AI FANTASY RPGS: WHEN AI TELLS THE STORY

The first AI fantasy RPG I ever loved ended when the goblin I had been befriending for two hours suddenly became a medieval tax collector, demanded I pay him in cryptocurrency, and then dissolved into a pile of spiders. I sat there at my desk laughing until my ribs hurt, because nothing in any scripted RPG I had played that year had surprised me as hard as that. Also nothing in any scripted RPG that year had been that stupid. Both things can be true. That's the whole pitch for AI-driven fantasy roleplaying right now, and it's why I keep coming back to the genre even when it lets me down.

I've been playing AI fantasy RPGs since the original AI Dungeon dropped in 2019, which means I've watched this entire category go from a novelty web demo into an actual corner of the hobby with real tools, real communities, and real money changing hands. The tech has gotten better every year. The stories have gotten better every year. The fundamental problem has not gone away. I want to walk through where we are right now, what's working, what isn't, and where I think this stuff is actually headed, because I think most of the takes online are either breathlessly optimistic or smugly dismissive and the truth is a lot weirder than either.

The OG: AI Dungeon

You have to start with AI Dungeon because nothing else existed before it. Nick Walton's demo at a BYU hackathon got picked up, ported to GPT-2, and suddenly the internet had its first proper "type anything and see what happens" fantasy engine. You chose Fantasy as your setting, picked a character class, got dropped into a tavern, and the rest was up to you. I wrote about AI Dungeon and where open-ended AI adventures are heading in more detail elsewhere, but the short version is that it opened a door nobody has managed to close since.

Current AI Dungeon runs on much better models than the GPT-2 version, and you can pick between different tiers depending on how much coherence you want to pay for. The good news is that the writing is way better than it used to be. The bad news is that the company's relationship with its community has been rocky, with filter controversies and subscription changes that drove a lot of early adopters to alternatives. But as a historical artifact and as a still-usable tool, AI Dungeon is the foundation. Everything after it is a response to it.

NovelAI

NovelAI is what happens when you take the AI Dungeon concept and aim it at people who actually like writing. The interface looks like a word processor with a generate button, which is exactly the point. You write as much as you want, hit tab, the AI continues the scene, you edit what it gave you, you continue writing yourself, repeat. It's collaborative fiction more than it is a game.

The fantasy models NovelAI has trained are genuinely good. They handle tropes, they stay in voice, they remember lore entries you've defined, and they don't constantly derail into the absurd the way stock AI Dungeon sometimes does. If you care about narrative craft more than you care about systems and mechanics, this is probably where you want to be. I've run entire campaigns in NovelAI that read like actual first drafts of fantasy novels, with the caveat that I'm doing about 60 percent of the typing and the AI is filling in around me. That ratio is the trick. You can't hit generate and walk away. You have to be a co-writer.

Character.AI roleplay

Character.AI took the whole thing in a social direction that I initially thought was a dead end and have been quietly eating my hat about ever since. Instead of one model trying to run a whole world, you interact with individual AI characters, each with their own personality card. Users create and share them. The Dungeon Master personas are a whole genre. The romance personas are a whole other genre. The "I am a grumpy elven blacksmith in a generic fantasy village" personas are weirdly some of the best roleplay partners I've ever had.

The reason Character.AI works, at least for the parts of it that work, is that a single persona is easier to keep consistent than an entire world. The AI has a narrower job. It just has to be this one character. When you ask Gandalf a question, Gandalf answers. He doesn't also have to remember the layout of Moria and track which orcs are dead and which aren't. That constraint turns out to be productive. For one-on-one fantasy roleplay, especially character-driven stuff, I'd take a good Character.AI persona over a raw AI Dungeon session most days.

Mantella and the Skyrim AI modding scene

This is where things get wild for me, and where I think the next decade of the genre actually lives. Mantella is a mod for Skyrim that hooks up every NPC in the game to an AI language model and a voice synthesis engine. You walk up to a random guard. You talk to him. Not from a dialogue menu. You literally talk, with a microphone, in your own voice. He hears you. An AI generates his response in character, using his known Skyrim backstory, and a voice model speaks it back to you in a voice that sounds approximately like a Skyrim guard.

The first time I played with Mantella, I asked a shopkeeper what he thought about the Civil War. He gave me a five-minute speech about how the Empire had raised taxes on wool three times in the last year and he couldn't compete with the Khajiit caravans. None of that was in the base game. The AI made it up on the spot, but it made sense within the world, and it was rooted enough in what the shopkeeper canonically was that I believed him. I stood in his shop for twenty minutes just asking him questions about his life.

The catch, and it's a real catch, is that the AI has no mechanical connection to the game. The shopkeeper can tell me a beautiful story about his dead wife but he can't give me a quest to avenge her. The AI can improvise anything it wants but the game world doesn't know what the AI said. That disconnect, between narrative and mechanics, is the central problem every AI fantasy RPG is trying to solve right now, and none of them have fully cracked it.

Inworld AI and the studio-facing tools

Inworld is less of a product you play and more of an infrastructure layer that actual game studios are starting to build into their titles. It provides NPC brains with personality, memory, and goals, and it exposes hooks that let the game mechanics react to what the NPC decides to do. The early integrations have been mixed, but the direction is right. A couple of the recent indie RPGs using Inworld have given me glimpses of what it looks like when the AI character and the game systems actually talk to each other, and the glimpses are exciting even when the overall product isn't fully there.

