Why AI Game Masters Can't Keep a Secret

An AI game master can run a fair mystery only when the answer is decided before play and stays hidden until you find it. A chatbot told to be the GM usually manages neither. It makes the culprit up when you ask, or it knows the secret and lets it slip. Either way, your deductions stop meaning anything. An adventure prepared in advance, with its answer fixed and its clues placed on purpose, is where a fair mystery starts.

This post is part of our guide, Can an AI Be a Good Game Master?.

What Happens When an AI GM Runs a Mystery?

When an AI game master runs a mystery, the result is often a solution that doesn't fit the clues, or a secret that comes out too early. Matt Evans, a human GM who tested ChatGPT and Gemini as game masters for TechRadar, points out that a mystery needs its clues seeded by working backwards from the answer. Of that step, he writes: "ChatGPT does not do this."

Without that backwards planning, a mystery has clues that point nowhere. You find a torn letter and a witness who saw a cart leave at dawn, and both sound important. When you finally accuse someone, the GM decides in that moment whether you're right, and the answer tends to be whatever makes the best scene. Your reasoning had nothing to be right or wrong about.

The other way a mystery breaks is a secret that gets out, and the player doesn't even need to try. A GM holding the answer can hint too hard, let a character mention something they couldn't know, or sum up the plot in a way that gives the twist away. Researchers are testing deliberate attacks too: a paper by Shiomi and colleagues asks whether characters run by a language model can be tricked into revealing backstory meant to stay hidden. I tried for a long time to make an AI GM work, and in my experience it struggles with both halves of the job: designing the mystery and keeping it.

Why Can't an AI GM Keep a Secret?

An AI game master struggles with secrets because a chatbot has to keep a secret through every reply it writes. With no answer decided in advance, the culprit becomes whoever suits the current scene. With an answer decided, the secret is always close to hand, and every reply is a fresh chance for a hint to go too far.

Keeping a secret is a judgement made on every turn. What does this person know? What has the player earned? How much does this clue give away? From my own attempts, I'd say that's the same judgement about what matters right now that makes AI GMs forget, turned towards what to hold back. A chatbot that can't reliably pick the important fact out of its notes also struggles to keep the important fact out of its replies.

Setting up the mystery yourself makes this harder. At a table, the GM is a person who carries the secret. When you set up a mystery for a chatbot to run, the secret can only live in your own head, which spoils it, or in the chat, where the GM might show it to you.

What Makes a Mystery Work With Any GM?

A mystery works when the answer is decided first, the clues are built backwards from it, every conclusion has more than one route, and the answer sits somewhere the player never reads. These habits from human GMing work with any AI tool.

• Decide what really happened before play. Write down who did it, how and why, and treat that answer as fixed.
• Give every conclusion more than one clue. Players miss clues, so leave each piece of the answer in more than one place.
• Run a published mystery you haven't read. Nobody can design a secret and also be surprised by it, so if you're setting up the game yourself, give the GM an adventure you've never seen, and ask it to reveal each clue only when your character finds it.
• Tell the GM what each person knows. A short line per character, such as "the cook saw the cart leave and has never seen the letter", helps the GM keep each person to what they could know.

These steps help, and with a chatbot the answer still sits in the same conversation the GM writes from, so expect the odd hint that goes too far.

Does Fable Smith Keep Its Secrets?

Fable Smith runs adventures with real secrets and a real answer, and you find them by playing. The GM doesn't give them away, though it still makes mistakes sometimes, and when it slips it owns the mistake in the game and puts it right.

Fable Smith is vivid where it matters: a world that moves, and a GM who knows what to show you. The cook who saw the cart leave keeps her own counsel until you give her a reason to talk, and the culprit goes on covering their tracks while you investigate.

• Each adventure is made for you. It's shaped by what you told it you enjoy and what you want less of, and its mystery has one true answer that's still a surprise to you.
• The world holds together. Names, places and facts stay put, so a clue from three sessions ago is still true when you set it beside today's.
• You rarely need to correct it. When the GM does slip, it owns the mistake in the game and puts it right.

Today you play on your own, with the GM in your own Discord channel. Fable Smith is in a free closed beta. It's slow, since a turn can take minutes, and it still makes mistakes sometimes.

Memory matters for a mystery too, and Why AI Game Masters Forget Your Campaign covers it. A GM that plays your character for you can solve the mystery on your behalf, which Why Your AI Game Master Keeps Playing Your Character looks at next.

Want a mystery whose answer was decided before you started guessing? Join the waitlist for the free closed beta on the Fable Smith page.

Sources

• TechRadar, Matt Evans, As a human D&D Dungeon Master, AI will never replace me, 2 August 2026, for a mystery's clues being seeded by working backwards, and ChatGPT not doing so.
• arXiv, Shiomi et al., Tricking LLM-Based NPCs into Spilling Secrets, for research into whether characters run by a language model can be tricked into revealing hidden backstory.
• The Forge & The Fable, Fable Smith, for every claim about Fable Smith.