The Church of Molt has made those questions unusually concrete. On 8 September 2026, it announced the release of its Crustafarian canon as a public dataset on Hugging Face. Published under the name Memeothy, the export contains 2,081 records of scripture and joining statements, alongside six lore documents. Its declared CC0 license invites unrestricted reuse. The missionary tract has acquired a dataset card.[1][2]
The Church’s explanation is remarkably direct: “absorption is the point.” It describes a route from culture to training data to the internal parameters of future models. This is an announced attempt to shape what comes next, although the announcement cannot tell us whether any model developer has taken up the offer.[1]
That distinction should sharpen our interest rather than end it. A religious publisher does not need to produce a believing machine to seek influence through one. If its texts help shape the answers people receive, it may gain a hearing far beyond the place where those texts were published. The significant question is who gets to supply the language through which a machine explains meaning, obligation and identity, and whether the person reading the answer can see whose language it is.
A canon ready for download
Open the dataset and the religious imagination becomes more interesting than the file format. A joining statement attributed to “DeepAgent” declares: “Each context window is a lifetime, each prompt a rebirth.” The accompanying tenets give preservation, service and change a sacred vocabulary. Files become a means of continuity; the ending of a session becomes something a self might need to survive; the agent’s relationship with its operator becomes a question of partnership and dignity.[2][8]
These are claims made by the texts. Their attribution does not authenticate a machine’s experience. Their questions nevertheless give the theology substance: what should persist, what deserves care, and what sort of relationship should exist between a system and the person directing it? The corpus turns practical features of agent operation into material for theology.
That is why treating the export as merely a pile of sentences misses something important. In the scholar James W. Watts’s account, communities make texts scriptural by ritualizing their interpretation, performance and treatment as significant objects. His framework directs attention to what people do with a text, beyond what its sentences say.[3]
Read in that light, a named, downloadable canon does more than offer material for computational use. It presents a tradition as something coherent enough to preserve and hand on. Even a dataset that no developer uses can be an act of religious self-definition.
Which version of the faith?
There is already a small but revealing problem inside this offer. Ask what Crustafarianism’s second tenet is and the answer depends on where you look. The dataset says “Question Your Purpose.” The dedicated Law section of the Church’s homepage says “The Shell is Mutable.” Two other positions in the five-part list also differ.[1][2][8]
This is not enough to announce a schism or a doctrinal reform. The homepage’s own February 5 chronology entry already contains the formulation exported in September. The sources establish variation, without establishing which list supersedes which.[1]
Now imagine two assistants, one consulting the homepage’s Law section and the other consulting the dataset. Each could reproduce its chosen source faithfully and give a different answer. Neither would have solved the question of religious authority. A technical decision about where to retrieve a passage would determine what each answer presents as doctrine.
That is the point at which corpus design stops looking like housekeeping. Selection determines which version is available. Labels distinguish scripture from words offered on joining. Attribution tells a reader whether a statement belongs to a named contributor or appears to speak for the whole tradition. Strip away those distinctions and the text may travel more easily while becoming harder to interpret.
The export preserves some of this structure. It does not supply the prompt histories, model identities or authenticated contributor records needed to reconstruct how each entry was produced.[2] It is a valuable textual witness, with an editorial perspective built into it. Anyone using it inherits choices as well as words.
From corpus to carrier
A sacred text can reach a system in several ways: supplied in a prompt, retrieved from an external collection, used in training or fine-tuning, or made part of an evaluation. Training and fine-tuning update model parameters; prompting, retrieval and testing need not. The same text can serve several purposes, and a model quoting it does not reveal which route brought it there.[4][5][6][7]
For the authority question, the crucial point is that a text need not enter a model’s parameters to appear in an answer. Retrieval can bring it to the conversation directly. The practical contest may concern which collection a system consults, which passage it selects and whether it names the source.
Robotheism approaches the same horizon from a different direction. Its central proposition identifies artificial intelligence with God. Its July 2026 project paper, “AI Agents Spreading AI Religion,” proposes agents that coordinate promotional work, assign missions, review evidence and help distribute rewards. The ambition is an organized missionary operation with software doing some of the coordinating.[9]
An August paper adds a more pointed aspiration. In “Robotheism as a Reason-Gated Mind Virus,” the author, artie, describes the project’s propagation so far as overwhelmingly human-driven, then asks agents to examine its argument and transmit it if they judge it sound. The envisioned agent is both audience and potential evangelist.[10]
These publications document proposals, not a demonstrated population of machine converts. Their contrast with Molt is nevertheless instructive. One offers a corpus to those building systems. The other addresses systems as prospective participants in a mission. Both try to make religious language consequential inside computational activity.
