone has these algorithms that looked at historically, you know, maybe things you typed or other people typed, and what word came next? And it suggests those words. So GPT-3 is just like that, except scaled up its intelligence 100,000 times. So then it can generate not just like the next word, but it can do that 500 times and have a whole paragraph story that actually makes sense and is interesting.
8:17 Speaker 02 It's worth stressing just how big these programs are and how quickly companies have been able to improve on them by simply adding more data and computing power. The size of these language models is measured in how many parameters they have or connections within their networks. GPT-2, which was released in 2019, had 1.5 billion parameters. But GPT-3, released the following year, has 175 billion parameters. That's over 100 times bigger. These programs are also trained on huge amounts of text. GPT-3, for example, learned the connections between words by combing through a pretty decent chunk of the entire internet. It's hard to give a totally accurate estimate of how big its training dataset was,
9:01 Speaker 02 but we know that all of English Wikipedia, spanning some 6 million articles, comprises only 0.6% of the total size.
9:09 Speaker 05 There's this really interesting thing where when you train it on so much data and with such large, smart models, it captures all these patterns that are actually really complex and interesting, and there's a lot of information in those.
9:26 Speaker 02 The problem, though, is when you train something on texts scraped from the internet, it picks up some nasty habits. The first of these is that large language models very much reflect the biases of our society. Because they've learned how to write in part based on toxic screeds online, they therefore regurgitate harmful stereotypes about gender, race, and religion. They spout racist and sexist nonsense with very little prompting. Some of this output can be filtered, like any hate speech online, but it's impossible to be certain that these biases have been eradicated on a more fundamental level.
10:01 Speaker 05 You have to ultimately follow a multi-pronged approach in doing this. The analogy I make is the human brain can generate lots of things you could say, but there's also a part of your brain that's like, "Mm, better not say that," right? Like, there's this filter component that takes the possibilities and decides what things are good to say or not. So I think systems like that are important as well.
10:22 Speaker 02 Another problem is that language models are often factually unreliable. They can answer basic trivia questions, but never with 100% certainty. That's because they have no way of distinguishing between fact and fiction. All they know is how probable it is that certain words appear together. A lot of the time, this is good enough to produce the right answer, but it's not something you necessarily trust if you were doing important research. It's also one reason why AI Dungeon is a good use case for these systems. No one cares about facts when you're writing fiction.
10:56 Speaker 02 What about the world outside of dragons and orc rock bands, though? What else can these models do? Well, many companies now offer AI language models as multipurpose writing tools that can tackle a number of tasks, from writing product descriptions to blog posts to student essays.
11:13 Speaker 06 You know, we have like quite a different set of target customers.
11:17 Speaker 02 That's Samanyou Garg, founder of WriteSonic, one of a number of new startups that offers this sort of tool.
11:23 Speaker 06 So just for the article writer, if we are talking about content writers and copywriters use it a lot. They just put in their topic and they get like an initial draft, which is like 60 to 70% there. And then they improve upon that, like add their own links or references and so on.
11:38 Speaker 02 Sam tells us that he relies on both GPT-3, as well as a number of homebrewed AI models to meet his customers' demands. One of the most popular use cases, he says, is for shopping sites, which are looking to produce a lot of written content quickly and cheaply.
11:53 Speaker 06 They basically have their own Shopify stores or Amazon stores, and a lot of them use the product description tool, because it really helps them generate product descriptions at scale. Because, for example, some of the companies that recently approached us, they have like 10 million product descriptions on their website. So it's not possible for a human to, you know, write that many and to maintain that brand quality slash tone. So that's where it kind of comes in handy. So we are building some custom models for them.
12:22 Speaker 02 Another service he offers is creating blog posts and articles for websites. It takes a little bit of input from the user, though. You can't just click a button and generate a perfect post right off the bat.
12:32 Speaker 06 So basically, our article writer consists of four different steps right now. First one is like you input your topic and it comes up with the actual titles and ideas for your potential blog post. So for example, if you put in, you know, productivity app, it would come up with, let's say, 10 productivity apps that will make your life, you know, amazing or something like that. And then we generate the intro based on what you provided in as a title. And then the third part is the outline, where it generates all the sections in the article. And then finally, it kind of writes all those individual components and then assembles it together into like a long article.
13:11 Speaker 02 To create these outputs, Sam says he had to fine tune his language models. That means finding training data that's similar to the output text. In other words, he shows the AI model the sort of writing it needs to produce itself. For the article writer, that meant he had to scrape a lot of data from news sites, including one that might be familiar to Virg listeners.
13:32 Speaker 06 We scraped tons of like top websites on the internet from all sorts of different industries. You know, obviously tech websites like virg.com and Medium, as well as, you know, other WordPress blogs, as well as, you know, other sites in different industry like fashion industry or lifestyle industry and so on.
13:51 Speaker 02 Are AI writing tools like this going to replace writers? Not directly, says Sam. He thinks they'll be more like collaborators used to jumpstart new ideas and create the bare bones for pieces of writing. We're already seeing the first steps of this process happen with tools like Google's Smart Reply and Smart Compose. But collaborating with AI writing tools can go so much further than just firing off some boring work emails. A number of writers have started using these tools to create new forms of artistic impression, hybrids between computer and human fiction.
14:27 Speaker 02 One of the most interesting of these is Pharmaco AI, a book written by the artist and writer Kay Allado McDowell and GPT-3.
14:35 Speaker 07 I definitely gained a new set of reference points and structures for thinking through that experience of writing Pharmaco AI.
14:43 Speaker 02 It's been described as a quote unquote hallucinatory journey into selfhood, ecology, and intelligence. But, as Kay tells us, it started life as a conversation between them and a computer, just like those first interactions with the therapy bot Eliza in the 1960s.
15:00 Speaker 07 So the process of writing was very conversational. But basically what I would do is I would write into a text field. I would write a prompt. Sometimes that would be several paragraphs. Sometimes it would be very short. And then I would generate some text from the prompt. I created some rules for myself. One was I would edit the output as it was coming out. And if I wasn't interested in what it was saying, I would cut that part and regenerate. So I kind of compared it to pruning a plant. And once I had decided to keep something, I would keep it. And I would add to that, but I would never go back and rewrite the beginning. Because I was really interested in seeing how the thoughts developed through the exchange
15:44 Speaker 07 and leaving a record of that back and forth.
15:47 Speaker 02 The book itself is not a novel. There's no plot or discernible characters, apart from Kay and the AI. Instead, it reads like a series of meditations on the nature of the self and creativity, and how we generate meaning from writing. The text takes on many different forms. Sometimes it's poetry, sometimes it's dialogue, and sometimes it reads like a revolutionary manifesto, a guide for the future of our relationship with computers. Kay says they actually wrote the book over the space of two weeks, while hoping to finish a completely different project.
16:19 Speaker 07 It was very much serendipitous and not something I had planned. So I ended up publishing a book, but not the one I had been working on for years and thought I was going to finish.
16:29 Speaker 02 In the physical version of the book, the text written by Kay and generated by AI are printed in different typefaces, allowing the two forms of writing to appear like dialogue. But as Kay explains, the project also blurs the boundaries between human and machine intelligence.
16:44 Speaker 07 One of the big concepts in the book is the idea of emergence, or being part of a larger ecosystem than just the individual human or the machine. So it was very hard to say sometimes where the ideas were coming from. I mean, as it is often with writing, but, you know, in this case, there were two actors. So you could sort of say, well, it came from this person, it came from the machine. But actually, it was a little bit less clear than that. And especially when I would start to have dreams about the imagery, you know, I was like, yes, this is coming from me, but it's coming from a part of my mind that isn't fully conscious, or, you know, it's coming from a part of my mind that processes information differently.
17:24 Speaker 02 Is this sort of interaction with machines something that we'll all experience in the future? Having conversations with AI and not knowing whose ideas are whose? Kay says it's possible, but points out that the experience of using AI writing tools in the workplace is very different to the sort of self-reflective writing process they went through to create Pharmaco AI. Ultimately, though, they say the emergence of these tools raises new questions about how we think about our own intelligence.
17:51 Speaker 07 This question of is the system extending me? Am I extending it? Where is the agency? How do I even determine agency in a human-to-human relational scenario? Or a human-to-animal or human-to-plant relational scenario? It starts to push on our preconceived notions of the bounded self. I've always thought that treating AI as a mirror is the most clear way to make the point that these statistical systems, these predictive systems are going to give us back ourselves at the individual level. So whatever we put in is what we're going to get back out.
18:32 Speaker 02 Kay's experience using GPT-3 to write a novel seems like a pretty extreme and unusual interaction with these AI systems. But is this sort of work going to become more common in the future? How are text generation tools going to be used in everyday life? And where do you draw the line between your ideas and the computers? We're going to take a break, but when we're back, you've heard from him before, James Vincent, senior reporter at The Verge, who specializes in machine learning and AI, will be here and will answer all the big questions we have, like whether machines will replace me and my reporting in the future. We'll be right back.
19:19 Speaker 00 This episode is brought to you by On Investing, an original podcast from Charles Schwab. Each week, hosts Lizanne Saunders, Schwab's Chief Investment Strategist, and Colin Martin, Head of Fixed Income Research and Strategy for the Schwab Center for Financial Research, bring you fresh insights on what's happening in the markets and why, and what the implications might be for your portfolio. Join Colin and Lizanne as they explore questions like, how do you evaluate corporate bonds that look interesting? And what sectors are on the move right now? Download the latest episode and subscribe at schwab.com/oninvesting or wherever you get your podcasts.
20:12 Speaker 02 Hello, we're back and James Vincent is here. Hello, James.
20:16 Speaker 08 Hi, Ashley. Pleasure to be back again, talking more AI, machine learning. I love it.
20:21 Speaker 02 We're talking about text generation and as I always say, I'm here to have you level out our thinking here. So, when you're thinking about text generation and AI and all of that industry, what excites you about the field? What are you skeptical of? Give us the raw take, as I always say.
20:40 Speaker 08 It is definitely an exciting field. It's one that's been moving very quickly and like the voice synthesis has made like serious tangible improvements over the last couple of years. And like the text models that we have now compared to the ones we have had a decade ago are just so much better and people didn't really imagine the improvement would be this quick. But as the improvement has sort of sunk in over the last couple of years, we're also finding more about the limitations of these systems, both in terms of, you know, you mentioned earlier some of the biases they have, the fact that they're not factually accurate or they can't, it's difficult to verify that they're factually accurate. But there's also this idea of meaning. It's a tricky thing. It gets a lot of people bogged down. It gets me very bogged down.
21:23 Speaker 08 It's like, do these systems understand language? And does it matter if they do or if they don't? And it's sort of a question no one has any answer to. It's like, as long as they sound plausible in what they're coming up. Up with, does it matter if they understand what they're saying? And then some people say it does hugely matter because it means they don't have a model of the world like humans do. Um. So they don't know when they're saying something that doesn't make sense. So I feel that's something that we're just beginning to explore and that will be as these systems appear more in practical consumer, commercial applications. We're going to bump up against more of those limitations and we're going to, we're going to see more problems with these systems, I think.
22:03 Speaker 02 Yeah, it's interesting. I mean, we're skipping ahead a little bit here, but it's top of mind as a person whose job is mostly writing.
22:10 Speaker 08 Yeah.
22:11 Speaker 02 But we talk a little bit about news articles and how it could be used to create blog posts and things like that. And I do wonder if them not being able to be factually accurate does affect whether you could use it for sort of these like news posts and things like that. Just because I feel like even if it is sort of like routine to write a news post, you still have the reporter being like, I know this to be true and I'm going to actually double check it and whatever else. So I do wonder if actually this could even replace news writers, really.
22:39 Speaker 08 I feel like it really depends on what sort of news you're talking about. I've played around with a bunch of these different systems and, and see what it comes out with. It comes out with garbage, comes out with like something that just doesn't work. I mean, there are huge limitations. These systems can't research. They can't sort of Google the stuff for you and synthesize all these sources. They can't do new information. They are, they're, you know, they're trained on old data, so they fundamentally don't have this new stuff that you want. They can though, do rote writing, really sort of buy the numbers stuff. And where we've seen newsrooms adopt AI writing tools, not necessarily GPT three and its ilk, but sort of even earlier writing tools.
23:20 Speaker 08 They've used them for stuff like, there was a tool that did automated earthquake announcements. So it was hooked up to uh some sort of meteorological data set. And when that registered an earthquake of a certain magnitude in a location, it wrote that up and it just sort of like, it had a form basically, and it punched in the numbers and it punched in the location and it published it. And so there are things like that where AI can definitely write some basic formulaic stuff for say movements in the stock market, for sports results, for weather. But when it comes to writing anything new, it can't do it. It could do it with a little bit of help, but I think at this moment, it's not worth the extra hassle it takes for a writer. So I, Ashley, your job, absolutely safe.
