Large language models are general-purpose machines programmed in plain language, yet a single model now answers to households, classrooms, congregations, and civic groups whose ideas of harm, propriety, and authority can be sharply divergent. No one set of rules will apply across communities and cultures. We are building an open, inspectable registry that lets any community, from a single family to a school district or city, express its values as rules its AI follows. Each rule can be read, contested, and amended by the people it governs, and rules can nest the way communities do, so that, for example, a classroom can refine a school’s policy without overriding it.
The registry can be represented and developed in plain English. No programming is required. It can thus be easily created, shared, modified, and implemented by normal consumers rather than solely by expert organizations. It is modeled on open source principles with Wikipedia-like moderators.
AI itself can be used to ensure the safety, consistency, and efficacy of the intermediary. It can thus be dynamic and evolving, adapting to unintended consequences, as well as enabling unintended use cases. Through careful design of the system, it can be extensible while still being controlled.