It's a Tuesday night. A volunteer is compiling the prayer list for Sunday's bulletin — twenty-some requests, handwritten on cards and typed into a group chat. Some are ordinary: a knee surgery, a job interview. A few are not. A marriage in trouble. A teenager's suicidal thoughts. A member's cancer diagnosis, not yet public.
The volunteer is tired and behind schedule. So she copies the whole list, pastes it into a free AI chatbot, and types: "Can you summarize this and format it for the bulletin?"
Nobody told her not to. Nobody thought to. And that's the actual state of church AI data privacy in most congregations right now — not a policy failure so much as a policy absence.
This post is about what happens after she hits enter, why churches are especially exposed, and what a workable rule looks like.
What actually happens to that data
I'm not going to pretend I can quote you the exact terms of service for every AI tool on the market — those change, and you should check the current policy of whatever tool your church uses before you trust it with anything. But the general mechanics of free consumer AI tools are well-documented enough to describe honestly.
Free tiers often train on your conversations. Many consumer-facing AI products, by default, may use what you type — and what you paste in — to improve their models, unless you've gone into settings and opted out (if that option even exists for the tool you're using). That means the prayer list doesn't just get summarized and forgotten. Depending on the product and its current policy, it may be retained, reviewed, and potentially used to train future versions of the model.
Retention isn't the same as deletion. "We don't show this to other users" is a different promise than "we don't keep this." Data can sit in logs, backups, and training pipelines long after the chat window closes — again, this varies by provider and by tier, which is exactly why "I assumed it was private" isn't a safe assumption.
Human review is a real possibility. Some providers reserve the right to have staff or contractors review conversations flagged for safety or quality purposes. That's a reasonable practice for a general-purpose consumer product. It is not a practice you want applied to a teenager's disclosure of self-harm.
Enterprise and API tiers typically behave differently. As of this writing, most major providers — OpenAI, Google, Anthropic, and others — offer business, enterprise, or API-level agreements that explicitly state your inputs are not used for training and are handled under stricter retention and access terms. That's the tier built for organizations handling sensitive information. It usually costs more, requires an actual account setup, and is not what a volunteer signs up for at 9 p.m. on a free consumer app.
The gap between those two tiers is the whole problem. The free version and the enterprise version can share a brand name and a chat interface and still carry entirely different data commitments underneath.
Why churches are uniquely exposed
Hospitals operate under HIPAA. Therapists have licensing boards and confidentiality statutes. Lawyers have privilege. Churches have none of that — no federal framework requires a congregation to protect what members disclose to their pastor or to each other.
What churches have instead is trust, offered voluntarily, often at the most vulnerable moment in a person's life. A prayer request is not customer data. It's a disclosure someone made because they believed the room was safe. When that disclosure gets pasted into a tool with murky retention practices, the breach isn't regulatory — it's pastoral. It's a broken confidence with no HIPAA fine attached, which somehow makes it easier to overlook and no less serious.
Add to that the reality of who's doing the typing. It's rarely the pastor. It's the volunteer running the bulletin, the admin assistant scheduling counseling appointments, the small group leader posting requests to a shared doc that later gets pasted somewhere for "help organizing." Nobody in that chain necessarily knows what a free AI tool does with the text they hand it, and nobody assigned them to find out.
Your church's whole asset is that people believe you can be trusted with their worst moments. That trust is not rebuilt by an apology after the fact. It has to be protected on the front end, structurally, by people who decided in advance what does and doesn't go into these tools.
The three-tier rule
You don't need to understand machine learning to run this. You need one rule your whole staff and volunteer base can hold in their heads.
| Tier | What it covers | AI tools allowed |
|---|---|---|
| Green — may go into any tool | Publicly shareable content: sermon outlines (before personal study), generic bulletin copy, event descriptions, general research questions | Any AI tool, including free consumer tiers |
| Yellow — enterprise-grade tools only, with review | De-identified administrative data: attendance trends, budget summaries, scheduling, first drafts of communications that don't name individuals | Only tools under a business/enterprise agreement with a no-training data commitment — verified, not assumed |
| Red — no AI tool, ever | Prayer requests, counseling notes, member records, giving/financial data, anything naming a specific person's struggle, diagnosis, or crisis | None. This stays with a human, on paper or in a locked system your church controls |
The rule is simple on purpose: if a request, a name, and a struggle appear in the same sentence, it doesn't go into an AI tool — free or paid. Enterprise-grade data agreements reduce risk; they don't eliminate the pastoral problem of a machine processing someone's confession.
Never paste these into a free AI tool
Make this list visible — printed by the church laptop, pinned in the staff group chat, whatever gets seen:
- Prayer requests, especially anything involving crisis, grief, addiction, or mental health
- Counseling notes or anything discussed in a pastoral care conversation
- Member contact information, addresses, or family details
- Giving records, pledges, or any financial data tied to a name
- Background check results or safeguarding/child-protection records
- Staff personnel matters — discipline, performance, conflict
- Anything a member shared with the words "please don't tell anyone" attached to it
If you're not sure whether something belongs on this list, treat it like it does. The cost of over-protecting a bulletin announcement is nothing. The cost of under-protecting a prayer request is a broken trust you can't fully repair.
The data clause every church AI policy needs
Church AI policies tend to focus on doctrine — what AI may say about Scripture, whether a sermon draft needs review, which tools are theologically sound. All of that matters. But a policy that never mentions data protection has a hole in it big enough for a Tuesday-night volunteer to fall through.
Here's the clause, plain enough to adopt this week:
Data rule: No sensitive pastoral information — counseling matters, member records, private prayer requests, or giving data — is ever entered into third-party AI tools, free or paid, without a signed enterprise agreement confirming the data is not used for model training and is handled under appropriate retention controls.
Pair that rule with a short set of commitments your leadership can actually stand behind:
- Bible first, AI second, pastor last — AI assists after human judgment, never before it, and never in place of it for anything pastoral.
- We verify before we trust — any tool touching real member data is checked against its current, actual privacy policy, not the policy we assumed it had.
- AI never carries the first draft of pastoral counsel — people get their pastor, not a paraphrased chatbot.
- We disclose meaningful AI use to leadership, and to members when it touches something personal.
- Sensitive data has one home — a system the church controls, not a free tool's servers.
Write it down. One page is enough to start. A policy that exists only in your head protects nobody, because the volunteer at 9 p.m. has never had the conversation with you.
Where to go from here
This is the data half of a bigger conversation — doctrine, tool selection, and mode classification are their own questions, and I'll be writing a fuller church AI policy template in a future post. For now, the data rule above is the piece that can't wait: it's the one most likely to be broken tonight, by someone who means well and has never been told otherwise.
If you want help auditing what your staff and volunteers are actually pasting into AI tools right now — and building a policy your board can approve this month — that's exactly what I help churches do through the Navigate track. It starts with finding out what's already in the building, then closing the gap before it becomes a broken confidence instead of a hypothetical one.
Or if you just want to talk through what this looks like for your specific context, reach out. A prayer request is somebody's most vulnerable moment placed in your care. It deserves better than a free tool's fine print.