Fans of George R.R. Martin’s A Song of Ice and Fire novels have been waiting for the sixth book in the series, The Winds of Winter, for 12 years now. 12. Years. That’s such a long time. It’s a small miracle that anticipation remains so high. Chock it up to the quality of the novels, and the fact that HBO has kept hype alive first with Game of Thrones and then with its prequel series House of the Dragon.
This year, the tech trend everyone has been talking about is AI. Write a prompt or ask a question to a chatbot like ChatGPT, and it can give you a coherent answer, drawing on millions of words of text written by human beings archived on the internet.
That’s how these chatbots work: they look at things people have already written, look at your prompt, and then respond one word at a time, choosing the word most likely to come next. The results can be pretty dazzling, but there are limitations. For instance, we had ChatGPT try and write the first five chapters of The Wind of Winter for us, and while it produced something readable, it was all very vague, didn’t seem to have much knowledge of Martin’s books, and didn’t resemble Martin’s writing style.
AI A Song of Ice and Fire
Developer Liam Swayne did one better: he had ChatGPT write only the entirety of The Winds of Winter, but also A Dream of Spring, which Martin plans to be the final book in the cycle (I’ll believe that when I see it, but that’s another issue). He shared the results of his project here.
And when I say Swayne had AI write the entirety of The Winds of Winter, I mean it; the AI-penned version of the book runs 683,276 words long, which is actually far longer than even the longest book in Martin’s series: A Dance With Dragons, which tops out near 415 thousand words.
How did Swayne pull this off? He explained his process to IGN:
First he gave ChatGPT a single prompt to generate an outline for the first chapter of The Winds of Winter. Then, he repeated that over and over to create 45 total chapters. From there, he fed those outlines back into ChatGPT and asked it for more detailed outlines of the same chapters. And finally, he used those extended outlines as prompts to ask ChatGPT to write the chapters themselves, turning every bullet point of the outline into its own scene.
The results are…well, they’re extremely long. But could they serve as an effective replacement for the real thing?
Swayne himself thinks the AI did some things right, like building in a few twists. He enjoyed the bit where Jon Connington turned traitor against Daenerys Targaryen, “a twist I didn’t see coming but served the narrative quite well.”
On the other hand, despite being prompted to, the AI didn’t kill off any important characters, a hallmark of Martin’s writing. In the end, Swayne isn’t really worried that AI could take over the job of a write like Martin anytime soon:
Large language models can be very scary, but this project makes me more optimistic about the future of writers and AI. This project demonstrates that large language models like ChatGPT can take hundreds of pages of text into account when making a narrative decision, which could help writers quickly fix plot holes. It also shows that AI can only do what has been done many times before. This project has given me confidence that AI won’t be replacing unique literary works any time soon. I believe the AI had trouble writing character deaths specifically because most writers (and consequently, most of the training data) are hesitant to kill off major characters. This is part of what separates George R. R. Martin from other writers: his stories make unconventional, surprising decisions. At this point, AI can only do what is most commonly done, which means it struggles to create stories that aren’t by the books. To my surprise, I’m more confident than before starting this project that writers making creative and unexpected decisions are not replaceable.
For fun, I looked a little deeper into the new books. Long story short, I agree that writers aren’t in much danger…