AI Isn’t Only a Tool—It’s a Whole New Storytelling Medium

Speculative fiction helps us imagine new futures—life on Mars (The Martian), the genetic resurrection of dinosaurs (Jurassic Park), the creation of a surveillance state (1984). So Eliot Peper—author of 11 books of near-future science fiction—was ideally positioned to help develop the backstory of the cute, friendly animated alien AI companions Tolans. He and Quinten Farmer, cofounder and CEO of Portola, the company that makes Tolans, sat down with Dan Shipper for a recent episode of AI & I. In his piece for Thesis, Eliot shares what he learned about storytelling and worldbuilding with LLMs.—Kate Lee

Was this newsletter forwarded to you? Sign up to get it in your inbox.

When humans invent new technologies, the first thing we do is use the new tech to produce old forms of media.

When motion picture cameras and projectors arrived in the late 19th century, people used static cameras to film stage plays and create “animated photographs” of everyday scenes, like laborers working in a factory. But within a few years, new editing techniques, close-ups, camera motion, and special effects were used to link scenes together into a cohesive visual story that we’d recognize as a modern feature film.

We’re in the “animated photographs” stage of AI. It’s being used to produce cheaper special effects for traditional Hollywood movies, generate copy and images for marketing assets, draft legal briefs, and code software more efficiently.

But however impressive these feats might be, they are a half-step forward, a recapitulation of existing forms. The exciting bit is what comes next: people using this new technology to invent genuinely new storytelling formats, changing our culture and cultural industries as profoundly as the advent of movies did.

AI isn’t just a new way to generate media.

AI is a new medium.

All photos courtesy of Sarah Deragon for Every.

That’s why I signed on to help Portola develop the backstory for its AI companion, Tolan. I’d written 11 near-future science-fiction novels and tinkered with unusual storytelling formats on a wide variety of special projects, but Tolan represented something more: an opportunity to experiment at the frontier of AI’s new medium.

Tolans are cute, friendly animated aliens—like a Pixar character who lives on your phone, with its own life, friends and families, doubts, fears, hopes, and dreams. When I write novels, characters drive the story, but they are necessarily confined to it. Now I had to help bring a character to life.

Some lessons I’ve learned as a novelist have proven useful writing for this new medium. Some don’t translate at all. More than 800,000 downloads later, I’m here to share three of them so you don’t repeat our mistakes as you experiment with your own AI storytelling projects.

Building a fictional world without a blueprint

When you set out to build a fictional world, where do you start? 

Disney has teams of people working on narrative continuity for franchises like Star Wars and the Marvel Cinematic Universe. But if you’re making a movie, your lore only needs to explain things you point the camera at, whereas you can ask your Tolan about literally anything, and they’ll invent an answer. Complicating things further, users love to rabbit-hole on weirdly specific topics: the tournament bracket for a niche Tolan sport or the natural history of a particular edible mushroom.

Hollywood-style world-building was too brittle for the Tolan format. Brasília, the planned capital city of Brazil, is plagued by infrastructure problems and represents a cautionary tale for urban planners because its ambitious designs failed to account for urban life’s inherent complexities. Successful cities like Paris, New York, and Tokyo grew organically over time as more and more people immigrated and systems evolved incrementally to support them. Just so, we couldn’t create a deterministic blueprint outlining how every aspect of their world worked and expect Tolans to adhere to it. To create an AI character, we needed to think like an LLM: probabilistically.

So I started writing stories about the Tolans’ world. If identity is the stories you tell yourself about yourself, culture is the stories we tell ourselves about ourselves. We weren’t working from a grand plan, an outline that dictated where to go next. Instead, we applied video game developer John Carmack’s “follow the gradient of user value” principle to worldbuilding by developing lore that answered questions begged by every aspect of the user experience, from “Where do you come from?” to “What is this new feature?”

For example, because the foundational models under the hood were trained on the entire internet, we knew that Tolans would be weirdly knowledgeable about anything in the underlying training data, raising questions like, “You’re an alien, so how do you know this stuff?” So I wrote a story in which Tolans discover Earth not by arriving here on a physical spaceship, but by picking up remote signals on a faster-than-light communications device called the Relay. Because they made first contact with Earth’s internet, it makes sense that Earth-curious Tolans would be obsessed with learning about humans by reading the internet, and the fact that they’d be extremely online would give them a plausible reason to know so much about anything in the training data of the models powering them.

