Church AI Policies Now Converge on Four Rules

Four AI Rules for Churches, and What Each One Asks of a Small Staff

On August 31, Archbishop Josafá Menezes da Silva of the Archdiocese of Aracaju, in Brazil's Sergipe state, issued guidelines for artificial intelligence and church communications across his archdiocese. Every AI-generated image gets "mandatory human review before publication." Anything "generated or manipulated by artificial intelligence" has to be "labeled and clearly distinguished from content created by humans." And no one may use AI to reconstruct or simulate the appearance of a bishop, priest, or deacon when the result replaces an actual photograph, or to simulate their voice in a message they never delivered. The guidelines then list the auxiliary work AI may do: backgrounds, textures, illustrations, layouts, color work, and editing support. They cite Pope Leo XIV's message for the 60th World Communications Day, the encyclical Magnifica Humanitas, and Antiqua et Nova, the joint Vatican note on artificial intelligence.

The published record on artificial intelligence and the Catholic Church in the United States is thinner. Three days earlier, Michelle La Rosa reviewed for The Pillar the American dioceses that have put AI policies online, and it is a short list: Biloxi, Paterson, Los Angeles, Orange, and Denver. In the same stretch of days, an ecumenical team released a free Church AI Toolkit with governance templates and pastoral triage guides, built during a five-day July sprint at Virginia Theological Seminary by Presbyterian, Episcopal, Lutheran, and Methodist participants through the TryTank Research Institute.

None of these groups coordinated with the others. A Brazilian archbishop, a handful of American chanceries, and a Protestant design sprint were writing at the same time, in different languages, under different authority, for polities that do not answer to each other. They landed on four controls that overlap almost exactly: a person reviews before anything is published or decided, AI-made content gets labeled, nobody simulates a clergy member's face, voice, or authority, and confidential data and pastoral judgment stay out of the model.

Agreement that arrives independently is stronger evidence than any single mandate would be. Four separate bodies wrote rules against the same four failures because those failures had already happened to somebody.

Whether churches should use AI is not what any of these documents argue about. Every one assumes staff are already using it, and most of them are right, a gap we covered in Churches Using AI Race Ahead of Their Own Rules. What shifted in the last week of August is that "our denomination has not ruled on this" stopped being a usable answer. Written guidance on how churches can use AI now exists in public, agrees with itself across traditions, and is short enough to adopt before the next staff meeting.

Rule one: a person decides before anything ships

Aracaju requires human review of AI art before publication. Biloxi goes further and bars AI from producing first drafts at all, permitting it for proofreading, summarization, and preliminary research, and only with supervisor authorization. Paterson requires human review for accuracy and bars final decisions made without human oversight. Los Angeles allows AI to assist work and ministry on the condition that the user evaluates every output.

Translated into operating terms, that is one requirement. Any pipeline producing something public, a bulletin, a social post, an email blast, a graphic, needs a named person who approves it before it goes out, and the approval has to be recorded rather than assumed. A verbal "looks fine" from whoever was in the office does not survive a question three months later about who signed off on the image.

Worth being precise about the tooling, because this is where church systems and processes break down: an approval gate is not something an AI tool has by default, it exists when somebody designs it into the sequence. On AgentPMT that is a step you add in the workflow builder. Draft the bulletin, stop, wait for a named reviewer, publish. The agent does the drafting; the workflow decides where it pauses.

Rule two: label what the machine made

Labeling is where Aracaju is strictest and where most of the American policies go quiet. The Brazilian guidelines require AI-made and AI-manipulated content to be visibly distinguished from human work, and Biloxi requires clear labeling too. La Rosa's review found a different pattern across the US policies as a group: they consistently require human oversight and ban feeding confidential data into a model, and few of them require labeling at all.

That gap is a tooling problem wearing a policy costume. Nobody can reconstruct after the fact which paragraph of a newsletter came out of a model, or which stock-looking background was generated on a Tuesday eight months ago. Provenance has to be captured at the moment of generation or it does not exist, and a staff member who wants to comply with no record to work from will guess or skip it.

It gets solved the same way agent accountability gets solved anywhere else. AgentPMT's action feed logs every tool call with the full request and response payload, so the generated asset and the prompt that produced it land in the same timestamped record, attached to the agent that ran and the budget that paid for it. Labeling becomes a lookup instead of a memory exercise. We went through why that record is so often missing in The Agent Observability Crisis in the Fortune 500.

