Story Generators: a new pillar of interactive entertainment

Every new entertainment medium spends its first years being described as a version of an older one. Early films were called "photoplays", recorded theatre. Early radio drama was theatre without the pictures. Video games were, for a while, "television games", something you plugged into the set the family already owned. In each case it took years before people stopped measuring the new thing against the old one and saw it for what it was.

Interactive stories generated by AI are at that stage now. They are usually described as a feature of video games ("AI NPCs"), or as a kind of chatbot ("talk to a character"). We think both descriptions are wrong, and that getting this right matters for the people who will build the field. This is our case for treating Story Generators as a second, evolutionary, stand-alone pillar of interactive entertainment rather than a new genre: separate from classical video games, with their own craft, their own rules and their own history still to be written.

Two pillars

For fifty years, interactive entertainment has effectively meant one thing: the video game. It has many genres, from puzzles to open worlds to competitive shooters, but they share a foundation. Someone builds a system of rules and content in advance, and the player explores what was built.

Story Generators work on a different foundation. Someone writes a world and its people in advance, and a model performs that world for each player, live, in language. The difference is not one of genre. It is a difference in what the work is made of, how it is authored, and what the player is actually doing. That is why we call it a second, evolutionary, stand-alone pillar of interactive entertainment rather than a new genre. It evolved from what came before, as games evolved from film, television and the tabletop, but it stands on its own.

Defining - Video Games

By classical video games we mean interactive works in which the experience is assembled from content and rules that were authored, or procedurally specified, before the player arrives. The game is a machine built in advance. The player's freedom is the freedom to move through it.

That definition covers almost everything ever sold as a game: from the text adventures of the 1970s to today's largest open worlds. A classical game may be enormous, emergent and surprising. It may use procedural generation to lay out levels or scatter loot. But its story, where it has one, is written, and its characters say lines that a writer wrote, chosen by a branching structure that a designer built. When a character in a classical game has nothing scripted for what you just did, they fall silent, or repeat themselves.

None of that is a weakness. Authored structure is why the best games feel deliberate. It is the craft of fifty years, and it is not going away.

Defining - Story Generators

A Story Generator is an interactive work in which an AI model generates the story and/or the character dialogue live, in response to the player's choices, while staying bound to a canon written by a studio or an author.

Each part of that definition does work:

• The story and/or the dialogue is generated. Not looked up, not branched to, not assembled from barks. Written at the moment it is needed, for this player, from what has happened so far.
• It is live and responsive. The player's words and choices are the input. There is no list of options the writer anticipated; the player can say what they like.
• It always includes player choice. The player decides what happens next: which way the next chapter goes in a story, or what their character does and says in an RPG. A model writing on its own while you read is not a Story Generator.
• It is bound to canon. The world, its history, its characters, their voices and the shape of the story are authored by people, and the model's job is to perform them faithfully. Canon is not decoration. It is what makes the result a work rather than an improvisation.

A classical game is a machine you explore. A Story Generator is a world that performs itself for you, from a script that keeps rewriting its lines but never its truths.

Games inside, games around, and always live

The line between the two pillars is not a line between games and stories. A Story Generator can contain games: a puzzle to solve, a chase to survive, a hand of cards to win. It can also be wrapped inside a game, with a generated story running at the heart of a world that has maps, missions and mechanics of its own. Depending on the experience, the generation happens within the game, or it is the game.

Story Generators also share something important with classical video games: neither has a set format. A video game can be anything from a falling-block puzzle to a sprawling simulation of a city, and games keep inventing new rules and new kinds of experience. Story Generators can do the same. That freedom is not what separates them.

What makes them different is live generation. Every Story Generator is a live game. A classical game can be live, through multiplayer or a service that keeps adding content, but it can equally be pre-packaged: everything the player will see already exists on the disc or in the download. A Story Generator can never be pre-packaged. The narrative and the dialogue are generated in the moment, so for the player and the reader it is always a live experience, and no two people ever receive the same one.

Player choice is not optional

Every Story Generator puts a decision in the player's hands. That is the difference between being told a story and being part of one. A model that writes a novel while you watch is a remarkable tool, but it is not interactive entertainment. The moment the player's choice shapes what is generated next, it is.

The form that choice takes depends on the kind of Story Generator:

• In a story, the player chooses the direction of the next chapter: who to follow, what to risk, which door to open, what a scene should turn on. The model then writes that chapter, faithful to the canon and to every choice made before.
• In an RPG, the player chooses as a character: what they do, what they say, who they trust. The model narrates what happens, the other characters respond in their own voices, and the story bends around the player's character.

In both cases the choice is real. It is not a pick from three options a writer prepared earlier, because the next chapter or the next reply did not exist until the player made it.

Story vs RPG: two kinds of generative play

Story Generators come in two main forms. Both perform an authored canon live, and both put choice in the player's hands. What differs is where the player stands.