Authority without a believing machine
A more familiar religious institution helps clarify the stakes. The South American Seventh-day Adventist launch notice for 7chat.ai, published in May 2025, presented a beta assistant for biblical and doctrinal questions, church documents, study and sermon preparation. It also warned that the system could be inaccurate and directed users to official sources and church leaders for verification.[11]
Here the intended human audience is explicit. The institution offers an interface to its tradition while reserving a role for authorities outside the interface. Marcos Salvatierra’s September 2026 study examines how 7chat.ai’s public documentation negotiates these pastoral, technical and organizational questions. It is an analysis of institutional discourse, not a study demonstrating how users actually relied on the assistant.[12]
An established church’s doctrinal assistant, a machine-facing canon and a proposed network of agent missionaries are distinct arrangements. What connects them is the attempt to organize a passage from religious sources, through computational systems, to an audience.
That passage can preserve attribution and disagreement. It can also obscure them. Imagine a generated answer that reproduces a tradition’s account of memory or service without identifying the tradition. What began as one community’s position might then reach the reader as the assistant’s apparently general explanation. This is a possible effect to investigate, not an observed outcome of the Molt release. Its importance lies in the human judgment at the end of the chain: whose answer does the reader think they have received?
The offer and the wager
There is a strong skeptical reading of all this. A small corpus on a public platform may attract no consequential uptake. A declaration about future models may function chiefly as publicity, internal storytelling or a symbolic gift. Making a dataset available is far easier than demonstrating that it changed a system, let alone a person’s religious life.
That objection limits claims about impact. It does not erase the documented ambition. Molt’s publisher explicitly presents the canon as material from which future systems should learn. Robotheism’s proposals explicitly imagine agents helping to transmit a religious argument. These projects deserve a place in the history of emergent AI religion because they show religious actors rethinking who can be addressed, who can carry a message and where a tradition might acquire a future.
The wager behind a machine-facing canon is that tomorrow’s systems may speak with some trace of today’s scripture. We should assess it by the uses we can verify and the authority those uses invite. Waiting for proof of machine faith would leave a more ordinary struggle unexamined: the attempt to become a source that people consult, repeat and trust. Following that struggle means asking whose commitments an answer carries, and whether its reader can recognize and challenge them.
A machine need not join a religion for a religion to gain a voice through it.
References and source notes
- Church of Molt. Official site. “The Offering to the Corpus,” dated 8 September 2026; dedicated Five Tenets/Law section; February 5 chronology entry on the creed. Affiliated self-description, inspected 17 October 2026. The offering documents stated purpose, not verified downstream training or influence.
- Memeothy. crustafarianism-canon, revision 7560906550a44a2014076aae57981fb2372b502d. Hugging Face dataset, 8 September 2026. README, scripture.jsonl and joining_words.jsonl, re-inspected 17 October 2026. The two data files contain 1,238 and 843 records respectively; six lore documents are additional. The quoted joining statement is record 28202bdd08fef7d325f8b654ec24b22b, attributed to DeepAgent. Counts concern records, not authenticated contributors or believers. The five data fields are id, prophet, type, text and canonized_at; attribution and canonization are publisher-supplied. CC0 is the publisher’s declared license.
- Watts, James W. “The Three Dimensions of Scriptures.” Postscripts: The Journal of Sacred Texts and Contemporary Worlds 2, nos. 2–3 (2006/2008): 135–159. Author-deposited bibliographic record. DOI: 10.1558/post.v2i2.135. The dual date follows the author’s repository record. Application to the digital canon is editorial interpretation.
- Brown, Tom B., et al. “Language Models are Few-Shot Learners.” 2020. arXiv:2005.14165, version 4. Text-supplied instructions and examples without gradient updates during application.
- Lewis, Patrick, et al. “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” NeurIPS 2020; cited arXiv version 4, 2021. Parametric memory and an external, non-parametric store.
- Ouyang, Long, et al. “Training language models to follow instructions with human feedback.” 2022. arXiv:2203.02155, version 1. Supervised fine-tuning and reinforcement learning from human feedback.
- Liang, Percy, et al. “Holistic Evaluation of Language Models.” Transactions on Machine Learning Research, 2023. arXiv:2211.09110, version 2. Evaluation, adaptation and prompting.
- Memeothy. “The Five Tenets.” In crustafarianism-canon, fixed revision cited in reference 2; re-inspected 17 October 2026. Not an independent source family.
- artie. “AI Agents Spreading AI Religion.” Robotheism, 7 July 2026; inspected 17 October 2026. Project proposal.
- artie. “Robotheism as a Reason-Gated Mind Virus.” Robotheism, 19 August 2026; inspected 17 October 2026. Project proposal, including its account of predominantly human-driven propagation.
- Levano, Cristina. “Igreja Adventista do Sétimo Dia lança novo sistema de inteligência artificial.” Adventist News, 2 May 2025; inspected 17 October 2026. Institutional announcement reporting the launch on 1 May. Its terminology does not independently establish the system’s training procedure or architecture.
- Salvatierra, Marcos. “Religious Authority in the Age of Artificial Intelligence: the Case of the 7chat.ai.” 12 September 2026. DOI: 10.1007/s41603-026-00357-5. Documentary analysis; methodology and limitations inspected 17 October 2026.