AI Hallucinations Are a Real Problem for Sermon Writers — Here's How to Manage Them

If you've used ChatGPT or another general-purpose AI to help draft sermon content, you've almost certainly encountered this: a plausible-sounding scripture reference that doesn't exist, a quote attributed to Augustine or Spurgeon that nobody can find, a statistic with no traceable source. The model generated it with the same confidence it uses for everything else — no hedge, no caveat, no indication that the citation was fabricated.

For general writing, this is an inconvenience. For pastoral work, it's a different problem entirely. Invented scripture, cited in a sermon, preached from a pulpit, can undermine congregational trust in ways that take months to repair. And the embarrassing part is that the pastor didn't make it up — the tool did.

Why AI Fabricates Citations

Language models don't "know" scripture the way you do. They've been trained on a large corpus of text that includes Bible passages, commentaries, and theological writing — but the model's representation of that knowledge is statistical, not structured. When asked to provide a verse "about perseverance," the model predicts what a scripture reference in that context would look like, based on what it's seen before. If the precise verse doesn't come cleanly out of that prediction, the model fills the gap with something plausible.

The same dynamic applies to quotes from historical figures. Ask an AI for a quote from John Wesley about sanctification, and it may give you something that sounds exactly like what Wesley would have said — without it being anything Wesley actually wrote. The model has enough context to construct something plausible; it has no mechanism to distinguish "I have this exact text in my training data" from "this is the kind of thing this person might have said."

The Scale of the Problem

Informal surveys of pastors using AI tools regularly find that 1 in 4 AI-generated scripture references is inaccurate or invented. The rate climbs for less-quoted passages, Apocrypha, and translations the model has seen less of. The more obscure the source, the higher the fabrication rate.

Most fabrications are close enough to real passages that they're not caught on a quick read. "The Lord is my shepherd; I shall not want" is safe. "Come to me, all who are weary, and I will give you rest" is safe. "For I know the plans I have for you, plans to prosper you and not to harm you" — that one is often slightly paraphrased in ways that differ from every actual translation. The difference is invisible in a sermon manuscript.

The Approaches That Don't Work

The usual advice is to verify everything manually — which is correct as far as it goes. The problem is that it puts the entire verification burden on the pastor, makes the AI assistance net-negative if every citation needs to be fact-checked anyway, and creates a workflow where the useful parts of AI assistance (drafting structure, suggesting transitions, tightening prose) are mixed with a landmine-detection pass that requires full concentration.

Some pastors respond by telling the AI not to include citations. This helps with the fabrication problem but removes scripture from what should be a scripture-driven document. The AI can still produce useful structural drafts, but you're now manually adding all your own scriptural support.

What Actually Works: Verified Sources + Refusal

The two-part solution is straightforward in principle but requires the tool to support it:

• Verified scripture sources. The tool fetches verse text from an authoritative Bible API rather than generating it from the model's training data. When you reference Romans 8:28, the tool inserts the actual text from the translation you specified. No generation, no fabrication, no approximation.
• Anti-fabrication rules. A rule that tells the AI to refuse to generate a citation or statistic it cannot source — rather than producing something plausible and hoping you catch it. When the AI doesn't know the exact source, it says so instead of inventing one.

The goal is not to remove AI from the drafting process. The goal is to confine generation to the parts of the document where fabrication risk is acceptable, and use verified sources where it isn't.

Presentations Mode

Sermon manuscripts sit in a category of writing that shares more with keynote presentations and academic papers than with novels: every factual claim needs to be traceable, every quote attributed correctly, and the author's credibility depends on the accuracy of their sources. The drafting problems are the same as in fiction writing (structure, flow, voice), but the citation problems are categorically different.

This is why OpusDraft's Presentation mode — designed for sermons, keynotes, and any document you'll deliver out loud — enables the anti-fabrication rule pack by default. Scripture is pulled from a verified Bible API in the translation you specify. Quotes from external sources are flagged as needing manual verification rather than generated on the fly. The AI drafts structure, flow, and prose; the facts come from sources you can point to.

The Workflow

The practical difference is immediate. You draft a sermon outline. You reference a passage — Romans 8:28, Jeremiah 29:11, Psalm 23. The tool inserts the verified text from your chosen translation. The AI builds around it. You never have to cross-check whether the model invented the verse, because the verse didn't come from the model.

When you do use the AI to suggest a quote or statistic, the anti-fabrication prompt tells it to acknowledge when it's uncertain rather than fill the gap with a confident fabrication. You can add your own verified external quotes to your project's canon and reference them explicitly — the tool cites what you gave it, not what it invented.

AI assistance for sermon writing works. The tooling just needs to be built for the specific risks of pastoral content, not for general-purpose generation.