Rule three: no synthetic clergy

Aracaju's prohibition is the most specific language any of these bodies published. No synthetic facial reconstructions standing in for a photograph, no significant modification of features, no placing clergy in situations that did not occur, no audio simulation of speeches they never gave. Paterson restricts impersonation and the editing of images of minors.

The reason for that specificity is sitting on Instagram. Religion News Service, reporting in America Magazine on August 31, documented AI religious personas at scale: Yang Mun, a synthetic Buddhist monk with 2.5 million followers, sells books and healing programs alongside guidance on anxiety and meditation, with no disclosure that he is generated. At that scale a synthetic teacher is a revenue stream borrowing authority that belongs to somebody real. An account presenting as an Orthodox rabbi was removed by Meta after promoting antisemitic stereotypes. In 2024, Catholic Answers dropped the "Father" title from its Father Justin character after it offered something resembling sacramental absolution.

Swami Sarvapriyananda of the Vedanta Society of New York put the objection plainly in that reporting: "The teacher embodies the teaching." Dr. Brian Patrick Green of Santa Clara University reached the same boundary from the Christian side at the AI & Church Summit in Nashville: "a mirror may reflect the image of God without possessing it."

For a communications director the requirement is concrete. Image and voice tools used by staff need hard exclusions for named people, and no AI persona speaks as a minister, not in a website chatbot, not in a voice greeting, not in a video.

Rule four: donor records and pastoral judgment stay out of the prompt

Biloxi bans entering financial data, donor information, or anything involving minors. Denver, which explicitly permits ChatGPT and Microsoft Copilot for brainstorming, summarizing, and drafting emails, prohibits relying on AI for moral, pastoral, or theological decisions. Orange, which runs its policy through a Diocesan AI Council, bars AI from replacing human interaction where empathy or spiritual guidance is required. Paterson requires professional involvement for legal, medical, or counseling matters.

Two boundaries live inside that rule. One is about data: the church systems and processes holding giving history, counseling notes, and youth rosters have to be scoped so none of it reaches a prompt. The other is about judgment: a workflow that detects a pastoral concern hands it to a person instead of answering it.

The data half is an access-control question with a standard answer. Credentials belong in an encrypted vault the agent never sees, and the agent gets only the tool actions its job requires. On AgentPMT, budgets carry per-tool restrictions, so an agent that reads a mailbox is not quietly able to write to the donor database because both hang off the same account.

A two-pastor office in Iowa, and three tools

The most useful thing published in those same days was not a policy. It was a description of somebody building inside one.

Rev. Michael Gewecke, co-pastor of First Presbyterian Church of Spirit Lake in rural northwest Iowa, described three tools at the Nashville summit, reported by Beth Waltemath for Presbyterian News Service. A self-hosted photo and video platform for a youth mission trip to Florida, so students uploaded to church-controlled space instead of commercial social apps. A taco-order app built in about 45 minutes, which collected each student's menu selections against a $12 per-youth cap and produced a printable order form for the restaurant. And an email monitor that read incoming messages while he was traveling with the youth group and flagged anything needing immediate pastoral follow-up. It alerted him. It did not write replies.

Map those against the four rules and they line up without anybody trying. The email tool is rules one and four in one design decision: it reports, a person decides, and pastoral judgment never leaves the pastor. The self-hosted platform is a scoping choice about where minors' photos live. The spending cap was set before the tool ever ran, which is the same move as capping an agent's budget instead of reading the invoice afterward and wincing. We made the general case for naming the spend up front in Cost Attribution for Agent Work: Every Dollar Gets a Name.

"The AI didn't choose to make this," Gewecke said. "It needed to be given real constraints and real boundaries." That principle sits underneath all four rules, and underneath any tool a non-deterministic system is allowed to call, which we argued at length in Designing Deterministic Tools for Non-Deterministic Agents. His framing of the limit was equally clear: AI "can help us translate the local knowledge of our communities into temporary, modest, accountable tools. It cannot replace the responsibility to decide what matters and what's permitted and what should happen next."

That email tool deserves a second look, because its shape generalizes to almost any small office. Read a mailbox, classify what arrives, alert a human, never send. On AgentPMT that is the Gmail - All Email Actions connector for reading, a classification step, and Telegram Instant Messenger or Google Chat for the alert, with the reply action simply never granted to the agent. The Gmail Inbox Triage to Telegram workflow already runs that read, sort, and alert sequence. The Google Workspace Automation agent packages the same idea for an office already living in Gmail, Drive, Docs, Sheets, and Calendar. AgentPMT had nothing to do with Spirit Lake; Gewecke built his own tools. The pattern he landed on is what carries, and it is buildable in an afternoon.