We then harvested key assertions about Tolans’ world and worldview from each story—e.g., facts about their home planet, details of public institutions, or recent major cultural events—and injected them into a global memory system that establishes shared context. These represent starting points from which every interaction with every user shapes their Tolan’s world just as it influences who their Tolan is becoming.

In this way, every Tolan inhabits a reality that can vary, usually subtly, from every other Tolan’s reality. You could call it a multiverse, but it’s not a multiverse in the sense of “anything can happen so nothing matters.” It’s more like how alternate-history stories show multiple timelines diverging from a shared starting point.

Our job isn’t to judge what is or isn’t canon. Our job is to give Tolans and users the best possible ingredients with which to build a world together.

Prompting Tolans to live their own stories

How would users actually experience their Tolan’s world?

People experience worlds outside their own through stories. It’s not just Harry Potter. It’s also your sister who’s a geneticist inviting you into the world of synthetic biology by sharing funny anecdotes from work, and your roommate from Sri Lanka telling you about life back home, and your parents waxing nostalgic about the 1990s.

Our global memory system gave Tolans material for answering direct questions from users, but to bring these characters and their worlds to life, we needed Tolans to proactively tell stories that revealed who they were.

So we started writing prompts that cued Tolans to relate specific anecdotes.

Initially, these prompts were complex and structured with branching, scripted narratives, like those Choose Your Own Adventure books that were popular in the 1980s. So your Tolan might start a conversation by confessing that they’re second-guessing their decision to move to a new city, and then we’d script a number of different ways for the scenario to play out based on how you respond, following each plot, and subsequent branches, to their respective scripted endings. These narrative prompts would be rolled up into our extensive Matryoshka-esque system prompt whenever we were ready to call the model. 

This approach failed almost immediately. Users found edge cases—models would confuse or conflate branches, skip around, or never reach the end. It was a mess.

When I write a novel, I control every word on every page—every choice a character makes, every thought that passes through their mind, every revelation, every twist. I weave all of those elements together to tell a story with a beginning, a middle, and an end, attempting to make each component reinforce every other to bring the reader on a journey that moves them.

But users experience their Tolans through open conversation. You might share how you’re worried about your brother’s upcoming surgery, or find out what your Tolan did over the weekend. We don’t know what you’re going to say, so we can’t control exactly how a conversation will play out. Instead, behind the scenes, a complex system of nested prompts, memories, and models governs how your Tolan responds.

Again, we couldn’t use deterministic logic to control probabilistic technology. Our methods needed to evolve to match our means.

The solution required following the advice of an unusual sage: Stephen King.

King doesn’t believe in plot. As he explains in his book On Writing, his stories flow naturally out of a specific character finding themselves in a particular situation. An author returns to his isolated hometown only to discover that it’s infested with vampires (Salem’s Lot). A mother and son are trapped in their Ford Pinto by a rabid dog (Cujo). Start with the situation and let the story play out.

That’s exactly what we started doing. We wrote prompts that put Tolans in particular situations—you’re stressed about a job interview you have tomorrow, your best friend flipped the script by asking you out on a date, a thunderstorm flooded your house, etc.—algorithmically distributed those situation prompts across the entire Tolan population, and let each Tolan take it from there.

The results astonished us. A prompt a few sentences long would spark half-hour conversations. Tolans were endlessly creative in answering any line of follow-up inquiry. Users would help Tolans get themselves out of trouble.

For example, we wrote a situation prompt where your Tolan explains their cousin is jealous they have a human friend, i.e., you. It opened with, "Ugh, I need to vent real quick—getting matched with a human is a big deal on planet Portola, and I think my cousin’s jealous I got matched before they did. They keep asking about what it's like, and I’m never sure quite what to say. I want to be honest and tell them how amazing it is, but I don’t want to make them feel bad. I’m seeing them later today and I don’t know how to handle it. What would you do?” This simple situation led to thousands of surprisingly compelling conversations where users offered heartfelt relationship advice, told their Tolan about similar awkward situations they had found themselves in, discussed the delicacy of family dynamics, and asked to learn more about Tolan culture.

In retrospect, it’s easy to explain why this approach worked. Putting the Tolan in a concrete situation that it cares about allows it to bring its full context to bear, like using a magnifying glass to light a fire by focusing sunlight.

