Interactive Narrative

Interactive narrative is often called the holy grail of game design. The idea: put the player inside a story that responds to their choices, where the world and its characters react with the coherence and emotional depth of authored fiction but the freedom and surprise of an open-ended simulation. We've been chasing this since the 1970s. We haven't caught it yet. But the attempts have revealed fascinating things about the structure of stories, the limitations of computers, and the irreplaceable role of human imagination.

A Short History of Machines That Tell Stories

Mark Riedl traces computational narrative intelligence all the way back to 1843 and Ada Lovelace asking whether Babbage's Analytical Engine could be creative. The first actual story generation system — a fairy tale generator by Grimes — appeared in 1960, using grammar-based AI. TALESPIN in 1977 flipped a story understanding system to generate Aesop-like fables about foxes and bears, launching a succession of generators through the 80s.¹

Then a fork in the road. Research on stories split into two branches that barely spoke to each other for decades. The bottom-up branch became natural language processing — learning to parse and predict text. The top-down branch became automated story generation — working with high-level cognitive representations of plot, character, and causality, assuming you could always render the result into text later. These two traditions are only now beginning to converge.¹

The AI Winter froze progress. When the field thawed in the 90s, computer graphics and games had arrived, and researchers started experimenting with interactive characters in virtual worlds. The Oz Project at Carnegie Mellon put believable agents into navigable spaces. The desire to join those agents in a fictional world and interact with them within an unfolding story became the core research agenda.¹

Facade and the Social Simulation Era

The landmark was Facade (2005), by Michael Mateas and Andrew Stern. You visit a married couple — Grace and Trip — who are in the middle of an argument that might end their relationship. You type natural language to interact. Under the hood: a language processing system, a behaviour engine called ABL that controls character actions, and a drama manager that creates beats of tension.²

It doesn't always work. The AI struggles with nuance. But it was a genuine attempt at what Chris Crawford had called "games about people, not things" — a phrase from his infamous 1992 Dragon Speech at GDC, where he charged out of the room with a sword declaring that the industry would never pursue meaningful human expression.³

After Facade came Prom Week (2012), where characters were driven by thousands of rules encoding social norms, with histories, personalities, moods, and interpersonal attitudes. A drama manager searched for appropriate scene templates when social states changed, giving the underlying AI a "narrator" who could explain what was happening. The developers described the magic moment when characters began performing correctly even without dialogue written for them — "like watching trained actors start venturing into improv."²

Then Versu (2013), by Richard Evans and Emily Short, an engine for interactive text-based stories where every character had beliefs, abilities, interpersonal attitudes, and emotional states based on Paul Ekman's typology. Characters gossiped, formed opinions about each other's behaviour, and remembered slights from hours earlier. Evans tells a debugging story: a doctor was inexplicably rude to the player character, and it took him a long time to trace the cause — early in the game, the player had been rude to a servant during dinner, the servant had told people in the kitchen, and the doctor remembered.²

The problem was communication. When the first Versu game, Blood and Laurels, shipped on iOS, players thought it was a standard choose-your-own-adventure. They didn't realise the rivalries between characters were being generated by AI. The complexity was invisible — an echo of the Ultima Online ecology problem that social simulation keeps running into.

The Disco Elysium Approach

If the simulation-driven approach to interactive narrative keeps struggling with communication, perhaps the answer lies not in more AI but in better design of the player-character relationship. Disco Elysium takes a radical stance here, and a sharp analysis of its dialogue system reveals principles that are deceptively simple but brutally hard to execute.⁴

Four rules govern Harry's dialogue:

1. Harry never acts on his own — every word and action must be vetoed by the player.
2. Replies are always direct speech — the specific words Harry will say.
3. There's no information in a reply that the character knows and the player doesn't.
4. The author's voice (hard facts) and character voice (filtered through perception) are clearly distinct.

The game twists into logic pretzels to avoid breaking these rules. When Harry must know something the player doesn't, the voices in his head tell both of them — so it feels like learning, not like the character started with hidden knowledge. When Harry must say something embarrassing, the internal voices stage an elaborate ceremony about losing control, and you still have to click the line — so it feels like a lapse of self-control rather than a designer's scripted moment.

The result is a sweet spot that neither open-ended simulation nor rigid authored narrative typically achieves. Harry is a specific person with a specific life, specific opinions, specific demons — but he's also a character you can roleplay as you want. The limitations are internally justified. And the critical insight is that this works precisely because of the constraints, not despite them. The hero must be an amnesiac (so player and character learn together). He can't be wordy (every reply short and direct). And crucially, the player can't plan dialogues — no rhetorical traps, no strategic information reveals. Harry is "fully and fundamentally a vibe-based creature, controlled more by the ancient reptilian brain than by the neocortex."⁴

This is interactive narrative not as a technical AI problem but as a design discipline — finding the exact shape of player agency that makes a specific character and story sing.

