Abstract

We describe a method for building a human-reviewed religious commentary corpus from transcripts of conversations between users and a conversational AI agent grounded in a scriptural canon, instantiated for the Pāli Buddhist canon. Components: (I) a domain-grounded conversational agent that declines claims of religious, sacramental or ordained authority and directs users to human teachers; (II) a chat, voice or avatar dialogue surface; (III) a consent-based recording layer in which every conversation is private by default and the user is asked, per conversation, whether it may be considered for publication at one of four tiers (attributed, pseudonymous, anonymous, excluded), non-published conversations contributing only to differentially private aggregate statistics kept outside the corpus; (IV) multi-label classification of published transcripts into three categories mirroring the canon's divisions: teaching dialogue, conduct guidance (never binding rule) and systematic analysis; (V) a staged human review pipeline (raw, reviewed, ratified) run by an expert advisory body of recognized religious authorities independent of the developer, with periodic ratification sessions. Consent withdrawal deletes a per-conversation encryption key, erasing the transcript from primary, index, backup and derived stores; it cannot recall third-party copies, as the user is told at opt-in. Ratified content is retracted by annotation rather than deletion, and consent withdrawal always overrides ratification. Ratified material is commentary, never scripture. Also disclosed: periodic model audits, model-version provenance labels on content, a cooling-off period for opt-ins made in crisis, and adaptations to other religious traditions. Stated limits include privacy untested at scale, contested authority and dependence on model quality.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS