The Power of Choice: How Branching Narratives Work in AI RPGs

Choice is what separates a game from a book. When you choose, you create. And when your choice genuinely changes what happens next, you're invested in a way no passive story can match. The whole appeal of role-playing games rests on one lopsided fact: in a book you witness, and in an RPG you decide. I'm convinced that's why an hour spent steering a story sticks in the memory so much longer than an hour watching a film. You were doing something the whole time.

There's research pointing the same way. In a set of studies on why video games pull us in, a team led by psychologist Richard Ryan found that perceived in-game autonomy and competence go hand in hand with how much people enjoy a game. Autonomy, in plain English, is the feeling that your choices are really yours.

Building choice has always been a headache for game makers, though. You could tell much of the history of game design as the story of designers trying to give players meaningful choice without writing a separate game for every decision. Each era's answer hit its own ceiling. Below, I'll walk through the old models and show how an AI Game Master pushes past those ceilings. I'll also be straight with you about the new limits that tag along for the ride.

The Old Model: Choose Your Own Adventure

Traditional branching narratives, video game dialogue trees included, work by pre-writing every possible branch. The classic Choose Your Own Adventure paperbacks had you flip to page 47 or page 53 depending on a single decision, and they were a phenomenon: the series grew to 230 titles in more than 40 languages and sold over 250 million copies. Mass Effect's celebrated dialogue wheel was the same idea with voice acting and 3D graphics. The maths is brutal, though. With five decision points and three options at each, you need 243 unique story paths to cover every combination (that's three multiplied by itself five times). Real games never write 243 paths. They write a few, then quietly funnel divergent decisions back to the same trunk a few scenes later.

The game writer Sam Kabo Ashwell has a name for that funnel. In his catalogue of choice-based structures he calls it "branch and bottleneck": the game branches, but the branches regularly rejoin, usually around events common to every version of the story. It's a perfectly sensible design. It's also why so many choices end up feeling smaller than they looked.

The result is the illusion of choice: a handful of pre-written responses dressed up as a vast decision space. Players learn fast that most choices don't matter, and the few that do lead to the same small set of outcomes. Savvy players start gaming the system, picking dialogue options because they're chasing the "good" ending rather than because the line fits their character. At that point the story stops being a story. It becomes a flowchart with prose attached.

There were heroic attempts to break out. Tabletop RPGs solved branching by putting a human GM in charge of improvisation, so the GM became the engine that turns endless player choices into fitting responses. Some video games (the Telltale and Quantic Dream catalogues spring to mind) leaned hard on "consequences will follow" framing, yet in my view they quietly hid the same trunk-and-funnel structure underneath. Quantic Dream's Detroit: Become Human even made the flowchart literal. At the end of each scene it shows you a flowchart of the route you took, along with the branches you could have taken. I find that rather charming, and refreshingly honest, because it shows you exactly where the walls are. All the same, the underlying problem of finite, hand-written content never went away.

The AI Model: Responsive Generation

In EchoQuest, the AI Game Master doesn't pre-write branches. It reads your action, whatever you say or type, and generates the next story beat in response. The story gets built on the fly. It's constrained by your world's rules and the current context, but it's never picked from a list of ready-made options.

That means:

• Any action is valid. There's no "I don't understand that" for sensible in-world actions. If you can describe it, the GM can adjudicate it. Want to bribe the dragon? That's probably a charisma check, and a real d20 roll decides it. Want to teach the dragon a song? Same deal. And if you try something truly impossible, the GM is told to handle it in character ("The stone door doesn't give, however hard you shove") instead of throwing an error at you
• Choices compound. What you did in scene 3 can still be shaping scene 47, because the GM carries recent context and the structured game state records explicit decisions as story flags. I even wrote a line into the GM's instructions telling it to bring old threads back "when it hurts or helps the most"
• Side paths are real. If you'd rather investigate the merchant guild than follow the main quest, the GM follows you there and builds something out. The "main quest" is a direction the world leans in, not a rail. In fact, the GM's instructions say, word for word, to "follow the player's creativity instead of forcing them back onto a path". Stepping off the rail doesn't break the game
• Failure produces story. In a pre-written branching game, the only response to "you fail your check" is whatever the writer prepared. In an AI RPG, failure is a creative moment, because the GM improvises a consequence that fits your exact situation instead of a stock fail-state. I learned how much this matters from my own game. For a while, dropping to 0 HP in EchoQuest did, well, nothing at all. Now the GM sees that you're down and has to narrate a real setback, like being captured or waking up hours later with a price to pay

You still get suggested choices, mind you. Each scene ends with three to five of them, read aloud and mapped to the number keys, and one is always an open-ended option for doing something else entirely. The GM is told never to repeat a choice from the previous turn and to skip limp stoppers like "Wait and observe". They're only suggestions, though. Type anything you like, or press V and say it. Getting even this part right took some fiddling. On May 16 I realised the choices were being announced and focused while the narrator was still talking, so screen reader users heard two voices at once. Now the choices wait their turn.