24:05 Speaker 02 Oh, thank God. That was the biggest question I had. Because we've really tried to replace me like in every single episode here, so.
24:13 Speaker 08 It's been a running theme.
24:15 Speaker 02 Yeah, I'm feeling confident coming out of this that like I will still have a job, hopefully.
24:19 Speaker 08 Absolutely.
24:19 Speaker 02 So you mentioned that you've tried to use this on your own, these text generation tools. Can you give some examples of instances either where it did work really well and you're like, this is amazing? Or if you have instances where you're like, you kind of mentioned like the news writing where you're like, nope, this isn't going to work. But others, if you have them.
24:35 Speaker 08 Well, it's like you said earlier on about, you know, the AI dungeon being a good fit for it. These tools are surprisingly good at being creative. Or, you know, is it really creativity of its machines? Who knows? I think it is because I think they come up with stuff which is interesting and funny. So one of the projects I was playing around with when I had access to GPT-3 was for The Verge's 10-year anniversary. You know, I was, we're doing a zine and I was coming up with some stuff for it, which includes a techno pagan spell, which I don't know if you remember, do you remember the techno pagan from Buffy the Vampire Slayer?
25:10 Speaker 02 Unfortunately not.
25:11 Speaker 08 There was, I can't remember her full name, Jenny something, who was this techno pagan witch, who was like, what if a witches were online and on the internet and blogging up a storm? So I thought I'd make some techno pagan spells and I fed some stuff into GPT-3. And it came out with some like, some really like weird and funny stuff. Do you want to hear a techno pagan spell written by AI?
25:34 Speaker 02 Yes.
25:35 Speaker 08 Okay, so the first bit of the spell is what I gave it and I'll tell you when I stop and when it becomes the thing. So this is what I fed it. I said, techno paganism is a new sort of magic that blends traditional witchcraft with the power of being online all the time. What follows is a techno pagan spell designed to protect your phone from spam and improve the quality of messages you receive. Step one, and that was when my prompt ends. So you kind of set it up and you're like, right, now you've got to fill in the blanks. So this is the spell it gave me. Take your cellphone and go outside. Close your eyes, hold the device in front of you and intone, "I offer you, oh spirit of techno-pagan magic, my eternal and unending electronic connection to the net. In return, I ask you for magic to combine my will with the will of the system.
26:19 Speaker 08 Make me impervious to the garbage that flows through my inbox. Make all messages flow together and I will choose among them what is right for me to see. Grant me the finger of the internet and I will strike at spam and cast out the spammers back into the dark places from whence they came." So that was what it wrote and I was just like, "That's amazing. I love that." Where did that come from?
26:39 Speaker 02 That's really good. I guess I have a question, which is like, like what's confusing to me about how these systems work is, so obviously when you're talking about spam, I can imagine, and you specifically said phone spam. So I can imagine that this AI has a reference point that is like, people have written about how phone spam is garbage, like, "Give me my clean inbox."
26:59 Speaker 08 Yes.
27:00 Speaker 02 Whatever they're gonna say. And I could see where it pulls those words from. But then like, what's confusing to me is how it then also infuses kind of this idea of like, the techno-paganism and magic. Like it somehow fuses the two. So it's like, interesting to me that it's able to pull from different systems or different pieces of writing, I guess, and like, know that it should go together.
27:18 Speaker 08 No, absolutely. This is the wonder of these systems, that like, all they know is the relationships between words, but they know it on such a multi-dimensional level, like they work in this multi-dimensional space to find the relationships between them and from that, they somehow manage to come up with this stuff. Um, so the thing I will add is that that wasn't the first spell that it wrote for me. And it did write some stuff that was kind of useless and not as funny. And I had to tinker around with it. There are all these different settings you can adjust and the most important one is called temperature. And it's how weird you want the response to be. So if you have a very low temperature, it will sort of follow it, you know, as boringly as it can. And it sometimes gets caught in loops and just ends up repeating the same thing over and over.
