How grounding, citations, a re-check pass, and a crisis gate keep a church AI from inventing doctrine — and the four questions to ask any vendor.
"Won't it just make stuff up?"
That's the first question every church asks about AI, and it's the right one. The honest answer is: a general chatbot will, and a properly built church assistant won't — but you can't tell which is which from the marketing. So here's what's actually under the hood, in plain terms, so you can ask a vendor the right questions.
"Grounding" means the assistant is only permitted to answer from a specific, controlled set of content — for a church, that's your website, your sermons, your podcasts, your PDFs. Nothing else.
The thing to understand is that there are two ways to do this, and only one of them actually works.
The weak way is to *ask* the model — in its instructions — to only use the church's content. This is a suggestion. The model can and does ignore it, especially when the question is one your content doesn't cover. It reverts to what it learned from the internet, in your church's name.
The real way is to build the constraint into the system, not the prompt. The assistant retrieves the relevant pieces of *your* content first, and the answer is generated from those pieces. If your content doesn't contain the answer, there's nothing to generate from — so it says it doesn't know, rather than reaching for the internet. The difference between these two is the difference between a chatbot you'd trust on your site and one you wouldn't.
A grounded answer should come with its source attached. Not as a footnote you have to dig for — as part of the answer itself.
This does two things. It lets the person verify the answer ("here's the page that says it"), and it gives you, the church, a way to audit what the assistant is telling people. If an answer can't point to a source in your content, it shouldn't be sent. Citation isn't a nicety; it's the receipt that proves the assistant stayed inside the lines.
Even a grounded, cited system can drift — the model can summarize your content slightly wrong. The stronger build adds a second pass: before an answer goes to the person, a separate check compares it against your actual content. If the answer isn't supported, it doesn't go out.
This is the belt-and-suspenders step most tools skip because it costs a little more per answer. For a church, it's the step that turns "almost always right" into "doesn't send what it can't back up."
Set aside doctrine for a moment. The single most important thing a church assistant does has nothing to do with answering a question correctly. It's what happens when someone messages it in real distress.
A crisis gate is a hard-coded check that runs *before any AI does anything.* It scans for signals of acute distress — in English and in Spanish — and when it sees them, it doesn't hand the message to the AI at all. It routes the person straight to the 988 Suicide & Crisis Lifeline, the Crisis Text Line, and the church's own pastoral-care team.
The reason it has to run first, in code, is that you cannot trust a language model to recognize a crisis reliably and respond appropriately every single time. So you don't let it try. The gate catches the message before the model sees it. The assistant's job in that moment is to get a hurting person to a human, fast — not to be the human.
You don't need to become an engineer to evaluate a church AI. You need to ask four questions and listen for specific answers:
If a vendor can't answer all four cleanly, that tells you what you need to know.
A church AI that won't invent doctrine isn't a matter of a better model or a sterner prompt. It's a matter of architecture: grounding built into the system, citations on every answer, a second check before sending, and a crisis gate that runs first. Those decisions are made before the assistant ever says a word — and they're the whole difference.
AskMyChurch is built by Vision Genesis in Knoxville, Tennessee. See exactly how the trust layer works at askmy.church.
Updated 2026-06-26 · AskMyChurch by Vision Genesis · Knoxville, TN
See it answer — try a live demo →