
How AI-narrated campaigns handle unexpected player decisions
The moment the plan falls apart
Imagine a siege scene. The campaign's author expected you to sneak through the sewers, bargain with the quartermaster, or fight at the gate. Instead, you type: "I climb onto the barrel cart, shout that the granary is on fire, and hide in the confusion."
Nobody wrote that option. There is no "shout about a fake fire" button, and no climbing rule for barrel carts. A rigid system would shrug and tell you to pick from a list. A good game master does what a person at a tabletop would do: take the idea seriously, work out what it would plausibly cause, and keep the world coherent while it happens.
That moment is where AI-narrated RPGs tend to win or lose your trust, so it deserves a close look.
The skeptic's question: does the AI just say yes to everything?
The honest concern about AI narration is a pair of opposite failures. Either the system refuses anything outside its script, which feels like a menu with extra steps, or it agrees to everything, which means nothing you do really matters and the world bends to whatever you type.
Fabulator's design aims between those two. It treats your input as something to translate, and it keeps a firm line around the things that must stay true.
Translator, not restrictor
When you type freely, the first step is an intent parser. Its job, in the design documents' own framing, is to be a translator rather than a restrictor. It takes "threaten the merchant with a club" and structures it into clauses: who does what, to whom. It does not shrink your options to a fixed list of verbs.
The parser also separates what your character says from what they mean. If you tell the merchant you will help while secretly stalling so your allies can surround the caravan, those are recorded as two different things: the surface action and the true intent. The surface action is what the world sees. The hidden intent only matters to NPCs who could actually observe it. That keeps bluffing and deception playable instead of flattening everything into the literal reading.
Compound actions work too. "I swing at the guard while shouting for my allies to flank" can be split into a combat clause and a social clause, resolved separately, then merged. If the two conflict, precedence rules decide, so a swing that drops the guard interrupts the shout aimed at him.
What if there is no rule for it?
Your barrel cart climb is a good example. If the world has no climbing module, the action is not rejected. It goes to a pass-through resolver, and in narrative-first play the outcome can be judged lightly and handled by the narrator. If the world does use a ruleset, a mechanical check can apply instead, such as a skill roll against a difficulty.
Either way, the system is designed so a missing rule does not become a dead end. Validation is described as permissive: it rejects actions that break world logic, such as casting magic in a zone where magic does not work, or that threaten the integrity of the data. It does not exist to block clever or social solutions.
Where the world pushes back
Creative freedom is only fun if the world has edges. Fabulator treats the language model as a proposer. Whatever it suggests, whether a state change or a new character, is checked before it is committed. Changes have to stay within sensible bounds, so an ally cannot flip to an enemy in one turn, and they have to respect canon, so the dead stay dead. A character cannot suddenly know something they never learned.
So your fake fire alarm can work, and the world will respond. Guards may run toward the granary, the quartermaster may remember who started the panic, and the story can carry that forward. What you cannot do is type your way into winning the war outright by declaring that the enemy general drops dead.
When the system is not sure
Some inputs are ambiguous or carry large consequences. For turns with big proposed changes, irreversible canon edits, or low confidence in the reading, the design includes a confirmation step. The system tells you what it understood and lets you continue, edit or cancel. It is meant to be rate-limited, so ordinary play is not constantly interrupted.
If a proposal fails validation, the system regenerates it with the problems noted, up to a small number of attempts. If it still does not work, you are asked to rephrase rather than being silently handed a broken result. Your text is preserved so you can adjust it.
Why this matters for the story
A game master that handles surprises well makes your choices feel like they count. Characters or places the narrator invents along the way can become part of the campaign's canon, and are checked against existing lore first, so the world grows from your decisions instead of contradicting them.
Try a surprise yourself
The best test is to play and do something nobody planned for. Browse the World Seeds on the Discover page, choose a setting that fits your taste, and start a campaign. Then ignore the obvious path and see how the world answers.