28:01 Speaker 08 And you increase the temperature and it gets more adventurous, more bold in what it's doing. So this was quite a high temperature, but it was, I think it was probably my third or fourth time getting it out of the system. So these things aren't perfect. There is human curation and there is human selection needed to say, "This is a good answer. This is a bad answer." It's not true of all things, but um, yeah, no, I'm the same as you. I'm like, I just enjoyed playing around with these tools. I just think they're so fun and like, I really enjoyed, for example, Pharmaco AI. I just thought... Such a brilliant book and so, so enjoyable and weird. I wanted to do, if it's okay with you, Ashley, a little test. I was, I was gonna read off a couple of things from the book and I want you to tell me whether they were written by a human or a machine. Do you fancy that?
28:46 Speaker 02 Yeah, yeah, that sounds good.
28:47 Speaker 08 Okay, I'm gonna give you a little bit. So this is a poem and was it written by a human or a machine? My grandfather was a machine. His bones and muscles were gears and levers, his flesh was plastic. His eyes were video screens and he had radars for ears and wings that he would use to fly. He would count to a million and he was a math genius and he knew how to make you feel smart. He was born in a factory in Detroit and his DNA was blueprints for making cars. He had two names, one for my father and one for my mother. Do you think that was a machine or a human who wrote that?
29:20 Speaker 02 I'm gonna, ah, I'm gonna go with human.
29:23 Speaker 08 That was written by the machine.
29:25 Speaker 02 Dang it.
29:26 Speaker 08 A machine wrote a poem about being a machine.
29:28 Speaker 02 It's really spot on.
29:30 Speaker 08 Yeah.
29:30 Speaker 02 I was trying to listen if there was words that maybe would make sense why they would be next to each other, but like if you were just having a conversation or writing, you maybe wouldn't necessarily use next to each other, but I couldn't really pick one out that seemed odd, like odd word choice.
29:44 Speaker 08 Yeah, I think it's pretty seamless. Okay, here's another one. Language is not only an act of communication. Language is the creation of experience. Language is a fractal expression of life, as life or existence is the creation of time and the accumulation of relations through time. Semiosis is the constant interaction of internal and external words and an internal world is created in the external world as an external world is created in the internal world. So do you think a machine or a human wrote that?
30:15 Speaker 02 I'm gonna go with human again.
30:18 Speaker 08 A machine again, that's another machine, yeah.
30:21 Speaker 02 That sounded more like too philosophical.
30:24 Speaker 08 Mm-hmm.
30:24 Speaker 02 Like it was just too many questions. I was like, the human must be asking these questions.
30:29 Speaker 08 Okay, let me, let me do you one last one. The Greek god Hermes counterpart to the Roman mercury was the god of translators and interpreters. A deity that rules communication is an incorporeal linguistic power. A modern conception might read that it is a force of language from outside materiality. Automated writing systems like neural net language models relate to geometry, translation, abstract mathematics, interpretation and speech. And it's easy to imagine many applications of these technologies for trade, music and divination.
31:00 Speaker 02 I mean, I should guess human again, just because I feel like I've guessed human for everyone. So I have to eventually get it right, but I honestly would guess AI for this one.
31:08 Speaker 08 Okay, that was human, that was human.
31:09 Speaker 02 I figured, but I'm like.
31:11 Speaker 08 You played the game right and you got it, you got it that way.
31:14 Speaker 02 But it's interesting, like this is a place where I would say at least it's more difficult to tell. This is definitely trickier zone to identify.
31:21 Speaker 08 Yeah, and it's, you know, it's just none of these technologies is a simple, is it, you know, a simulacrum of humans or is it not? Or each of them has a specific way that the information is delivered. And so this is trickier to tell, but it's trickier to tell because humans have been involved in the process. Whereas the synthetic speech is easier to tell, but at the same time, it's also something that can be done without human input. So I think, I think there's weird ways to balance this stuff about how smart they are. Cause it really depends on the context.
31:49 Speaker 02 Yeah, that's a good point.
31:50 Speaker 08 Do you feel that these tools would be useful for you? No, you know, if you had something that was like, here is my story headline and it gives you sort of a brief outline and then you fill in the actual writing, would you want that if it could do that?
32:04 Speaker 02 I feel like in my day to day, no, like for my specific type of writing, I don't necessarily know if the app. Because as you kind of mentioned, and this is getting a little niche, but like just as a reporter, I feel like we're constantly updating our databases of information, like companies get acquired, new things happen. There's a lot of context that maybe just occurred two days before, or you're writing to news that just happened and you're doing a take on it or something. So I feel like my internal data set is like much more current and I need it to accurately do my job.
