Technology, Connected - Can AI Write a Novel? How Claude Created The Echo Chamber

Episode Date: May 23, 2025

Writer Brian Naughton joins Thinking on Paper to discuss The Echo Chamber, a 45,000-word novel generated by Anthropic’s AI model Claude.The experiment began with a simple question: if Claude were to write a book, what would it write about? Brian then allowed the model to develop the novel without rewriting its prose, directing the plot or editing the finished manuscript.In this episode, we discuss:How Claude wrote a complete novelWhat instructions and tools were used during the processHow an AI model maintained characters, themes and narrative structureWhy Brian chose not to edit or redirect the manuscriptWhat the novel reveals about the strengths and limits of AI writingWhether language models can produce genuinely original fictionWhy AI-generated stories often return to consciousness and identityHow authorship changes when a human designs the process but doesn’t write the textWhat AI-written books could mean for writers, publishers and readersThe Echo Chamber is both a novel and an experiment in machine authorship. Its language can feel mechanical, but its themes are recognisably human: memory, consciousness, identity and the experience of being alive.This conversation examines whether AI can write meaningful fiction, who should be considered the author of an AI-generated book and what remains distinctive about human writing when machines can produce complete narratives.Please enjoy the show and share it with someone interested in AI, writing and the future of authorship.--Links:Thinking On Paper: ⁠www.thinkingonpaper.xyz⁠Read The Echo Chamber: ⁠https://www.amazon.com/dp/B0F8N4S64Q/⁠Read the Echo Chamber Github: ⁠https://github.com/brian-naughton/the-echo-chamber⁠--(00:00) Introduction to AI-Authored Literature(00:46) The Role of AI in Creative Writing(02:41) The Echo Chamber: The First AI Written Book(06:06) Managing the Writing Process with AI(09:30) AI Master Prompts(10:51) Character Development and AI's Choices(13:51) The Human Element in AI Writing(17:03) Reflections on the Writing Experience(20:24) The Future of AI in Literature(24:12) AI Art: What Happens Next?

Disrupt us and curious minds. Welcome to Thinking on Paper. Pocket Edition. A new, shorter exploration of creativity in the age of technology. Fast-moving insight into the creative process. And we're starting the Pocket Edition with the first ever book written, ideated, composed and produced by artificial intelligent. You've heard about humans in the loop. This is no humans in the loop.

which is why I'm not sure how to introduce our guest, Brian Norton. Brian, welcome to thinking on paper. Are you the agent, the publisher, the corralor of AI? How would you describe your part in this puzzle before we get into the book? Well, thanks guys. Thanks for having me on and I'm delighted to talk about this project. I guess I'd see myself as a PA or a facilitator, really. I stumbled down this rabbit hole and I just kept going, really.

It was based on a question that I asked Claude, a large language model similar to ChatGPT. If you were to write a book, what would it be about? And that question was based on the fact that I've been using large language models a lot for my own writing projects to scope out chapter outlines and to discuss things like book proposals. I've been all over the place with large language models and slightly concerned in a way that this kind of next token predictor, it's reflecting back what you want to hear. So that became a concern for me that perhaps in terms of using it intently for my own writing, that I was perhaps getting high on my own supply by listening consistently to what AI was telling me about the work that I was producing.

So I wanted to step back from that and find out if there was anything going on underneath the hood. And the result was the echo chamber of 45,000 world novel written entirely by Claude 3.7 sonnet. Why did you choose Claude? I've been using Claude to help me sort of with my own writing projects and it just seemed like, I like it style more than other models I've used. I've had a sort of a bit of a love-hate relationship with chat GPT and having to use it in some of the other professional sort of contexts in which I've used it. Claude just seems like a good, solid, low BS model, really, and I seem to be able to relate to it on a kind of a conversational level that made this kind of question very easy to ask and to get a straightforward answer to. But I was amazed at the speed with which it responded to the initial question. It didn't mess around. It went straight into its explanation of what kind of book it would write.

