Skip to content

How Nexus uses AI

We're a handful of developers who run tables together. Nexus uses AI. If that makes you wary, good. This page is for you. These are the lines we draw for ourselves and the record we expect to keep when they change.

The short version

You run the game. Nexus carries bookkeeping. AI helps in narrow places, such as drafting a name or finding a rule, and only when you ask. The one exception is the session recap: with voice transcription on, Nexus writes it when the session ends and shares it with your table, unless you set your campaign to hold it for your read first. Either way you can correct it any time. Nothing else enters your campaign unless you accept it, and the product remains useful with AI switched off.

Five commitments

  1. The GM owns the story. AI suggestions are drafts to take, edit, or throw out. The session recap is the one thing we publish on our own, and you choose how: share it when the session ends, or hold it for your read first. You can correct it any time. Nothing auto-runs your game.
  2. We never train generative models on your content. We never train generative models on your campaign, compendium content, or anything else we hold; our providers process API requests under terms that prohibit training on that traffic. If you opt in, your sessions help tune small classifiers that decide when the co-pilot spends an AI call. The one carve-out is BYOK: that traffic runs under your own agreement with your provider, not ours. Read your provider's terms before connecting a key.
  3. Your campaign stays yours. An AI request carries the context it needs and nothing else. Unless you opt in to the small classifiers, we tune features on aggregate outcomes, not campaign content. Each request's prompt and response are also recorded in PostHog, our analytics processor, so we can debug failures. Those records never train a model.
  4. AI costs are visible. Usage is measured in the product. The Steward plan bundles a shared pool of credits into its flat monthly price with no overage charge; BYOK subscribers run their own Anthropic, OpenAI, or Gemini key and pay that provider's cost with no markup.
  5. AI can be off. Campaign-level switches govern text, image, and transcription features. Compendium, maps, campaign management, and live play continue without an AI request. The campaign setup screen asks whether your table uses AI helpers, with no answer picked for you.

Small classifiers, and what opting in means

A classifier is a small model that does not write anything. It reads a short stretch of table talk, the last few lines, and answers one narrow question with a score. Is the table chatting about pizza, or playing? Is a roll about to happen? A generative model, the large language model behind suggestions and recaps, does the opposite job: it writes text.

We put classifiers in front of the co-pilot's automatic calls so it can skip the expensive call when nothing at the table needs it. Scoring a stretch of talk costs a fraction of a cent, far less than one language model call. Fewer wasted calls keep costs down, which keeps the free tier useful, and mean fewer interruptions for the GM.

On our own playtest, the classifiers skip about half of the roll suggestion calls and about 1 in 9 of the ambient suggestion calls, while still catching 95% of the moments that matter. They let a share of the calls they would skip through anyway, and those are how we check they are not missing things. If one starts missing moments at your table, it backs off for your campaign on its own.

Less energy. A classifier decision, plus the step that turns words into numbers, is a small fraction of the compute a large model spends writing a reply. Every call skipped is one the large model never runs, so a session spends less energy as well as less money.

Where your words go. The scoring runs on our own servers. Turning a stretch of talk into numbers uses a small embedding model from a third party (Google's Gemini embedding model), far lighter than a full language model call and under API terms that bar it from training on our traffic. When a classifier says skip, that stretch of the session is never sent to the larger model at all, so less of your table talk leaves our systems.

Opting in. Training these classifiers is off unless the GM turns it on in the campaign's settings. If you opt in, transcripts of that campaign's sessions and your moment corrections are used only to train these classifiers: never a generative model, never sold, never used to write content. We fit the classifiers on our own machines, and they ship as numbers in our code with no transcript text inside. Preparing that training data runs each stretch of talk through the same embedding model, under the same terms.

Without opting in. Your campaign still tunes its own classifier. We keep a log of what it decided and whether a skipped moment turned out to matter, with no transcript text in it, and use that only to loosen or switch off the classifier for your campaign. Nothing in it trains anything.

Opting out. Turn it off any time. From then on your campaign is left out of every training export, earlier sessions included. A classifier already trained keeps what it learned until the next retrain, which leaves your campaign out.

Tell your table. Players' voices are on those transcripts. Let everyone know before you opt in, the same as you would for recording.

Check our work, don't just trust it

Every AI-generated image produced since September 2026 ships with a signed C2PA provenance manifest recording that it was AI-generated, which provider and model made it, and — where the image was generated from a licensed artist reference — which artist's reference it was derived from. That manifest is checkable by anyone who reads it back out of the file, not just asserted by us. See Security if you find one that doesn't hold up. Images made before September 2026 are labeled but carry no signature.

AI-generated content is labeled where you see it. You always know when you're talking to the co-pilot instead of a person, and public compendium images generated by AI stay visibly marked as such until a human artist supersedes them.