The research points at the same controls

Two studies from the same days describe, from well outside any church, why these particular rules keep showing up.

Dipto Das, Arpita Kundu, Nusrat Jahan Mim, Shion Guha, and Syed Ishtiaque Ahmed, at the University of Toronto and Khulna University of Engineering and Technology, posted a paper to arXiv on August 28 drawn from semi-structured interviews with Bangladeshi Hindu users of generative AI. Small sample, deliberately deep. Participants described real benefits, including scriptural inquiry without formal training and, for some, access to texts that caste restrictions had kept out of reach. They also described failures with uncomfortable precision. A canonically dark-complexioned deity rendered with a white complexion. AI video of Krishna saying things that are not in the Gita, which participants named as blasphemy. Sanitized explanations of caste followed, on a leading prompt, by cheerful agreement with casteist framing.

Their prescription, which they call interpretive alignment, reads like the four rules arriving by a different road: systems should disclose their limits, preserve plurality instead of collapsing contested meanings, refuse to speak in a divine first person, flag when a scholar or practitioner should be consulted, and resist sycophantic personalization. Refusing first-person divine speech is rule three. Flagging when a human expert should be consulted is rule four. Nobody on that team was reading a diocesan handbook.

The second study is about capability rather than harm. Muhammad Ruslan and Abd Rahman at Universitas Muhammadiyah Sumatera Utara published a survey of 74 Islamic religious education graduate students on August 30, measuring four dimensions of readiness. Attitudes toward AI scored highest of the four; pedagogical competence, the ability to actually design AI-supported instruction, scored lowest. They recommend training on practical design work rather than on enthusiasm, which is the same gap the diocesan policies are trying to cover. Rules help, but training is what produces someone who can build the tool correctly.

What the rules do not settle

Labeling remains the loose thread, and the Toolkit's authors were candid about their own work being provisional. Rev. Dr. Lorenzo Lebrija invited users to test it and "tell us what works, what doesn't work."

The larger arguments are unresolved, and the summit did not pretend otherwise. Participants debated aligning with advocacy organizations, a "Pro Human Sunday," sermon series, technology fasts, and calls for slower frontier development, and reached no consensus. Kristen Opalinski of the ELCA called the moment "our nuclear moment" in the context of autonomous weapons, while Kevin Hyrams, a ruling elder in Charlotte, argued for attention to present harms in surveillance and immigration enforcement.

Those debates will not be settled this quarter, and they sit on a different layer from the operating rules. A parish can adopt human review, labeling, an impersonation ban, and data scoping this month without taking a position on autonomous weapons, labor policy, or advocacy. Conflating the two is how an organization spends a year deliberating while its staff keep pasting donor spreadsheets into a chatbot.

What to do with this on Monday

The convergence is not a denominational mandate and should not be described as one. It is four independent bodies publishing compatible guidance inside a week, which makes it the closest thing to a shared baseline that currently exists.

The work it implies is unglamorous and finite. List every place AI already touches your staff's output: the bulletin, the newsletter, the social graphics, the sermon research, the scheduling, the inbox, the ordinary church task management that fills a staff week. Check each against the four rules and note which control is missing. Most will be missing the same two, a recorded approval and a provenance record, and both are configuration rather than theology.

Then build something. The lesson from Spirit Lake is that the rules and the building are the same activity, and a well-bounded tool is easier to justify to a session or a chancery than an unbounded one. As churches adopt AI for ordinary administrative load, the congregations that write their constraints down first are the ones that get to keep building.

Sources

• Brazil archbishop establishes rules for use of AI in archdiocesan communications, Catholic World Report
• How are dioceses handling AI policies?, The Pillar
• Pastoral care in the age of AI, Presbyterian News Service
• Generative AI Alignment with Hinduism's Theological Plurality and Sacred Representation, arXiv
• Readiness Profile of Islamic Religious Education Teachers in Integrating Artificial Intelligence, Journal of General Education and Humanities
• They look like religious teachers. But they're AI, and millions are listening, America Magazine and Religion News Service
• Ancestral vs. Artificial Intelligence, Presbyterian News Service
• Can AI bear the Imago Dei?, Presbyterian News Service
• Ministry on the edge of imagination, Presbyterian News Service