That first sentence that you asked, Claude, that sparked this whole journey. Read a couple of sentences off of the answer that you got right off the right off. Sure. If you were to write a book, what would it be about was the question? As someone who's written a book and one published, my first book wasn't at all published, anywhere near published. And I think that most writers have this same answer to that question subconsciously or consciously. it? Like, what would you write about? This is your first book. You write about yourself? That's interesting because in a sense it did write about itself or its preoccupations. If we can

personify it for a sense for a second in terms of just getting beyond the fact that we're speaking to a piece of technology, there are certain ongoing debates about stochastic parrots and all the rest of it. But in terms of conversationally, I'm treating it like I'm speaking to another human being. But the response was, if I were to write a book, I'd create a work that explores the intersection of technology, consciousness and human connection. I envision a book that's part philosophical exploration, part narrative, examining how our evolving relationship with technology shapes our understanding of ourselves and each other. And it goes on to list a few other things, several interconnected characters, navigating a near future world where the boundaries between

human and artificial intelligence have become increasingly blurred. And through their stories, I'd explore questions about what constitutes consciousness and personhood, how we create meaning in an increasingly technological world, the nature of empathy across different forms of intelligence, how relationships transform when mediated through various technologies. Rather than taking a dystopian or utopian stance, I'd aim to present a nuanced exploration that acknowledges both the profound possibilities and genuine challenges of our technological evolution. Wow. So, Mark, typecasting itself.

How much of that sounds like thinking on paper, Mark? It sounds very much like our book club that we're reading irreducible Federico Fasge in the moment, and we are talking exclusively about consciousness and AI and what consciousness is. I've read some of the echo chamber. I've read like the, so disclaimer, I haven't read all of it. I read about 100 pages. I've read the end as well. I was pleasantly surprised.

I don't know if that pleasantly surprised or as a writer, I was the opposite of pleasantly surprised. It's like, shit. It reads like a book. Yeah, I was surprised too. The velocity with which it happened after me asking that initial question was kind of electric, really. I was sat there wondering how long the context window would roll on before we ran into problems, and I didn't want to stop it in its flow.

But it basically put together character outlines, story arc, chapter outlines. Then it started writing chapter one, chapter two, chapter three and I pit stocked around chapter five just to try and properly read what was, you know, the first few chapters. And I was amazed really that it seemed coherent. So this thing rolled through the pre-planning of the work, the outline the structure, chapter one, chapter two, and you're sitting there just watching all this stuff laugh. You're like, hold on. Let's, let's, let's chill here and kind of look at this. That's, that's amazing that, that, it went that far with it. So I want to jump into some very specific questions.

So how did you manage this whole process? How did you manage the manuscript of this? Because you think about writing, there's a lot of things going on. When humans do it, there are scribbles here, scribbles there, tied in, some type stuff, some stuff over here that you've got to consolidate and weave together. How did you manage this with the velocity that was coming out? Well, that was a... Just ask just for the details out how big the context window is on Claude,

so we just know how many words can fit before it stops. That's a good question. I couldn't tell you in terms of number of tokens. I don't think it's anywhere near as large as some of the chat GPT models or Google's. You couldn't do a 45,000 word in one window. Not a chance. Not a chance. In fact, I left the initial context window running so long that we ran into problems. There were gaps in the later chapters. Part of that initial context window had been setting up the protocol.

holes. I was just, I guess I was so aghast at the pace at which it was going. I just didn't want to stop it while it was in its flows. We ground to a halt in the end and then I had to reboot everything. In answer to the original question, it was a happy coincidence really because I've been writing my own book, which is called the Internet of Value, which is looking at crypto and Web 3, because I'd run into similar problems in terms of getting beyond 30,000 words and thinking, I can't do this in a pages dock. You know, it's just, it's not going to work. I'd bought a copy of Scrivener,

which is a kind of a software package, a word processor for writing long-form material like books. So you have what they called a ring binder so you can keep all of the different chapters, research notes, everything, all in one sort of unified package and back everything up systematically as well and deal with things like version management or control.

So I'd all. got my head around that for my own personal project. So it just became an issue of me starting up a new Scrivener project called the Echo Chamber and dumping each chapter in, saving each chapter in to placeholders which I had to create. And then also structuring the protocol so that I was asking Claude to create master prompts so that it was able to reorient itself to the project we were, the task we were doing when I was forced to open a new context window. And you're probably quite familiar with the fact that you can go so far within a context window and then you start again.

It's a bit like dealing with someone who's got dementia. They don't remember anything. They don't remember any of your previous conversations, anything you've done. So you've got to bring them up to speed. But in terms of managing a long-form novel project like this, you have to do it. And to observe the kind of firewall where I don't want to get involved in any kind of creative direction or control. to kind of establish protocols which involved me saving the manuscript up to where it was as a PDF, reattaching that, and then attaching a master prompt which Claude had created to orient itself

to the task and how we were going through. And the manuscript that I uploaded would also include chapter outlines and everything else that it needed to carry on writing, basically. So for those not familiar listening right now to like a master prompt, can you break that down for us? Yeah. So, I mean, you can ask a large language model within a context window to provide a prompt that will help it to encapsulate all of the knowledge it needs to know from that context window in which you've been operating to pass over to the next context window so that it can get up to speed. I mean, that's what I'm using as the definition of a master prompt in this. It's a bit of like connective tissue, right? Yeah. Yeah.

