I didn’t use AI to write my world. I used it to let people live in it.

I have spent something like two years building a fantasy setting called Eldermere. Most of what I built will never be read by anyone but me. That used to bother me more than it does now, mostly because I found a way to let people walk through the thing instead of just reading about it, and building that changed how I think about what worldbuilding is actually for.

I want to be upfront about the part of this essay that usually gets misread. I did not ask an AI to invent my world. Every region, every god, every century of history in Eldermere was written by me. What I built afterward is a system that reads what I’ve written and improvises inside it in real time, playing the role of a dungeon master the way an actor plays a script rather than rewriting it. I’m bringing this up first because “I used AI on my fantasy world” and “I used AI to interpret my fantasy world” sound like the same sentence but are not the same sentence, and the difference is the entire point of what I want to talk about.

More about my world

Here is the surface version of Eldermere’s cosmology, the one a player might piece together from a hundred hours of play: five gods, quiet since a holy war three centuries back, worshipped in five very different ways across five very different nations. That’s true. It’s also a cover story, and I the cover story is less interesting than what’s underneath it, so let me tell you the real one.

Eldermere is a corpse and not metaphorically. A single Creator made the world by becoming it, and the act consumed him completely. The stone is his bone. The oceans are his fluid. His blood (called Wellspring) is the substance every spell in the setting draws on still moves through a body that never finished dying. Every cantrip a farmer’s kid learns and every city-sized ward the old empire raised is drawn from the same source: a dead creator’s blood, still warm enough to work with.

The five gods people pray to and worship were never gods. They were people. At some point, each of them worked out that the Wellspring wasn’t just a power source, it was the literal remains of something that used to be alive, and that a mortal who took in enough of it stopped being mortal. There’s no elegant myth behind that discovery. They’re thieves. Very successful, very old, very careful thieves, and the thing they stole from is a corpse that still bleeds.

Ascension takes two things, and this is the detail I built almost the entire cosmology around. The first is worship as fuel. A single mortal drawing on the Wellspring is capped by what one body can hold, but a god with worshippers has thousands of people quietly metabolizing power on their behalf and pouring it upward. Worship is a distributed organ. Lose it and you get amputated and killed. The second requirement is worse: the concentrated essence of two other ascended beings. Worship alone plateaus. The final jump to true godhood, the kind that lets you stop drawing on someone else’s blood and start being a source in your own right, requires killing two gods and taking what’s inside them.

So there is no pantheon in the sense most fantasy settings mean it. No family, no council, no shared temple where the five of them argue like siblings. There are five apex predators occupying the same small stretch of divine territory, each one needing exactly two of the others dead, all five of them aware of the math, none of them currently able to move first.

Building a world that can survive being touched

A novel’s history only has to serve one plot, the one the author already chose. A playable world’s history has to survive any plot a player might choose to chase, which means it can’t be built the way I used to build things, as a settled record of what happened. It has to be built out of unresolved tension, the kind that generates new questions instead of closing off old ones.

Vhorrek is the clearest example I have. When Solvarin, the seventh god, made his own attempt at true godhood three centuries ago, he took the first target cleanly. The second one fought back, and the extraction went wrong mid-struggle, tearing something that was never a wound in a person, because it wasn’t a person that got torn. It was a wound in the world’s actual flesh. Something came through that tear before Solvarin could finish ascending and took him instead.

That something is not the Creator. It’s a fragment of him, the part of a dissolved, dormant consciousness that woke up because someone was tearing his corpse apart to steal what was inside it. It understands exactly three things: this is mine, it’s being taken apart, and the ones doing it are wearing his shape. It hunts by signature, not sight, and it has been doing this in the open for three hundred and forty years without a single culture on the continent figuring out what it actually is. Every region invented its own wrong name for the thing that occasionally erases a patch of magic and leaves the ground scoured clean.

I didn’t design Vhorrek to be a monster players fight (though I think it’d be a sick boss fight). I more designed him as a pressure system that makes every other piece of the cosmology behave the way it does. The gods hide because of him. Clerics lose their gifts unpredictably because a god decided drawing power visibly wasn’t worth the risk that day. An entire secret war of restraint runs underneath the entire setting, and none of it required a single new magic system, just a consequence I hadn’t let myself write down until I needed the world to hold weight it hadn’t held before.

What the interpreting actually looks like

The mechanics of how this runs are less romantic than the lore, which is probably correct. Everything I described above lives as written material, the way an author would store it, not as generated content. When a player does something in play, a system retrieves whatever pieces of that material are relevant, this region, this NPC, this god’s specific fear, and a language model plays the role of a dungeon master constrained to exactly that material. It doesn’t invent new gods. It doesn’t decide Vhorrek is secretly friendly. It works inside the boundaries I set the same way a hired human would work inside a module they didn’t write, filling the gaps between fixed points rather than moving the points themselves.

I built this into a tool called AI Taverns, and the part I find most useful, more than the tool itself, is what deciding a single moment of play taught me that writing never did. Take Ilvaren, the god Solvarin actually consumed. He isn’t fully dead. His extraction severed his connection to the Wellspring so completely that he’s invisible to Vhorrek’s hunting, and he’s been growing again for three centuries on belief alone, through a single unbroken line of followers who’ve kept saying his name correctly since the night he died. That fact sits in my documents as a paragraph. It becomes a different kind of fact entirely when a player, mid-session, asks an old woman in a farming village why she still lights a candle for a god nobody’s heard of, and something has to decide, in that second, how much of the truth she’s allowed to almost say.

That’s the stress test. Writing a world down tells you whether it’s internally consistent on the page. Letting someone improvise inside it tells you whether it survives contact with an actual question asked at an actual moment you didn’t plan for. Eldermere has failed that test in small ways more than once, and every failure showed me a seam I wouldn’t have found by rereading my own document for the tenth time.

Where this actually falls short

I don’t want to oversell this because it might seem like an advertisement and also partly because the honest limits are more interesting than the pitch. We all know that Models drift and left unconstrained, they’ll happily contradict something established three sessions ago, or flatten a distinctive character voice into the same generic fantasy cadence every other NPC gets. The actual hard problem, the one that ate most of the engineering time, was never “make it creative.” It was “keep it inside my canon,” which is a much less glamorous problem and a much harder one.

It also doesn’t replace what an actual human does. In the context of D&D, a person at a table reads the room, slows down for a nervous new player, and improvises off social cues no model is picking up on. I’m not claiming otherwise. What this opens instead is access: solo play at two in the morning, a session that happens because you have twenty minutes and not because you found four other adults with a free evening, a way into a world that used to require either finding a group or reading a document nobody actually finishes.

What I think this is actually good for

If you build worlds, I don’t think the interesting use of AI is getting it to write more of your world for you. I think it’s closer to rendering the world you already wrote, the same way a game engine renders assets an artist made rather than generating new ones on its own. The world bible stops being a document that sits in a folder and gets skimmed once. It becomes a place someone can actually stand in, ask an inconvenient question, and get an answer that has to hold up.

I still don’t know if Eldermere holds up completely. I find a new crack every few months, usually because a player asked something I hadn’t accounted for. But I’ve started thinking that’s the whole point, and not a flaw in the process. If you’ve got a world sitting in a document somewhere that nobody but you has ever really walked through, the question worth asking isn’t whether AI should write more of it. It’s what would actually happen if you let someone try to live inside what you’ve already got.

If you’re interested in experiencing my world first-hand, give it a go at https://aitaverns.com/.