How to Use AI for Sermon Writing Without Inventing Scripture
AI writes confident, plausible-sounding scripture references that don't exist. For anyone who preaches, that's not a software problem — it's a stewardship problem.
I've used AI in my sermon preparation. And the first time it fabricated a scripture reference, I almost didn't catch it.
The passage sounded right. The verse number was plausible. The sentence fit the argument I was building. I was moving fast on a Saturday night, reading for flow rather than accuracy, and the AI had produced exactly the kind of line I was looking for — confident, well-placed, theologically coherent.
It was wrong. The reference didn't exist in that form. The model had blended two similar passages and produced a hybrid that matched the pattern of scripture without being scripture.
If I stand in the pulpit and say God said something He never said, that's not a software problem. That's a stewardship problem. And from that point on, I started thinking differently about how I use these tools.
Why AI fabricates scripture
Language models are trained to produce text that matches the patterns of their training data. In theological contexts, authoritative writing cites scripture. It attributes quotes to recognized figures. It sounds like someone who knows the tradition.
The model has absorbed an enormous quantity of theological writing. It has learned the pattern of citation, attribution, and exegesis. What it has not learned is where its knowledge ends. When it can't recall the exact text of a verse — or when it has absorbed several similar verses and produces a blend — it does what it's designed to do: it produces text that fits the pattern. The pattern says "cite a verse here." So it cites one.
This isn't deception. The model has no concept of truth. It has no awareness that it's producing something inaccurate. It's doing the one thing it knows how to do: generating plausible text. And in theological contexts, plausible text sounds like scripture.
The specific risk for sermon writers
Factual errors in a business document get corrected in the next version. A fabricated citation from the pulpit is different. Your congregation is trusting that what you say God said is what God said. That trust is not easily rebuilt when it's broken, and it's especially vulnerable to errors that are quiet — not wildly wrong, just slightly off in ways that erode confidence over time.
The risk scales with how fast you're working. Under deadline pressure — a Saturday finish before Sunday morning — review gets compressed. You're reading for flow, not for accuracy. That's exactly when a plausible-but-wrong citation makes it through.
What responsible AI use looks like for sermon prep
AI genuinely helps with the parts of sermon preparation that don't require verified accuracy: structuring an argument once you've established your text, drafting transitions between points, offering illustrations you can choose from or discard, tightening prose that's running long.
What it shouldn't be doing is generating your exegesis, producing your citations, or putting words in the mouths of historical figures. That part belongs to you, to your study, and to verified sources.
The practical discipline: bring your studied text, your central claim, and your illustrations into the AI conversation. Let it work on structure and language. Then verify every citation it produces — not because you're worried the AI is lying, but because it's confident even when it shouldn't be, and confidence is not accuracy.
When the tool matters
Not all AI tools handle fabrication the same way. A general-purpose language model will produce scripture when the prose calls for it. A tool built for high-stakes writing can enforce an anti-fabrication layer — refusing to generate invented citations and flagging the gap instead, so you know to supply the real source.
If you're using AI for sermon preparation, that distinction matters more than almost any other feature. A tool that treats "don't invent scripture" as a soft suggestion will violate it when the prose momentum calls for a citation. A tool that enforces it will stop and tell you where you need to do your own work.
The goal isn't to keep AI out of sermon preparation. It's to use it where it's genuinely useful — and to make sure the parts that require accuracy stay with you.