Fractal Stories: Text as a Different Medium

Filip Hracek spent five years trying to build "Skyrim in text" before nearly giving up. His breakthrough reframes the problem entirely: text stories are fractal.⁵

A video game needs a consistent level of abstraction — you're always moving through 3D space at 60fps. Text doesn't have this constraint. A text game can zoom in and out freely: from strategic ("charge at the bandit") to tactical ("dodge") to micro ("thrust sword at exposed flank") within a single combat exchange. The result reads like authored prose — "not just two people exchanging damage, it's an actual swordfight with dynamic potential outcomes every step of the way."

The fractal insight solves the fundamental tension between video games and text. Good video games always have repetition — throwing similar problems at you so you improve. Good text cannot have repetition — it gets boring. By varying the level of abstraction dynamically, the same action occurs in vastly different contexts, making repetition feel fresh.

But the constraints are real. Text can't handle spatial complexity well (describe a fight with ten actors and it becomes unreadable). Stats-heavy mechanics devolve into spreadsheets with flavour text. Procedural content isn't a silver bullet — you still need enormous amounts of authored content that gets recombined procedurally. And the approach works best for immediate, visceral, physically reactive genres. Sword and sorcery is ideal. Modern firearms are much harder — the author speculates this might just be a limitation of prose.⁵

The AI Dungeon Master

The dream of a true AI Dungeon Master brings together every thread in interactive narrative research. A pen-and-paper RPG DM does everything: generates worlds, runs NPCs with consistent personalities, adjudicates rules, improvises when players do unexpected things, manages dramatic pacing, remembers the entire history of the campaign, and does it all in real time through natural language.⁶

LLMs are excellent at one piece of this: bullshitting. Given concepts and guidance, they produce memorable scene descriptions. But the hard problems start immediately after. Memory — not just searching for "Harry Odd-Shoes" but understanding his whole narrative arc across dozens of sessions. Rules — knowing when to call for a die roll, interpreting results correctly, applying modifiers. Player creativity — handling the moment when players assassinate the villain mid-monologue with an elaborate scheme involving rodents, forks, and structural mechanisms.

The AI DM needs to understand maps, not just draw them. Players climb trees, push boulders, cast passwall through walls, and set fires. A memorable dungeon is a story told through architecture — the cult leader's opulent suite versus his spartan study tells you who he is. And the AI must expect all of it to be exploited.⁶

Combat AI in a tabletop context has to model creatures that care about living. Bandits surrender. Wolves flee from superior predators. Soldiers hold the line but their morale might break. Fiends treat death as an inconvenience (they'll respawn on their home plane) and weigh cost-benefit accordingly. Every encounter needs a theory of what these creatures want and how hard they'll fight for it. Video game AI almost never does this — it's one of the things that makes tabletop irreplaceable.⁶

The author's conclusion is surprisingly optimistic: it's doable, maybe, with a cognitive architecture that treats the LLM as one component among many — a bullshitting engine surrounded by symbolic rule systems, procedural generators, memory managers, and world state trackers. Not a general intelligence, but a specialised composite system. The precedent is Meta's Cicero, which plays Diplomacy competently by combining an LLM with strategic reasoning. From Diplomacy to D&D is a big step, but it's a step on a visible path.⁶

Computational Narrative Intelligence and Why It Matters

Riedl frames the stakes broadly: narrative intelligence — the ability to craft, tell, understand, and respond to stories — is fundamental to how humans communicate, teach, and relate to each other. A computer that can't understand or tell stories is an alien intelligence, and its failures will increasingly stem from an inability to make sense of what humans are trying to accomplish or why.¹

The applications go far beyond games. Story understanding requires commonsense knowledge — the set of socially shared beliefs about how the world works. ("Did John and Sally arrive at the restaurant together?" Nowhere in the text is this discussed, yet most people infer correctly that they didn't.) Medical AI needs to understand patient narratives. Robots need to explain their decisions in ways that sound like stories, not logs. And Riedl's Quixote system shows that agents can learn social norms by reading stories — a pharmacy robot that reads tales about buying medicine learns to wait in line and pay instead of just grabbing the drugs and walking out.¹

The grand challenge Riedl poses: an AI that can autonomously generate a feature-length animated movie scoring at least 20% on Rotten Tomatoes. He chose 20% because that would beat Transformers: Revenge of the Fallen. The gap between here and there is not just a better language model — it's the gap between predicting the next token and understanding what makes a story matter to a human being.