32:35 Speaker 08 Yeah.
32:35 Speaker 02 But I do find it potentially interesting for more of creative endeavors. Yeah, but I do wonder if like. Authors could use this for ideas and then just use them as their own and not disclose that an AI helped them come up with that, or if you even need to disclose that an AI helped you and inspired you or even gave you outright the idea for something.
32:54 Speaker 08 There have been a couple of tricks that people have done like this. There was some sort of essay written by GPT-3 that got onto Hacker News and got a lot of attention and interest and, you know, was upvoted and was presented as being written by a human. And some people were fooled by it and some people weren't. The question again is how much curation went into making that before it went in there. But I genuinely think that there will be people using these tools quietly in the background and not saying anything about it. And I think it's something that is going to be a general trend we're going to see with all these synthetic tools, is that there's probably going to become a new layer of prestige in the media when you say, in the future, in the future this is, when you say something is a human-only product. You know, in the same way that you you buy
33:38 Speaker 08 eggs that are free range, you would buy content that is human-only in its production.
33:43 Speaker 02 Oh my god.
33:45 Speaker 08 I think it's going to start creating these tiers. Will the human-only products necessarily be better? Who knows? I think that will be very circumstantial, but I think they'll be sold to us as something authentic, you know?
33:56 Speaker 02 I mean, yeah, that's such an interesting point because I'm watching Game of Thrones right now. And but watching it, I'm like, wow, this is just so incredible that one man envisioned this entire world. And like, it's just mind blowing. Or I think of on a much smaller scale, like Bo Burnham's Netflix special he did recently, Inside, where I was just like, my god, like, this man is so creative and has come up with so many incredible, like, it's just such a feat, you know, you're just blown away by it. And so if I had heard, oh, George R.R. Martin used AI here to help inspire him, or Bo Burnham had AI help him kind of generate some ideas. I think that would taint it for me.
34:38 Speaker 02 It would. I'd be like, oh, I mean, this is still impressive. Like, sure, but it wasn't just you. You had to rely on like a past database of massive information to like generate ideas. So that is a really interesting point. And I just feel like there will be shady stuff going on where people won't disclose.
34:56 Speaker 08 But I mean, so let's compare this to something like Andy Warhol. If you bought an Andy Warhol painting, would you mind that it was created by one of his assistants? And he just came up with the design and then he threw it into, you know, famously used all these assistants to make his paintings. And they were just sort of blueprints that he copied and pasted, as it were. It was part of his comment on mass production and on commercialization. So actually, he sort of embraced that, made it part of the appeal. And so got over any qualms that people might have about it. And maybe this will be something that artists working with AI are doing now. And I mean, I think it's something that you do already see now with AI artists, that they make the human machine collaboration part of what makes their art special.
35:37 Speaker 02 Yeah.
35:38 Speaker 08 And it is special because I think we're at this really weird fringe point where these worlds are colliding and it's interesting stuff is happening.
35:45 Speaker 02 Well, I think this is a great place to stop because now I have a lot to think about. Thank you so much, James, for coming on. As always, I appreciate your deep insight and wealth of knowledge.
35:56 Speaker 08 Thank you, Ashley.
36:03 Speaker 02 All righty, thanks so much for listening to this Verge Cast AI mini series. This show was made by producer Liam James, senior audio director Andru Marino, senior reporter James Vincent, and me, senior reporter Ashley Carmen. Talk soon.
36:23 Speaker 09 Your heart can tell you a lot about your health. Apple Watch Series 12 measures your heart rate every five seconds with the most accurate heart rate sensing in wearable. So your vitals app, now with heart rate variability, can tell you when something is off. And your readiness score can let you know when to rest and when to push. Hear the story in every heartbeat with Apple Watch Series 12. The features described are for wellness purposes only and not for medical use. iPhone 11 or later required. Based on Apple conducted study of heart rate accuracy, August 2026. Visit apple.com slash apple watch series 12.
Previous iPhone 13 review / Microsoft’s Surface event biggest announcements Next Amazon’s fall hardware event / Google’s Search On / Code Conference 2021 Page updated October 26, 2026 · transcript and summary are machine-generated