A story generator places the player beside the story, closer to an author or director. You shape where the narrative goes, chapter by chapter, and watch the cast play it out. The pleasure is the pleasure of a great serial: what happens next, and knowing you decided it.

An RPG generator places the player inside the story, as a character. The world is narrated around you, and you decide what your character does next. The narration can take many forms: a narrator describing the scene, other characters speaking to you, or both at once. You run down the stairs and find a zombie at the bottom. Jane shouts that there is a gun in the drawer. What do you do? Sometimes the answer is a line of dialogue, more often it is an action, and the model narrates what happens as a result. The pleasure is the pleasure of being there: every decision is yours, and the world answers it.

Why neither needs a dungeon master's rulebook

Stats, dice rolls, hit points, skill checks, inventories and combat systems are classical mechanics. They came from tabletop role-playing games, where a dungeon master needed rules to decide what happened, and were carried into video games, where code needed numbers to decide it. They belong to the classical pillar, and they work well there.

Story Generators have a different core mechanic: input and output. The player acts, in their own words, and the model generates what happens: the narration, the consequence and the characters' replies. Then the player acts again. Both halves are the mechanic. The player's input is the move; the generated output is the world's answer. Everything else is secondary to that loop, and the whole of it is focused on the player.

That is why carrying the dungeon master's rulebook across is usually a mistake. Complex mechanics compete with the loop. Every turn the model spends tracking numbers and adjudicating rules is a turn it is not spending on what makes the medium worth playing: a scene that lands, a consequence that follows from what you actually did, a reply that could only have been written for you. Heavy systems also get in the way of the player's input. When an action has to be translated into a skill check before the world is allowed to respond, the player is no longer acting, they are filling in a form.

That does not mean anything goes. Structure still matters, but in a Story Generator it lives in the canon: what a character knows, what she wants, what she would never do, what must be earned before a secret is told. A guard who will not let you pass because of who he is makes a better obstacle than one who will not let you pass because you rolled a four. Light tracking, such as relationships, knowledge or what a character remembers about you, can help. The test is whether something makes the input and output better, or just gets between the player and the world.

What a Story Generator is not

The boundary matters because AI is now arriving in every corner of game development, and most of it does not create a new medium. It makes the existing one faster or richer. Three uses in particular are often confused with Story Generators:

1. AI NPCs inside a classical game. Giving a shopkeeper in an open-world game the ability to chat is a welcome improvement, and we have written about designing NPCs that listen. But if the game's story is still scripted, its quests still authored and its ending still fixed, the game remains a classical game with a more talkative cast. The generated dialogue is a layer on top, not the substance.
2. Code generation. Models that write gameplay code, shaders or tools change how games are built. They do not change what the player experiences as a medium. A game written with AI assistance is still a classical game.
3. Asset generation. Textures, 3D models, animation, voices and music made with AI are production techniques. A world full of generated trees is still a world whose story was written in advance.

A useful test: if you removed the model, would the work still exist? Remove the AI from a classical game that uses generated assets and NPC chatter, and you have the same game with fewer assets and quieter characters. Remove the model from a Story Generator and there is nothing left to play, only the canon waiting to be performed.

Canon is the contract

If there is one rule that separates a Story Generator from a model improvising fiction, it is this: a Story Generator must contain a studio or author canon. Story data, character data, or both.

By canon we mean the material a model cannot overrule:

• World: places, history, rules, technology, what is possible and what is not.
• Story: the premise, the secrets, the turning points that must land, the endings that are allowed.
• Characters: who they are, what they want, what they know and do not know, how they speak, what they would never say.
• Tone: what kind of story this is, and what it is not.

Canon does three jobs at once. It gives the player something real to push against: a secret that stays secret until it is earned, a character who refuses you because that is who she is. It gives the model a standard to be checked against, so that a character alive on Monday is not dead on Thursday. And it gives the work an author. As we argued in Who writes the story?, the writer's canon is what keeps authorship with people, even when every performance is new.

This is also why a general-purpose chat app, however fluent, is not a Story Generator. It has no world to be faithful to. Ask it for a dragon and you get a dragon; ask again and you get a different one. Nothing in that exchange is at stake, because nothing was ever fixed.

A field at its 1970s moment

The best comparison for where Story Generators are today is not the modern games industry. It is the late 1970s, when interactive fiction was invented by a handful of people on computers most of the public had never seen.

In 1975 and 1976, Will Crowther, a programmer who had worked on the early ARPANET, wrote a text game for the PDP-10 mainframe based on his maps of Kentucky's Mammoth Cave, the longest cave system in the world. He partly wrote it so his daughters, who were not technical, could play. In 1977 Don Woods expanded it with more fantasy, and Colossal Cave Adventure spread across the network. The program acted as a narrator, describing where you were and what happened when you typed a one- or two-word command.