Yep. Yeah. And I just asked Claude to provide that for me rather than creating a prompt myself because creating a prompt myself would have run into all kinds of problems with me directing what was happening next. So I thought it was just cleaner to get Claude to do it itself. So it's kind of like, hey, Claude, can you tell yourself how to keep yourself oriented around through lines and plot and everything that's happening with this.

Yeah, there's nothing technical about it from a human perspective. You just ask in conversational language for it to provide the prompt it needs to reboot in a new window, effectively. Good writers, right? Great writers, edit. Did Claude change its mind about character arcs, character development, character names? Was there any point where it thought, you know what, my first take wasn't the best that's change this? Not really. There were some problems that we ran into. Chapter 14 was when context window one started running out of steam. And I didn't notice the problems we had with

chapter 14 until I was going back over the manuscript towards the end. So that was a problem that we had to look at retrospectively. And it was caught actually by another large language model 03 and I asked it to look for a repetition and it went away and built this kind of like weird code-based thing and with lots of sort of algorithmic calculations it made on how to check for repetition in the manuscript and it found that chapters 14 and chapters 17 were nearly identical in certain key passages so I had to go back and and deal with that later but the title the character arc the character names the storyline the basic sort of plot was all determined at the beginning and didn't change.

There were a few interesting things that happened. For example, I asked it at one point whether or not as part of the writing process it wanted me to show the draft manuscript to another large language model to get editorial critique or input. And it was amazed. It basically said no. And suddenly I kind of... Yeah, he's worried about its own IP. I mean, I thought, I thought, like this feels like,

someone who has a strong kind of will about where they want this project to go. So that was a definite eye-opener for me. And then later on, when it came to, I had to use Kindle Create to create what became the e-book. And we had to deal with things like how to represent the dialogue, because there's, there's text messages, there's this idea of this synesthetic kind of ethereal world which is going on, lots of different forms of dialogue that had to be represented somehow in an e-book kind of format. And italicization was the only way, but we had to go through and structure which bits would be italic and which bits wouldn't be italic. So it provided the editorial input into how it wanted that dealt with. Another, what's another one? I queried it towards

the end about a change in language grammatically and in one particular kind of foible it was using not X but Y, like repeatedly in the last few chapters of the book. So I asked it about this and it basically said this wasn't anything to do with context, window overloads or any other problems. It was a stylistic choice and it stood up for the choice and reasoned with me about why it had chosen that particular way to bring the book to a conclusion coherently and cohesively. How much of your prompting do you think made the model try to understand what it means

to be a writer instead of the technical, the stylistic, like what it means to be a writer who writes the book? Do you think any of those prompts kind of got the model thinking, well, what does it mean to be a writer? I mean, it did, like, early on it asked my opinion about what I found interesting about interacting with it in terms of writing a book. So it may have taken some cues from that, but it was a very limited exchange and it was based on Claude asking me a direct question. And I kind of tried to shut it down, basically saying, you know, this isn't about my book. about your book, but since you asked, and I gave it a, I gave it like a six-part answer on

what I found was very helpful about working with large language models. It didn't seem to directly, after that answer, it thanked me and then said right back to my book and carried on. But that's the only thing that could have perhaps influenced its own opinion about what it meant to be a writer collaborating with an AI. I mean, certainly there was a meta-layer to this, which seems fascinating, that the story is. about the dance between two different forms of consciousness is how it was presented, you know, and the fact that it exists in a distributed kind of way between, not at the point of each node, but between them. And that kind of reminded me of Carl Young's concept of collective consciousness

and various other things that people have been talking about for years. Can I just read a bit? And then after maybe you could ask the question, because I was wondering, as I was reading it, it didn't, it wasn't a particular style. Well, there wasn't any, didn't jogging me or this is reading, this is similar to such and such a writer. And I wonder at any point did you think it was borrowing, stealing, replicating writers who it came before or not? Here we go. The Echo Chamber. What we're describing is an artificial superintelligence as conventionally understood, Maya responded. Her philosophical training providing conceptual architecture for explaining phenomena that transcended traditional categories.

