Skip to content
← Insights
Nov 14, 20264 min read

A church AI policy starts with the task and the person responsible

New Church of England guidance brings AI decisions down to local practice. Define permitted inputs, check authority claims and name the reviewer for each task.

An open book and a small chapel-shaped notice board beside a sealed envelope on a cream table

A church newsletter and a private prayer request can pass through the same office computer. They carry very different expectations about who will read them. An AI tool approved for drafting an event notice needs separate approval before receiving someone's account of bereavement.

The distinction gives a congregation a useful starting point for its AI policy: define the task, identify the information involved, then name the person responsible for the result. Broad arguments about whether machines can preach leave those daily decisions unresolved.

Current guidance leaves room for local decisions

On 2 November 2026, the Church of England published a guide for churches considering generative AI. Its introduction explicitly describes the resource as guidance and suggests that churches may develop their own internal policies. It also acknowledges ethical and theological concerns about the tools.

That publication offers a current point of reference for Anglican congregations. A local decision still needs to reflect the task being proposed. Permission to improve the wording of a public notice can be specific enough for someone to apply consistently.

A useful trial begins with material already intended for publication. Ask the responsible editor to compare the draft with the original event details. A wrong venue address is a small-looking error with a direct effect on the people trying to attend.

Check the actual output before approving it. A fluent rewrite can accidentally change an invitation into a promise about accessibility or facilities. The person who knows the event should be able to correct it before it reaches the congregation.

Religious authority is part of the message

An answer about doctrine carries the implied authority of the person or institution offering it. Putting a church's name above a chatbot can give its responses a status that the underlying generation process has not earned.

The Vatican's January 2025 note Antiqua et nova places human dignity and responsibility at the centre of its account of AI. It warns about treating AI as a person and about substituting artificial interactions for human relationships. This is a Catholic theological and ethical statement, with its own tradition and intended audience.

A local application is to make the origin of an answer visible. If a study assistant retrieves passages, show the passages and their sources to the person preparing the material. An approved source collection can narrow what the assistant searches. A reviewer still needs to establish whether its interpretation follows from the material cited.

Keep the final responsibility explicit. Someone preparing a sermon needs to know which claims they can defend and which passages they have actually checked. A source link that resolves to a real page can still point to material that does not support the sentence beside it.

A robot-sermon study has a defined reach

Research published in 2023 examined responses to robot preachers and AI-attributed sermons. The publisher's account of the studies describes temple settings in Japan and Singapore, plus an online experiment involving Christian participants in the United States. Credibility was lower for the automated source in the comparisons reported.

Those results deserve attention when a congregation proposes delegating a public religious role. They concern the settings and outcomes examined in that research. They cannot, by themselves, establish a future rate of religious decline or explain why a particular church building closed.

The role also differs from an internal drafting tool. A congregation can evaluate an editor's checked newsletter separately from an autonomous system presented as a spiritual adviser. Combining the two uses under one acceptance statistic would obscure what participants were actually being asked to trust.

Before a trial, describe the proposed use in ordinary language to the people affected. Their response to that specific description will be more informative locally than a general survey question about liking AI.

Confidentiality needs a data decision

A private conversation can contain information about someone who never chose to use an AI service. Removing the speaker's name may leave enough detail to identify a family within a small community.

Establish which material may enter the tool before anyone uploads it. Check the service's retention terms and who can access its workspace. A setting controlling model training answers one question; storage and organisational access need their own checks.

For a trial involving public material, keep private correspondence outside the workflow. That boundary lets the team examine whether the tool improves the task without expanding the kinds of information it receives.

The policy should also explain what happens when someone includes private material by mistake. Name a contact who can assess the incident through the congregation's existing procedures. A volunteer should be able to report the error without first guessing whether it is serious enough to mention.

Write permission at the level of the task

A short policy entry can identify the approved use, the permitted input and the person who signs off the result. Keep the entry beside the tool's instructions, where someone doing the work can find it.

Try applying it to tomorrow's newsletter before extending it to another role. If a new volunteer can tell which details to check and whose approval is needed, the policy has become usable in the place where the decision actually occurs.

AIEthicsPrivacy