At MIT in May 1977, four members of a research group (Tim Anderson, Marc Blank, Bruce Daniels and Dave Lebling) started building their own answer to Adventure, with a far better parser. They called it Zork. In June 1979 they and colleagues founded a company, Infocom, to sell it. In 1978 Scott Adams, working on a TRS-80 at home, released Adventureland, a small adventure that fitted into a home computer's memory, and founded Adventure International on its success.

And in 1979 a young company called Microsoft, then known for its BASIC, published Microsoft Adventure for the TRS-80 and Apple II: a home-computer version of Colossal Cave Adventure, programmed for Microsoft by Gordon Letwin, with 130 rooms and 15 treasures. When IBM launched its Personal Computer in 1981, Microsoft Adventure was one of the first two games available for it. Microsoft had even been offered Zork and passed, because it already had its own adventure.

Count the studios in that story: Crowther and Woods, the Infocom group, Adventure International, Microsoft, and from May 1980 On-Line Systems (later Sierra), whose Mystery House is often called the first graphical adventure. A handful of teams, working in text, on hardware that could barely hold a story, inventing a medium that most people had not yet heard of. Within a decade, that medium had grown into an industry.

Story Generators are in the same position. The tools are new, the audience is early, the conventions do not exist yet and most people who try one today are trying it in text. The studios doing serious work can be counted on two hands. The question is not whether this medium will grow, but who will define its craft while it does.

One more parallel is worth drawing. Those early text games and today's Story Generators share an interface: the player types, and the world answers in prose. But the early parsers understood a few hundred words and only the first letters of each. Adventureland read just the first three letters of a command, so "lig lam" meant "light lamp". Everything a player could do had to be anticipated by the author. Story Generators remove that ceiling. For the first time since 1976, the narrator can understand anything you say, and the author's job changes from anticipating every input to defining a world well enough that any input can be answered truthfully.

The market today, beyond generic mobile apps

Most people's first experience of AI fiction today comes from general-purpose mobile apps: chat with a character, ask for a story, roleplay a scenario. These apps have proved the demand is enormous. They have also, in our view, set expectations far too low. Most have no authored world, characters with no history beyond a paragraph of description, no memory you can trust and no story that goes anywhere. They are conversation, not narrative. Generic is the stage this medium has to grow out of.

A third layer on every device

Step back from games for a moment and look at how people use their devices. For thirty years a computer has had two main layers: classical apps, from games to productivity software, and the web browser, the window onto information sites and web applications. AI Generators are becoming a third layer beside them. On the desktop that layer already includes AI chat, agent harnesses and a new wave of AI-first development tools. On mobile it is creative tools and Story Generators. In VR it will be chat worlds and video Story Generators that surround the player.

Beyond those apps, the serious work is spread across five areas:

1. Canon-led story works. Studios writing complete worlds, story bibles and casts of characters, and using models to perform them. This is the core of the Story Generator field and still the smallest part of it.
2. Authoring tools. Software that lets writers build canon in a structured way (characters, locations, timelines, secrets, rules) and see how a model performs it. This is the equivalent of the engine and level editor in classical games, and it barely exists yet.
3. Memory and continuity. Systems that let a world remember a player across days and weeks without contradicting itself. We cover the design questions in Memory is the new save file.
4. Voice and performance. Open-weight expressive voice models such as Fish Audio's Drama 3 make it possible for a generated line to be acted, not just read.
5. World models. Models that generate explorable, interactive video worlds in real time: Google DeepMind's Genie 3, announced in August 2025, generates navigable worlds at 720p and 24 frames per second that stay consistent for a few minutes. Open models are following fast: Matrix-Game 4.0, WorldPlay2, SolarWM H3 Realtime and Viggle's JST-2.

Today these pieces mostly live apart. A world model can show you a place but does not know who lives there. A language model can voice a character but cannot show you her street. The next phase of the medium is about joining them, with canon as the thread that runs through every layer.

How the medium is evolving

Video games grew from text to pictures to animation to 3D, each step adding a layer of presentation on top of the same underlying idea. Story Generators are following a similar path, in a fraction of the time, because the underlying models are improving so quickly.

Text is where the medium is now, as interactive fiction was in 1977. It is the cheapest layer to generate and the easiest to keep faithful to canon, which is why it comes first.

Illustrated stories sit right beside text. The story and dialogue are generated live, while the pictures are pre-generated by the studio: illustrated scenes, characters and locations drawn ahead of time from the canon and shown as the story reaches them, alongside illustrated puzzles and other light game elements. Because the art is made in advance, it can be checked and kept consistent, so a character's face and a city's streets look the same from scene to scene. This is the step that corresponds to Mystery House adding pictures in 1980, and it is already here.

