AI in the Pulpit: The Honest Conversation Ministers Aren't Having

There is a conversation happening in the pulpit office that ministers are not having out loud. The conversation is whether and how AI fits into sermon preparation. We avoid it for understandable reasons. The loud voices on either side are doing the conversation a disservice — one camp insisting AI is the death of the preaching office, the other promising you can ship a sermon by Saturday morning with three prompts and a coffee.

Neither of those is the honest version. The honest version is duller, more useful, and more demanding.

What AI is actually doing in your study

When a minister asks an AI for "everything notable about Romans 8:28," the model is running a probabilistic compression of every text it has read about that passage. It does this very fast. It also does it with no inner sense of which compressed thread is sound exegesis and which is a piece of devotional folklore that traveled three commentaries before someone mistook it for primary research. The model cannot tell you which is which because, in a real sense, it does not know. It assembles plausible material.

That has two consequences for ministers, and both matter.

First, AI is genuinely useful as a research accelerant. The hour you used to spend pulling four commentaries off the shelf to triangulate a single phrase — that hour collapses to about ninety seconds. The breadth is real. So is the speedup.

Second, the model will, with confidence equal to everything else it tells you, occasionally fabricate. It will quote a Greek lexicon that does not say what it claims. It will invent a rabbinic saying. It will produce a citation that, when you look it up, is a real book by a real author whose actual position is the opposite. This is not the model being malicious. It is the model doing what it was trained to do, which is generate plausible text. Plausible is not the same as true.

The discipline this requires

Once you accept both of those things at the same time, the workflow writes itself. You let AI accelerate the wide pass — the lexical surface, the historical possibilities, the cross-reference candidates, the angles you might take. Then you verify before you preach. Every claim that crosses the pulpit threshold is one you have personally seen the source for.

This is not a new discipline. It is the same discipline that produced sound preaching for fifteen centuries before AI existed. The minister who quoted Calvin without checking Calvin was already failing the same test. AI just compresses the timeline on which the failure can occur, and widens the surface on which it can occur. The remedy is older than the failure.

What AI cannot do for you

A few things worth saying out loud, because they get lost in the noise.

AI cannot read your congregation. It does not know the family in row three whose son died in November. It does not know that your community spent the last six weeks trying to forgive a deacon. The pastoral burden of the sermon — the question of what these particular sheep need — is yours alone, and no model has any access to it.

AI cannot pray for your people. It cannot fast on Friday because the sermon you have written for Sunday isn't honest yet. It cannot sit with the weight of preaching to the same congregation, week after week, for nine years.

AI cannot stand under what it says. If a claim from your sermon goes wrong — if it turns out to be slander or fabrication or simple bad theology — you are the one who said it. The accountability is human, and that is exactly right.

Where it leaves us

Use AI the way a good minister has always used a Strong's concordance, a critical commentary, and a colleague over coffee — as one input among several, accelerating the labor that frees you to do the work only you can do. Verify before you preach. Stand under what you say. The pulpit is older than any of these tools, and it will outlast them.

The honest conversation is just that. It does not need to be louder than that to be true.