What is an AI game master?

Ask anyone who's played tabletop games what made their best game master great and you'll notice something: nobody says "the descriptions." The descriptions are table stakes. What people remember is the ruling that felt fair, the villain who remembered, the promise that came due three sessions later. The GM's real work was never the talking. It was everything underneath the talking.

That's the right frame for evaluating the software, too.

The four jobs

Narration

Describe the world, voice the characters, make the scene worth being in. This is the job language models were practically born for, and every product in the category does it at least serviceably. It's also, for exactly that reason, the least useful basis for choosing between them.

Interpretation

A player says "I want to talk her down" or "I check the ledger for anything odd." A game master has to understand the intent, not the keywords — and modern models genuinely do. This is the quiet revolution of the last few years: the parser problem that plagued forty years of text games, the tyranny of guessing the right verb, is simply gone.

Adjudication

Now it gets harder. When an action is risky, someone has to decide whether it works — and how it's decided is what separates a game from a story with buttons. A model asked to narrate an outcome will drift toward the dramatically satisfying one, because satisfying text is what it's optimised to produce. That's not adjudication; it's authorship wearing a referee's shirt. Real adjudication means the outcome is settled by something outside the prose — dice rolled before the scene is written, thresholds the narrator can't argue with. We've written up how our dice work, but the principle matters more than our implementation: if the storyteller decides the roll, there is no roll.

Memory

The hardest job, and the one where the category most often quietly fails. A human GM keeps notes. A language model keeps a context window, which is not the same thing — it's a horizon, and things fall over it. Worse, a model can assert in fluent prose that something happened which never did, and nothing in the architecture objects. The fix isn't a bigger window; it's an external record the prose can't overrule. That argument gets a full piece of its own.

A field guide to the species

Products calling themselves an AI game master cluster into four shapes, and knowing which one you're looking at saves a lot of disappointment.

Shape	What it optimises	What it drops
Character chat	The conversation continuing	Stakes, endings — it's a companion, not a campaign
Open sandbox	Total freedom, any setting	Consistency and consequence over long play
Rules emulator	Fidelity to a tabletop system	Accessibility — you bring the group and the patience
Structured engine	Consistency, stakes, endings	Freedom to leave the authored frame

None of these is the wrong answer. They're answers to different questions — the same taxonomy that runs through our survey of what's worth playing on a phone, where each shape has a best-in-class representative.

Six questions that sort the field

• Are the dice real? Ask whether the number exists before the prose does. If you can't tell, assume it doesn't.
• Does anything persist? Take a wound, close the app, come back tomorrow. Is it still there — as a number, not a vibe?
• Can you lose? A game you cannot lose is a story you're annotating.
• Does it end? Endings are what give middles their weight.
• Who wrote the world? Authored settings hold together; improvised ones hold up until they don't.
• What happens when you break a promise? The best single test of memory there is. If the answer is "nothing", you have your answer about everything else.

The human comparison, honestly

Is any of this as good as a great human game master? No — and it's worth being precise about why. A human reads the table. They notice you've gone quiet, that the joke landed wrong, that tonight needs to end on a win. That's not narration or bookkeeping; it's care, and no current system does it.

What the software offers instead is availability. It's there at midnight, for twenty minutes, with no scheduling and no social cost when you cancel. For the enormous number of people whose alternative isn't a great human GM but no game at all — the lapsed player, the reader who never found a group, the solo roleplayer graduating from oracle tables — that trade is not a consolation prize. It's the whole point.

Where we sit

Unwritten is a structured engine, and by now you can guess the pitch: authored worlds, dice rolled before the prose, state the narrator can't overrule, promises with deadlines, stories that end. The narrator writes the scene; the engine decides what's true. If the four jobs above are the job description, our bet is that the last two are where the product actually lives — and it's free to test whether we're right, fifteen turns a day.