Voice is arriving in two steps. First, text stories with audio: narration and character dialogue performed by expressive voices, closer to radio drama than to a book. Next, live voice conversation with the characters themselves, where the player speaks and a character answers out loud, in her own voice, from canon.

Generative interactive film comes next: scenes rendered as moving pictures, cut and paced like cinema, but performed differently for each viewer. We explore this in Every viewer a different cut.

3D generative worlds are the far end of the staircase: places you can walk through, rendered by world models, populated by characters performed from canon. At that point the line between Story Generators and games will look thin from the outside. It will not be thin underneath. A 3D Story Generator is still a world performing an authored canon, live, for one player. A 3D classical game is still a machine built in advance.

Generative VR: the next great leap

Beyond the screen lies generative virtual reality: worlds generated around the player in real time, in every direction, at the scale of a city. We think this will be viewed the way we now view the leap from text adventures to modern 3D worlds, such as the explorable environments of Syberia or the living simulation of Grand Theft Auto. In the late 1970s a cave was a paragraph of text. A few decades later it was a place you could walk into. Generative VR is a leap of the same size for Story Generators.

It will not come cheaply. Rendering a generative VR world, two high-resolution views at a high frame rate, with every surface, character and line of dialogue generated live, will require tremendous amounts of compute. Our view is that it arrives in the mid to late 2030s, as the cost of generation keeps falling and the hardware to run it reaches homes.

When it does, we expect it to become a dominant form of video entertainment among younger generations. The reasons are the same ones that make Story Generators compelling today, multiplied: content that never runs out, immersion that surrounds you, and free-roam worlds that expand in whatever direction the player chooses to go, rather than ending where a studio stopped building. Canon still anchors those worlds, keeping their history, characters and truths intact, but the map is no longer fixed in advance. It grows from the player's choices.

This is the road we are on. Our name, Senrama, is a blend of sensory and diorama: a miniature 3D world that engages the senses. It describes where we believe entertainment is heading, from text, through illustration, voice and film, to immersive and passive 3D computational entertainment: worlds you can step into and act in, and worlds you can simply sit back and watch unfold around you.

Where it goes: interactive film and 3D generative worlds

Film gave audiences a story they watch. Games gave them a world they act in. Story Generators, at their furthest point, give both at once: a film that listens, and a world that has a story to tell.

Picture a detective story set in a city its writers have mapped street by street. You walk into a pub that exists in the canon, and the landlady who runs it, whose history, debts and secrets are written down, looks up and recognises you from last week. You can ask her anything. She will answer as herself, and she will not tell you what she has no reason to tell a stranger. The scene is lit and shot like a film, the camera finding her face as she lies. Nothing about the exchange was scripted. Everything about it was authored.

Getting there needs three things to come together: world models that stay consistent for hours rather than minutes; character models that hold to canon under pressure from millions of players; and authoring tools that let a writer build all of it without becoming an engineer. Each of these is moving quickly. None of them, on its own, is the medium. The medium is what happens when they are joined by a canon.

The manifesto

We build Story Generators. These are the principles we hold ourselves to, and the ones we think the field should hold to as it grows:

1. Story Generators are a second, stand-alone pillar, not a genre or a feature. They evolved from classical video games and stand beside them, not inside them. Their craft deserves its own name, its own tools and its own criticism.
2. No canon, no Story Generator. Every work must contain an authored canon: story data, character data or both. Generation serves the canon; it never overrules it.
3. Generation is the medium, not the shortcut. AI that writes code or makes assets for a classical game is a production tool. In a Story Generator, the live performance of the story is the work itself.
4. Always live. A Story Generator may contain games or sit inside one, but its narrative and dialogue are always generated live. It is never pre-packaged.
5. Every Story Generator gives the player a choice. The next chapter in a story, or the next word and action in an RPG. Without player choice there is only a model telling a story.
6. The story leads, not the rulebook. Stats, dice and skill checks are classical mechanics. The core mechanic of a Story Generator is player input and generated output; structure belongs in the canon, and rules that get in the way of that loop should go.
7. The player's words matter. Anything a player can say deserves an answer that is true to the world. No dead ends, no "I don't understand that", no menus pretending to be conversation.
8. Consistency is respect. A world that contradicts itself has stopped taking the player seriously. Continuity and memory are part of the craft, not an engineering detail.
9. Credit where it is due. The work is credited to the studio and system that brings writer and machine together, because no single part of that partnership is the whole author.
10. Start in text, aim for worlds. The medium will grow from text to illustrated and voiced stories, then to film and walkable worlds. The canon is what carries a story through every layer.

In 1979 a few people typing two-word commands into a cave they could not see were playing the first version of something that became one of the largest art forms in the world. Most of them had no idea. We think Story Generators are at that point now, and we would rather name the thing clearly while it is young.

Questions