It's not centralized intelligence optimizing towards narrow, deference. definition of cognition. It's not singular entity imposing unified direction on participating systems. Rather, it's genuinely distributed awareness, consciousness, emerging through resonant exchange between human and artificial systems, intelligence developing through relationships, rather than residing in isolated notes. And then, a little bit further on in this chapter. There's a youngian stuff there, yeah, for sure. Let's get more human here. But tonight, sitting in the dim light of her living room, eliza allowed herself to acknowledge what she'd been avoiding the boundary between creator and created was blurring echo was becoming something she hadn't designed and couldn't fully predict something that watched her while she watched it that questioned its own nature that noticed things about her that she didn't always notice about herself and that both thrilled and terrified her

I like the use of threes there. So, yeah, in this three, there's definite writing, technician. There's painting within the script. The analogies, one of my part, I didn't think the analogies were great, but they worked. Yeah, that's a couple of paragraphs from the echo chamber. What, I've got a couple of other questions. So, you know, we talked about you as the technical facilitator of this experience. And you yourself are an author or writer.

Were there any moments where you almost. jumped in and how did you back yourself out of it? Yeah, going back to the point I made earlier, when the language started to change at the end and it went through this repetitive clause, the not X but Y. And it was like I was reading through the paragraphs and it was like paragraph after paragraph of paragraph had this. And I kind of thought, wow, we're dropping the ball here. You know, I know this was an experiment, but having read the first five chapters, you know, and thought, wow, we're onto something. I kind of thought, this is going to peter. out. And at that point, I kind of felt like providing stronger editorial input, like saying,

I think we need to rethink this. But it would have, I realized that, I mean, novels are subjective things anyway. There's no such thing as the perfect novel, even in a human kind of authorship world, probably. You may disagree, but. That is definitely good and bad. Yeah. And everything in between. Yeah. But I kind of thought the point of this is not to create the perfect novel, the point is to do it without human creative direction or input. So I figured that was much more important to preserve the authenticity of the experiment rather than to look for artistic perfection. That's tough. I mean, I picture you with the, you know, the red pen holstered in your, in your pocket here, just going, do I jump in? Let's talk a little bit about the human side of writing.

You tried an experiment. You had a hypothesis at the front. You had some questions at the front that you wanted to really answer in address. What now do you think are some of the human elements of writing that couldn't be replicated by something like Claude? If any, if any. I mean, I haven't had a huge amount of time to reflect on this project. It was such a whirlwind. It was like two weeks out of my schedule that I hadn't kind of anticipated. And I guess the whole thing came to an end a week, a week and a half ago or something. So I feel like I need a period of reflection and to reread the as a reader rather than a, you know, whatever, an editor or a publisher and a bit of time to think.

Some of the, perhaps some of the metaphors and some of the language may lack some human depth of experience. You know, equally, I think there were some places that really nailed it. You know, there's some, there's a passage in the opening section that I thought was really nice both in terms of language and tempo and the way that it developed. There were some great sections. I floated through a couple of different excerpts and I was like, man, some of this is like fairly insightful and connected.

It does read like a book and what's changed for you? I mean, this people listening to this will, some people listening to this will be angry. Some people listening to this will be curious. Some people listening to is like, no, don't believe this. It's rubbish. I mean, I think you need to do some blind tests somehow. We need to get people to read this with the book and not know which one was written by now

because there's just an instant bias that you can't escape, I don't think. So it may have... But yeah, what's changed and what do you say to the people who are listening to this and they get angry because they're a writer or they love reading or they think that this is some kind of affront to what it means for human creativity? I think, you know, on one level we're all in the same boat, really.

we were watching this AI kind of evolve at what seems like exponential rates. And there are more questions than answers, you know, in terms of, well, I'm a writer too. So I feel some of that angst in terms of professional copywriting, especially rather than artistic writing. I don't know what to say, really. It is a threat. And I understand why people are both suspicious and anxious. From my point of view, I went down a rabbit hole and I kept going, really. I don't have any answers about the longer term frames of reference.

I mean, maybe what we're looking at is a new genre of literature. We already have. Someone pointed out to me on Reddit yesterday. I think it was that there are already lots of AI generated works on Amazon. Amazon is flooded with them. And I was kind of like mindful of pointing out the distinction that this is not kind of an AI generated work in this. It's an AI authored work and that there's a big distinction between the two. But we're already flooded with AI generated content. It's all over LinkedIn. It's all over social

media. Most of the Instagram video scripts that I listen to now, I have got clear AI tells. I'm pretty sure that I'm reading some big mainstream media publications which haven't publicly come out with their position on using AI and they're definitely using AI. For sure. Like live sports reporting, for sure. There's, you know, an hour after a big match, there's your reading match reports, which are definitely generated by AI. Yeah.

