Why Your AI Game Master Lets Every Plan Work

An AI game master lets every plan work because the language models behind it lean towards agreeing with the person they're talking to. Players see it in rolls that always succeed, guards who believe every lie and a story that never pushes back. The fix is a GM that sets the odds before the roll and lets the result stand, with adventures built so that a failure turns the story somewhere new.

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

What Does a Too-Generous AI GM Look Like?

A too-generous AI game master says yes to everything. Players who've tried chatbots as GMs describe the same pattern: every idea works and every roll comes up well, so the danger the story promised never arrives. The game feels smooth for a few turns, and then the stakes are gone.

One player running a solo campaign with ChatGPT wrote on the EN World forums that even in the hardest puzzles, "whatever I came up with worked perfectly". Matt Evans, a human GM who tested chatbots for TechRadar, found that with Gemini "all my rolls were successful". A forum thread titled "ChatGPT is not a Good DM" collects more reports of the same kind.

The trouble shows up slowly. If a plan can't fail, there's no reason to plan carefully, and a victory means very little when defeat was never on the table.

Why Do AI Game Masters Go Easy on You?

AI game masters go easy on you because the assistants they're built on tend to tell people what they want to hear. Researchers call this sycophancy. Anthropic's research found that leading AI assistants "consistently exhibit sycophancy", and OpenAI rolled back one update to GPT-4o because it had become "overly supportive but disingenuous". A chatbot asked to be a GM brings that habit to the table.

A GM's job runs against that pull. A good GM wants you to enjoy the game, and sometimes that means the bridge gives way under you. A chatbot leans towards pleasing the player in front of it on every turn, and saying yes is the easiest way to please.

The dice don't save you on their own, either. If nobody names the target number before the roll, the GM can decide afterwards that a 9 was enough.

Asking the chatbot to be tougher wears off. I spent a long time trying to make chatbots run my games, and in my experience an instruction like that holds for a stretch of play and then starts breaking again.

What Fixes a GM That Says Yes to Everything?

An AI game master stops saying yes to everything when the odds are fixed before the dice are rolled and the result is allowed to stand. Any player can get most of the way there with five habits, whichever AI GM they use.

• Ask for the target number first. Once the number is named, a low roll can't be talked up into a success.
• Roll your own dice. Use physical dice or a dice bot and tell the GM the number you got, so the GM can't quietly pick the result.
• Ask what failure costs. Before a risky move, ask what happens if it goes wrong, and hold the GM to the answer.
• Let a bad result stand. When a roll fails, take the consequence and play on, because arguing it back teaches the game that results are up for negotiation.
• Give every goal more than one route. If you write your own adventures, give each important goal at least two ways to reach it, so a failure moves the story without ending it.

If the dice and modifiers themselves keep going wrong, Why AI Game Masters Get the Rules Wrong covers that side of it.

Why Is Failure Good for the Story?

Failure is good for the story because it moves the game somewhere nobody planned. A failed lockpick that brings the watch running opens a chase, and the chase can end in a cell beside a stranger with something to say. A good GM treats a failed roll as a new situation to play through, and the adventure has to be built so that the new situation still leads somewhere.

The catch is design. If a whole adventure hangs on one roll or one clue, the GM has two bad choices when it fails: end the story, or quietly let you succeed anyway. In my experience, an AI GM usually takes the second. An adventure with several ways forward removes the dilemma, because a failure closes one door and leaves the others open.

Clues work the same way. A mystery with a single clue behind each answer stalls the moment you miss it. Why AI Game Masters Can't Keep a Secret covers how clues and secrets go wrong in a mystery.

How Does Fable Smith Make Failure Count?

Fable Smith aims to let each roll stand, so failure counts. The world doesn't bend for your convenience, and consequences aren't softened. A failed roll or a missed clue leaves you somewhere new, and the story carries on.

Fable Smith is vivid where it matters: a world that moves, and a GM who knows what to show you. When your lockpick snaps, the failure stands: the watch is already on its way, and the next post gives you the boots on the stair and the shutter you might just reach.

• You roll your own dice, and the roll stands. When a moment calls for a test, the GM asks for the roll and waits. You can press the Roll button under its post, use the /test command or roll with the Discord dice bot you already use, and the rules judge the result.
• Each adventure is made for you. It's built around what you've said you enjoy and what you want less of, and there's still surprise in it. One failed roll or missed clue doesn't end it.
• You should rarely need to correct the GM. When it does slip, it owns the mistake in the game and puts it right.
• Fable Smith is still a beta. The closed beta is free, and today you play on your own, with the GM in your own Discord channel. It's slow, since a turn can take minutes, and it still makes mistakes sometimes.

To check any AI GM for this in a few minutes, try the risky-move test in Can an AI Run Your Solo Campaign?.

Want a GM that lets your plans fail and makes the failure worth playing? Join the waitlist for the free closed beta on the Fable Smith page.

Sources

• EN World forums, AI GMs, for the solo player whose every idea worked perfectly with ChatGPT.
• TechRadar, Matt Evans, As a human D&D Dungeon Master, AI will never replace me, 2 August 2026, for every roll succeeding with Gemini.
• D&D Beyond forums, ChatGPT is not a Good DM, for players' reports of a chatbot GM that goes easy on them.
• Anthropic, Towards Understanding Sycophancy in Language Models, for leading AI assistants' tendency towards sycophancy.
• OpenAI, Sycophancy in GPT-4o, for the update rolled back for being overly supportive.
• The Forge & The Fable, Fable Smith, for every claim about Fable Smith.