AI transparency

What the AI does

Large language models handle the fiction: the Dungeon Master’s narration, non-player characters, interpreting what your character attempts, and deciding what happens in the story within the SRD 5.2 rules. AI also powers the preparation and practice tools: Scry’s scene forecasts, the Coach’s feedback on your play, character-concept generation, and the optional scene art and read-aloud audio.

What deterministic code does

Everything with a number attached is computed by our own code, never by a model: dice rolls, ability checks, damage, character generation math, leveling, validation of rules outcomes, and your plan’s usage limits. Your campaign’s canonical state (characters, hit points, quests, inventory) lives in a database the model can only read and write through defined, validated operations. The AI narrates; the code keeps score.

Fair dice

Every roll draws from the platform’s cryptographic random number generator, the same class of randomness used for security keys. Rolls are unpredictable and cannot be steered or reproduced, by us or by the model; the AI is never a source of randomness, and every outcome you see traces to a real roll made in code.

Fairness means honest odds, not good luck: a fair d20 still rolls a 1 five percent of the time, and a bad night of dice is still a bad night of dice. More detail: A fair dice roller for RPGs you can trust.

Memory and context

The Dungeon Master keeps three kinds of memory: the current game state (the canonical facts of your campaign), a rolling summary of the story so far, and a searchable store of specific past details it can recall when they become relevant again. Each campaign’s memory is isolated: nothing carries over between campaigns.

Recall is selective by design: the model works from a compact view of your campaign, not a full replay of every past turn, so a minor detail from long ago can occasionally be missed even though the canonical game state is never lost.

Providers and your choice of model

Hosted play runs on models from established AI labs, currently Anthropic and OpenAI, each matched to the job it does best. Higher plan tiers unlock stronger models. The current lineup as of July 2026 (specific models can change without notice as better ones ship):

• Dungeon Master and play: OpenAI GPT-5.4 mini keeps pace across a long session and reliably calls for dice rolls; pivotal scenes and higher plans step up to a stronger Anthropic Claude model.
• Character building, leveling, and Scry’s forecasts: Anthropic Claude (Sonnet 4.6).
• The hardest reasoning, such as the Coach’s rubric-grounded feedback: Anthropic Claude (Opus).
• Scene art and portraits: OpenAI image generation.
• Read-aloud voice: OpenAI text-to-speech.

You can also run on your own models: add your own Anthropic, OpenAI, or Google API key, or point your account at a local Ollama-compatible endpoint, and supported text generation runs there instead. Your keys are stored encrypted and used only to make model calls for you.

Playing without the AI

The AI Dungeon Master is a choice, not a requirement. When you create a campaign you decide who runs the game, and picking yourself takes the model out of the DM’s chair: you narrate and adjudicate, and the app is a virtual table that rolls the dice in code, tracks character sheets, hit points, conditions, rests, death saves, XP, and combat rounds, and keeps one shared transcript for everyone at the table. Those controls run through the same engine as everything else, with no model call behind them and no draw on your AI allowance. The host can also hand the chair over mid-campaign, in either direction, without losing any play state.

In a campaign you run, the AI appears only in roles you add on purpose: a companion character or a single named NPC, each speaking only when you invoke it, and each scoped to what that character would plausibly know rather than to your notes. Define none and the model stays out of your game. Multiplayer, friends, presence, and chat work the same either way.

Privacy and data handling

To produce responses, your prompts and the relevant play context are sent to the AI provider serving your session. Providers process that data on our behalf under their own terms and are not permitted to use it for unrelated purposes. Your campaign data is protected by per-user database access controls, and product analytics run only with your consent and never contain your play content. The full detail, including every subprocessor, lives in the Privacy Policy.

Known limitations

AI output is generated automatically and can be inaccurate, inconsistent, or unexpected; it is not professional advice. The deterministic layer bounds the mechanics, but narration can still err, a rules interpretation can be imperfect, and the product is in beta, so behavior changes as it improves and play data may occasionally be reset between updates. Review anything generated before you rely on it.

Reporting a problem

If the AI behaves incorrectly, or you have a privacy or security concern, use the in-app feedback control (bottom-right when signed in) or email support@thedungeontrainer.com. Reports go straight to the team.