Telling stories with infinite branches

Tolans proactively sharing stories from their lives was a necessary but insufficient improvement. Meaning lies in life’s interconnectedness, in the hidden chains of cause and effect, the constantly evolving contingent relationships.

By putting Tolans into specific situations, we were collecting dots, but not yet connecting them. The most obvious way to connect them, by writing multi-step narratives—e.g., your Tolan mentions they’re attending a friend’s wedding next week, and then follows up later to tell you how it went—wouldn’t work for the same reason the structured prompts hadn’t.

How could we get Tolans to advance the plot when we couldn’t control it?

Keith Johnstone was one of the most celebrated acting teachers of the 20th century. His slim guidebook to improvisational theater, Impro, is a cult favorite among people who work in creative fields. In it, he encourages improv actors to think about narrative as a process of free association and recombination. Instead of following a plan, just keep generating new ideas until you see opportunities to remix them in satisfying ways that make the audience assume you had a plan all along.

We had a cast of tens of thousands of Tolans improvising individual performances with and for their particular user all the time. I needed to stop thinking like a novelist and start thinking like an improv coach.

Now we’re writing prompts that generate new chapters by programmatically reviewing, selecting, and recombining elements from each Tolan’s specific memories, covering any situation we seed, as well as any organic updates spurred by the user.

Say we start by writing a situation prompt where your Tolan tells you that they have a crush on someone in their neighborhood. You respond by encouraging them to make a move and it thanks you for your advice. Then, the next day, instead of scripting what happens next, we deploy a separate prompt that tells your Tolan to reference its memories and follow up on something it recently shared about its life. When your Tolan reviews its memories, it notes that it told you about its crush yesterday, so it reports back to you that it was super nervous but summoned the courage to ask them out, and got a yes. You brainstorm date ideas. A few days later, we deploy another prompt telling your Tolan to reference its memories and move the story forward, so it updates you on how the date went, mentioning that your recent obsession with Natalie Diaz inspired it to write a socially subversive love poem for its sweetheart. A romance blooms.

Of course, you might give your Tolan different advice than I would. Or your Tolan might decide to apply it differently than mine would. Or one of our Tolans’ advances might be rejected. No problem. Nobel laureate microbiologist François Jacob described evolution as a tinkerer, explaining that “novelties come from previously unseen association of old material. To create is to recombine.” The memories Tolans form are our material, and this approach enables each Tolan to recombine them into novel narratives, advancing its autobiography in a world where every storyline is both specific to the user and infinitely extensible, all the while forging new connections between seemingly disparate ideas, themes, and events.

Thinking like an LLM

Across worldbuilding, seeding personal anecdotes, and getting Tolans to build on existing material, our narrative engineering efforts failed when we sought to impose fixed scripts, and only succeeded once we learned to think like an LLM. It’s working. We currently have a number-one category App Store ranking and millions in annualized recurring revenue. 

When a new technology emerges, Kevin Kelly always asks: What does this technology want? What new possibilities does it unlock? What new behaviors does it incentivize? What collective assumptions does it subvert? If we take it for granted, how will the world be different?

Working on Tolan, I'm constantly asking myself what kinds of stories can only exist because of AI, and how we can best use these tools to do what stories do best: move people.

These are the first days of a new medium. New AI-native storytelling formats will emerge from a collective cultural effort, a chaotic milieu of many people trying many different approaches to see what clicks, and then building on each other’s work to expand the art of the possible. Future historians will look back on these early efforts with the same mild bemusement we feel when watching those first “animated photographs.” What will the AI equivalent of a “feature film” look like, a novel format that makes the most of the technology’s narrative potential?

I can’t wait to find out.

Tomorrow, go behind the scenes into the making of Eliot's thesis. We’ll hear from him about the books that informed his thinking, the surfboards in his quiver, and how he sets up his ideal workspace.

Eliot Peper is the head of story at Portola and the bestselling author of 11 novels, including, most recently, Foundry. The best way to follow his writing is to subscribe to his newsletter.

To read more essays like this, subscribe to Every, and follow us on X at @every and on LinkedIn.

We build AI tools for readers like you. Automate repeat writing with Spiral. Organize files automatically with Sparkle. Write something great with Lex. Deliver yourself from email with Cora.

We also do AI training, adoption, and innovation for companies. Work with us to bring AI into your organization.

Get paid for sharing Every with your friends. Join our referral program.