Narrative Nexus is listed as a conformant Generator Product on the C2PA Conforming Products List. To check any file yourself, drop it into the Content Credentials verifier.

We built that signing pipeline before the EU's AI-labeling requirement took effect for us, because a product that touches art and text needs provenance on its own merits, not because a regulator asked. That's an engineering commitment, not a legal certification — no such certification exists to hold, and a signed manifest is one piece of what the labeling rule asks for, not the whole of it.

Tools are not creators

A tool that finds a rule or drafts a recap leaves creative decisions with a person. We do not want an AI game master, AI players, or a substitute for the person behind the screen.

We do not sell AI-generated art. Our compendium uses original writing and openly licensed material, with attribution shown where it appears. Books we do not have the right to use are not fed into a model or reproduced in the product.

Human art has a different boundary

Public compendium placeholder images that are AI-generated remain visibly disclosed. They may only be superseded through approved human-art controls; a generated image is never silently presented as a human-made commission.

When an Artist Network commission is delivered, it is signed with a credential that names the artist, and deliberately carries no AI digital-source-type — the platform did not watch the work get made, so it does not assert how it was made. We mark what's AI so what's human stays credible.

Artists can tell us how they want to participate in the Artist Network interest intake. That intake is not an account, commission offer, license, or request for artwork. It does not grant us training, reference, publication, or other rights in an artist's work. Invited beta artists can keep portfolio pieces, drafts, experiments, references, and works in progress private in Studio unless they later approve an exact controlled public derivative.

Artist Network interest intake

Why use AI at all?

Some narrow uses help people stay in the game. A running recap lets a table pick up the thread after a missed week. Transcription helps when speech moves faster than attention. Rules lookup keeps play moving. None of that replaces a creator. It clears the path back to the table.

We use the smallest practical model, retrieve instead of regenerate, and cache aggressively. The campaign master switch turns off that campaign's AI features, and the rest of the product keeps working. It is campaign-scoped, not account-wide: safety checks on uploads and features outside any campaign keep running.

Questions we expect

Did you scrape sourcebooks? No. The compendium is original writing and openly licensed material. Full attributions are on our Legal and Licenses page.

I made a game you carry. Can I keep AI away from it? Yes, feature by feature. Our policy for creators lists every AI feature that can touch a game's text, what each one does, and which games have already switched which ones off.

What leaves my campaign for AI? The content the request needs, sent to a provider for that request. A copy of the request and response is also recorded in PostHog, our analytics processor, so we can debug failures. It never trains a model. See the Privacy Policy for processors and rights.

Do you learn from my campaign? From aggregate signals about whether a suggestion helped, not from campaign content. The PostHog records are for debugging and are not used for tuning. Any study of real content would require a separate, explicit opt-in.

What about environmental cost? It is real. We do inference only when requested, use the smallest practical model, cache aggressively, and keep a useful AI-off product instead of treating offsets as an answer.

Does the C2PA manifest make you compliant with the EU's AI rules? No. No such certification exists to hold. The manifest is a real, checkable signal we built ahead of the requirement — evidence of how we work, not a legal seal.

Will this page change quietly? No. Every substantive change gets a dated entry below.

Changelog

  • 2026-11-20: Updated the GM-owns-the-story commitment: the session recap still shares with your table when the session ends by default, and a GM can now hold it for a read first, then share it or let it go out on its own after 24 or 48 hours, or never.
  • 2026-11-15: The campaign setup screen now asks whether to use AI helpers at the table, with no answer picked for you. Leaving it unanswered keeps AI on, and you can switch it off any time in settings.
  • 2026-10-19: Added an opt-in, off by default: with the GM’s say-so, a campaign’s sessions help tune small classifiers that decide when the co-pilot spends an AI call. Explained how they work in a new section. We never train generative models on your content. The classifiers turn table talk into numbers with Google’s Gemini embedding model, which reads the table better than the OpenAI model they started on.
  • 2026-09-26: Corrected the GM-owns-the-story commitment: the session recap is written and shared with your table when the session ends, not held as a draft for your approval. You can correct it afterward.
  • 2026-09-19: Disclosed that each AI call is recorded in PostHog, our analytics processor, for debugging, and corrected the master-switch and data FAQ answers to say what the switch does not stop.
  • 2026-09-08: Every AI-generated image is now signed with a C2PA manifest by a trust-listed certificate, not just licensed-reference artwork. Added our C2PA Conforming Products List listing and a link to the public Content Credentials verifier.
  • 2026-08-22: Added the provenance section: signed C2PA manifests on licensed-reference AI artwork and Artist Network commission deliveries, and a BYOK carve-out on the no-training commitment.
  • 2026-07-27: Added per-ruleset AI feature switches and the creator policy page at /creators. Stonetop is switched off across the board.
  • 2026-07-24: Added the Artist Network interest path and the boundary between private artist materials and disclosed public compendium placeholders.
  • 2026-07-23: First published.