The thing Inworld gets that a lot of the text-only tools don't is that games have verbs. Fantasy RPGs especially have verbs. You swing a sword, you cast a spell, you pick a lock, you steal a horse. If the AI can only generate dialogue, it's missing most of what makes an RPG an RPG. The studios figuring out how to let AI NPCs actually do things, not just say things, are the ones worth watching.

Replica Studios and the voice problem

Replica isn't strictly an RPG tool, but it matters for the genre because voice is how you feel a fantasy world is real. Text-based AI adventures have been around for years and they work, but they ask a lot of your imagination. The moment an AI NPC actually speaks, in a voice that has emotion and timbre and pacing, the whole thing crosses a line that text alone can't cross.

Replica provides voice actors trained as AI models, with proper licensing deals so the actors get paid when their voice models get used. Studios are using these to generate dialogue on the fly for NPCs whose lines are being written by language models. The Mantella mod uses similar tech. It's the thing making dynamic NPC dialogue feel like a real game instead of a tech demo.

The recent wave

I'm not going to pretend I've played every AI-enhanced RPG that has shipped in the last two years because there have been a lot, and most of them are not great. But a few are worth naming. Suck Up is a vampire game where you talk your way into houses by convincing AI residents to invite you in. It's a concept demo more than a full RPG, but the "argue with an AI to get what you want" loop is genuinely fun and points at where dialogue-as-gameplay could go. There are also a handful of larger fantasy projects in early access or late alpha that are building entire campaigns around AI GMs, AI companions, and AI-generated quests. The quality varies wildly. The ambition is real.

What AI RPGs actually do well

Infinite content. This is the biggest one. A scripted RPG has as much content as the writers wrote, and once you've seen it, it's gone. An AI RPG can generate new content forever, responding to whatever you want to try. You can decide to spend your whole playthrough running a tavern instead of slaying the dragon, and the AI will let you, and it will have opinions about which ales sell best.

Personal stories. The narrative adapts to you, specifically. Not to a "player type" the designers anticipated. To the thing you actually said or did five minutes ago. When this works, it creates moments that feel like they belong to you alone, which is a sensation almost no scripted game can produce because the studio doesn't know who you are.

NPC depth. The best AI RPG moments I've had are not dragon fights or epic battles. They're long conversations with a side character who wasn't supposed to matter, where the AI improvised a full inner life for someone the scripted game would have reduced to a quest giver. You can ask questions and get answers. Real ones. Improvised ones, sure, but coherent enough to feel real while you're in them.

What AI RPGs are bad at

Narrative coherence over long arcs. The AI generates locally plausible text, but it doesn't build toward anything. The way I described this in the branching narrative games piece applies here too: authored stories build pressure and release. AI stories tend to just keep going, at the same intensity, forever, until you get bored and quit. A five-hour AI adventure rarely has a satisfying ending because the AI doesn't know what an ending is. It just knows what text looks like.

Hallucinations. The AI will confidently tell you that Ironclaw Keep is ruled by King Aldric, and then twenty minutes later tell you that Ironclaw Keep was destroyed a hundred years ago, and then thirty minutes after that bring back Aldric with no explanation. This gets worse the longer the session runs. Lore walls and memory systems help, but the underlying problem is that language models don't actually track state. They pattern-match on context, and context is limited.

Gameplay verbs. This is the killer. You go to swing your sword and the AI describes your swing. You didn't actually swing anything. There was no hit detection, no timing, no risk. The word "swing" happened. That's all an RPG is to the AI, a string of words, and for some players that's enough, but for a lot of us an RPG is a system you interact with mechanically, not just narratively. The best AI games are finding ways to bridge this, but the bridge is still being built.

Repetition. AI generators have house styles. Once you've read a thousand AI-generated fantasy scenes, you start to notice them. The overuse of "you find yourself," the weirdly florid descriptions of eyes, the dialogue that always resolves into vague wisdom. Good human writers break patterns on purpose. AIs fall back into them unless you actively fight the tendency.

Where the genre is headed

Hybrid. That's the whole answer. The future of AI fantasy RPGs is authored games with AI-responsive details, not AI games with authored framing. A human designer lays down the world, the major plot beats, the quests, the rules. The AI fills in the NPCs, the dialogue, the small talk, the reactions to weird player choices nobody anticipated. Each layer covers the other's weaknesses. You get the structural satisfaction of a real story and the responsive flexibility of AI all at once.

We're already seeing this in indie projects and mod scenes. Scripted quests with AI-expanded NPCs. Procedural dungeons with handcrafted boss encounters. AI DMs running authored campaigns. The big studios are slower to move because their pipelines aren't built for hybrid workflows, but they'll get there, and when they do, the category is going to look very different from both traditional RPGs and current AI tools. The interesting question is whether the studios beat the modders to it. Right now I'd put even odds.

If you want more context on how the text-driven side of this family tree evolved, I'd point you at my piece on choose your own adventure history. AI RPGs are the current branch of a very old tree, and understanding where the tree came from makes it easier to see where the branch is growing. We've been trying to make stories respond to us since people were picking pages in paperbacks. The AI stuff is just the newest attempt at a very old dream, and even when it stumbles, it's doing something that was impossible ten years ago. I'll take the goblin tax collector. I'll take all of them.