The Texture of Meaningful Choice

What makes a choice feel meaningful isn't only that it changes the plot. It's that it reveals character, namely yours. Sparing a villain says something about who you are. So does negotiating before fighting, or asking a bystander's name before they matter to the story, or tipping the ferryman more than he asked. Insisting on a slower, less efficient route through a city because you want to see it counts too. These choices don't change the plot dramatically, but they pile up into a portrait of who your character is.

Pre-written branching almost never honours them. Nobody has the development time to write a warm version and a sarcastic version of "I ask the bartender what his name is". There just isn't time. So pre-written games either leave such choices out or render them with the same canned text regardless of intent.

EchoQuest's GM is built to honour exactly this kind of choice. When you do something unexpected but true to your character, the GM responds in kind: NPCs remember, and situations reflect back what you've done. Under the hood, every named NPC carries a standing with you from -100 (sworn enemy) to +100 (loyal ally), and whenever it shifts, the GM jots down a note of fifteen words or fewer explaining why. "Convinced to let us pass" is the example in its instructions. So a character who's been polite to every shopkeeper in town will find the town treats them very differently from a character who's been brusque, even if both made the "same" major plot choices. The plot is identical. The story is not.

There are achievements for this style of play as well, such as one for resolving a hostile encounter without violence. Frankly, I like that they reward who you are rather than how hard you hit.

What Choice Reveals About Story

Here's a pattern I think most long-time roleplayers will recognise. People start out making choices because they're trying to "win", or to find the canonical path. Then, gradually, they start choosing because the choice fits who their character is becoming. The game stops being a tree to optimise and turns into a person to inhabit. I'd bet that shift, when it happens, is the single biggest reason anyone sticks with an RPG for months.

The shift happens because the GM keeps honouring character-driven choices over plot-driven ones. Stay true to your character's flaws and the world responds. Betray your character's principles for a tactical edge and the world responds to that, too. There's no global "right answer" the AI is nudging you toward. There's just a world reacting to who you are. Isn't that what you secretly wanted from those paperbacks as a kid?

The Limits of AI Choice

AI narrative isn't perfect, and I'd rather be honest about the gaps than oversell. Long-term consequence tracking is harder than short-term. A choice you made twenty sessions ago won't be remembered as precisely as one from two scenes back. Here's how it works in EchoQuest today. Your recent turns travel to the GM word for word, up to a generous budget of roughly 40,000 tokens. Older turns get folded, ten at a time, into a factual summary of 100 to 200 words. Key facts about your choices, faction reputations and NPC standings are stored explicitly in the game state instead of living only in the conversation. Even so, the texture of older moments fades. The GM may remember that you spared the villain in chapter four, but not the exact words you used.

On top of that, the AI doesn't experience your choices emotionally the way a human GM might. It won't be visibly moved by a sacrifice or caught off guard by a twist the way a person is. The performance of being moved is there in the narration; the actual being-moved isn't. For some players that doesn't matter at all. For others it's a real absence, and I think that's a perfectly fair thing to miss.

There's also a particular failure mode where the AI, faced with a genuinely ambiguous choice, writes text that hedges in both directions instead of committing. My hunch is that it's a side effect of training models to avoid being wrong, and in narrative terms it produces mush. I push against it directly. Since my rewrite, the GM's instructions say the narrator commits: no "perhaps", no "it seems", no "you sense that maybe". If something really is uncertain, a character gets to be unsure out loud instead. The result is far stronger than an AI left to its own habits, although the odd hedge still slips through.

Still, the AI never gets tired, and it never runs out of patience. Within a session, its responsiveness is remarkable. Plenty of people greet any new game by trying to break it, making absurd choices and attempting impossible things, and I fully expect EchoQuest to get the same treatment. Go ahead. The GM will keep answering in character, and if you regret a move, pressing U undoes your last turn. My bet is that after a while, breaking it stops being interesting and you start making the choices your character would actually make. That's the moment EchoQuest becomes the kind of RPG that ruins other games for you.

See the difference for yourself →

Related reading: Browse all EchoQuest blog posts.