Afraid an AI Will Make Something Up and You'll Preach It?

You’ve heard the stories, maybe even from a colleague: a pastor drops a compelling illustration into a sermon — a statistic, a quote, a historical detail — pulled straight from a chatbot. It sounds right. It isn’t. Someone in the pews looks it up afterward. That scenario, more than any theological objection, is what makes a lot of pastors close the laptop and decide AI just isn’t worth the risk.

That failure mode has a name: hallucination — when an AI states something false with the same confident tone it uses for something true. A citation to a paper that doesn’t exist. A Bible reference that isn’t in the text. It isn’t the model “lying” on purpose; it’s a structural consequence of how language models work when asked to answer from memory alone, with nothing to check itself against. Ask any large language model to answer purely from what it has “learned,” with no source material in front of it, and this kind of error becomes close to inevitable.

Why grounding in real sources changes the picture

This isn’t a problem preachers have to take on faith that someone, somewhere, is working on. It’s measured. Instead of letting a model generate freely from its own training, retrieval-augmented generation (RAG) forces it to search verified documents first and answer only from what it finds there. A 2021 study from Meta AI (Shuster et al., Retrieval Augmentation Reduces Hallucination in Conversation, EMNLP Findings) measured this directly: grounding a conversational AI in retrieved documents meaningfully reduced hallucination compared to letting it generate freely. The real question was never “can you trust AI” — it’s “what is the AI required to check its answer against.”

You’re not the only one being careful here

If this hesitation feels like it makes you the outlier among more tech-forward colleagues, the data says otherwise. Barna Group’s June 2026 survey of 442 U.S. Protestant pastors (Pastors Are Using AI More Than You Think) found that 87% already use AI in some form — but 71% still describe themselves as “cautious,” and 40% as “conflicted.” Separately, Pew Research’s 2026 age-comparison study found that only 6% of respondents 65 and older said they felt “very confident” using an AI chatbot, versus roughly 30% of those under 30. That gap isn’t a sign of falling behind — it’s what careful people do when the stakes are a Sunday morning and a congregation’s trust.

How Didymus Lab actually stops this

That’s the structural reason Didymus Lab restricts its primary-language biblical texts, commentaries, and ancient sources to openly and verifiably licensed material — CC-BY, CC0, and public domain works. The model isn’t answering from an unaccountable blend of “whatever it picked up somewhere”; it’s constrained from the source pool up to write only from material whose provenance is checked in advance. And a draft doesn’t ship as-is: it passes through a verification stage that catches citations to Bible verses that don’t exist and journal DOIs that were never issued, then through a separate adversarial audit pass — a second AI process built specifically to critically re-check the finished draft from scratch. A claim with no real backing has to survive both stages, not just one. This isn’t a promise that the model behaved honestly — it’s a structure you can go open and check for yourself.

The caution itself is warranted. What matters is redirecting it — not toward “should I use AI at all,” but toward “what is this grounded in, and what checks did it pass.” Open a sample report and you can trace that for yourself: every claim carries a footnote back to its source.

---

Next in this series: another common worry — “if AI writes the sermon, is it even my sermon anymore?”

Comments

Loading comments…

Leave a comment

Comments are reviewed before being published. Only your name and comment are stored — no personal data collected.

Submit comment