As an introduction to how our studio makes decisions about AI, I want to say that I understand that for some people, this will be the point where they step away from the project. And that’s “Fair ’nuff.” The AI topic is far too complex and ethically snarled for anyone to make any decision except for themselves. If our vision has a different alignment than yours, then I respect that we can’t share the same table, and I wish you much joy in your adventures wherever you find them.
The creative heart of Grimmloch is human. The lore is ours, written by us and used in our online roleplay community for the better part of seven years. The stories are yours. The 3D world is built from assets we buy from working human artists. Where we do use AI, the use is narrow and deliberate, and every decision is held to an ethical line with three parts: ethically sourced training data, environmental impact, and protecting your writing from ever being used to train a model.
Our line covers three categories, and every AI decision in this document was made against it.
The first category is ethically sourced training data, and it is about careful selection of the tools we use. This is why we work with Mistral for our language model, because we consider them to be an ethical company.
The second category is environmental impact, and the third is protecting your writing. Both of these are addressed by the same decision: we self-host the model ourselves. Self-hosting on Mac hardware carries roughly one tenth of the environmental impact of running the same model in a data center. It also means that what you write never leaves our environment, and it cannot be used to train anything.
There are three places a language model appears in the game.
People who are new to roleplaying often feel self-conscious about not knowing the lore, or about whether their writing is good enough to join in. The onboarding assistant lets a new player learn the backstory at their own pace and practice their writing without an audience, so they can enter the community without that sense of feeling inadequate. The lore it teaches from was written by us, not by an AI.
Most games handle player action through menus of pre-written responses. The parser reads what you actually wrote, which lets us drop the “choose your response” menus entirely and lets you write your character’s own words. In this role the model produces no writing of its own. Its entire job is to understand yours.
We use a language model to give an NPC its voice and personality. Every decision the NPC makes is made deterministically by our game systems. The model’s only creative output is the roleplay post in which the NPC speaks.
Our lore is not new. Its predecessor Grimmwold, our private Ultima Online roleplay server, ran from 2007 to 2015 and was the most popular of our world-building fictions.
The stories in Grimmloch are yours. We are not writing a story for you to consume, and no AI is either. Our job is to build the world and the systems that host your writing, and as stated above, your writing is never used to train a model.
We buy our 3D assets from working human artists, mostly on the Unity Asset Store, because paying real artists for their craft is the right way to build and it keeps our budget in the hands of independent creators. We do not use AI-generated 3D assets except in extremely rare cases, and to date we have done exactly one.
Our music asset library was purchased years before generative AI existed, some of it five and ten years ago, so its human origin is beyond question. And our first call for new music is always to license work from human musicians.
We would rather tell you what is undecided than project false certainty, so here is where things stand.
The first thing we intend to build is a way for artists working on commission to register with the game and produce custom character art for players. We will verify that these artists are actual working artists, because the commission space is full of scammers who generate images with AI and claim the work as their own, and we want the artists you hire to be real ones. For players who cannot afford commissioned artwork, and for NPCs, we are in negotiation with an AI art model that is trained entirely on public domain and licensed assets and that we can self-host. Nothing is final until it meets all three parts of our line.
Many of you love the idea that the stories you write might be turned into small songs you could overhear in a tavern or from a roaming bard. We love it too, because it is your own adventure showing up in the world. But we have not yet found a music model that meets our ethical line, and if we never find one that does, this feature will be cut. The songs matter to us, but the line matters more.
If any of this raises questions, ask them. This document exists because you deserve to know where we stand, and we would always rather explain a decision than have you wonder about it.
Have a question or comment? Join the discussion on our Studio Discord in the #design-discussions channel.