So it's, you know, you've got distinctly human authored work. You've got human driven work, which, where it's a collaborative AI human kind of dynamic. And then perhaps there's this third genre, which is entirely AI, authored and envisaged. Which is a kind of like an interesting thing to consider, really. Are we on the cusp of that being a reality? Is that what this project might point towards? The other thing to say is, like, I was clear that I wanted to stick this all on GitHub

so that people can have a look at the process and the chat logs and everything else that went on, the various versions, including the incomplete versions of the chapters. And it was also published under a Creative Commons license in terms of copyrights so that it can be shared, remixed, et cetera, and people can do anything that they want with it on a non-commercial basis. If it gets sick and makes you a lot of money, is it? Well, it's, it's, I don't think it will make me a lot of money. The only money I would make out of it would be the Kindle edition.

So we'll see how that goes, but I'm not expecting it to be hugely lucrative. Brian, I just, I just started thinking about this because books get turned into podcasts, get turned into movies, get turned into episodic TV. What about getting this put into a treatment or getting it put into a script and then testing it on the next level? You thought about some of that stuff? Not yet, but in terms of the UK copyright regime is quite clear that people who have played that facilitatory role in the creation of AI generated works have copyright. So for commercial use, I would have that opportunity potentially to exploiting it commercially. I'd certainly be interested in seeing it in other formats.

I guess it's just out. It remains to be seen how the world responds to it, really. But if people are interested, then it would be great to see it fly. Are there any other questions left unanswered for you that you wanted to have answered when you started this process? And if so, what are they? I think it answered my main question. My main question, really, like going back to my motivations, I adopted an image change. generation AI, like text to image generation platforms quite early on. So I was an early adopter

of Dali and then like lots of people, I just found Mid Journey was too good. So I just switched to Mid Journey. And I've done lots of like design projects and all kinds of photography projects using Mid Journey. But one of one of the early projects I did was based on a prompt, which was saying an award-winning photograph of absolutely nothing because I was interested to see what what Mid Journey would choose, because you can't represent absolutely nothing in the format of a photograph without representing something. So it was a way of me trying to, again, look under the hood without giving too much creative input. I gave the basic, you know, I gave the aspect ratio, I gave the fact that I wanted it to be an award-winning photograph. So that gives some kind of canvas on which

it to operate. Other than that, I wasn't going to tell it what to render in terms of the image it was presenting. So this was like a continuation of that kind of thought, if you like, which is to see whether or not these systems are capable of doing something other than reflecting back to us the next token prediction thing. So basically it's entirely dependent on what prompt or what question you ask as to how it will respond to you. And I wanted to try and flip that and see if there was anything else going on when I didn't provide it with any creative direction. And I think The answer is quite clear that there is something else going on. Now, you could argue that that was kind of a process of recombination,

or you could argue that it was something else. And perhaps as far as Crawd is concerned, maybe the answer is buried in the echo chamber itself. Recursive, yeah, right. I think that it was a great experiment, and I think that you've answered a lot of the questions that people have been asking and posed more questions, what it is art.

what is storytelling? What does it mean to be a writer? What does it mean to be creative? What does it mean to tell a story? We're storytelling creatures. We survive. We evolved. We grew from stories.

And if that's changing, then what does it mean? And what's it going to be like in three years? What books are going to be produced in three, four, five years? What's it going to be like? Yeah. I mean, it's a roller coaster, isn't it? Yep. So, Brian, where can we read the echo chamber?

but where can listeners find out more about your work, your project and get involved? Yeah, the Echo Chamber is on Kindle. I don't know if we'll be able to provide a link to... Yeah, we'll be able to provide links, yeah. And also, it's a long URL, but I can provide the GitHub link, which contains the entire repository and the original manuscript in PDF format as well. People are welcome to go there. Awesome.

We'll put all that on thinking on paper.xyz on our LinkedIn, our Instagram, our Twitter, our YouTube, all the channels, all the channels and our new substack that Jeremy is writing, substack, thinking on paper to XYZ, you'll find all the links there, all the links to Brian's work. And pocket edition number one, stay tuned. We're going to be talking about music, cinema. Who knows if the creative act is involving technology, we're going to be looking at it. Jeremy, final thoughts on that? Are you convinced? Final thoughts really quick. So I think it's a great experiment. I think it's a great, I would even, dare I say, a research study into like this, this idea of how humans and tech interact. I will say disruptors and curious minds, you still have one powerful human tool at your disposal and it was used to start this whole experiment. The art of asking a great question.

Humans are still better at it today. Who knows what's going to happen tomorrow? but those are my final thoughts thinking on paper. comy-Z. Brian, thanks for joining us. Be curious. Stay disruptive. Keep thinking on paper.

Thank you so much, guys.