Sermon Research Without AI Hallucinations

Most pastors who worry about AI aren’t worried about laziness. They’re worried about one specific Sunday: a statistic, a quote, or a “Greek meaning” goes into the sermon, sounds right, and turns out to be wrong. Someone in the pew looks it up afterward.

That worry is reasonable. A language model that answers from memory is built to produce plausible text, not to check itself. It can hand you a journal article that was never written, a Bible verse that says something different from what is claimed, or a Hebrew word with a meaning nobody in the field would recognize, all in the same calm tone it uses for true statements. (For the mechanics, see our earlier post on why hallucination happens.)

So the useful question isn’t “can I trust AI?” It’s “what did this answer have to pass before I saw it?” Here is a routine that works on any research, whether it came from an AI tool, a website, or a well-meaning colleague.

The five-check routine

1. Open the verse. For every Scripture reference, read the passage yourself. Check that the reference exists, that the verse says what is claimed, and that the quotation isn’t quietly borrowed from a parallel passage. This takes two minutes and catches the most embarrassing errors.

2. Confirm the source exists. For every book, article, or commentary cited, look up the author, title, and year. If there is a DOI, paste it into doi.org and see whether it resolves to the paper described. Invented citations often look perfect and lead nowhere.

3. Read the quote in the source. Never trust quoted wording you haven’t seen in the original. Paraphrases drift, and AI systems in particular tend to smooth a real author’s sentence into something more quotable than what was actually written.

4. Trace the historical and numerical claims. A date, a population figure, a statistic about church attendance: ask where it came from. A number with no traceable source shouldn’t go into a sermon, however tidy it sounds.

5. Test the original-language claims. If you’re told a Greek or Hebrew word “literally means” something, check it against a lexicon and against how the word is used in other passages. Look at the word in context, not just the dictionary entry. If you can’t read the language, ask whether the tool shows its evidence or just asserts a conclusion.

Run these five and you will catch most factual errors before they reach the pulpit. Keep the notes: a footnote you can defend on Wednesday is worth more than a clever line you can’t.

The one question to ask a research tool

When you evaluate any AI research tool, ask: what is this answer checked against?

An ungrounded tool answers from whatever it absorbed in training. A grounded one is required to work from documents it can point to, and every claim carries a footnote you can follow. Grounding doesn’t make a system perfect, and no honest vendor should claim it does. It changes where your time goes: from discovering whether something is true to confirming something you can already trace.

We built Didymus Lab around that second model. Reports are built primarily on openly licensed sources (CC-BY, CC0, and public domain), pass automated checks for nonexistent verses and unresolvable DOIs, go through a separate adversarial audit, and are reviewed by a biblical researcher before delivery. You still make the final call, as you should. What changes is that there’s a source trail to check.

You can see what that looks like in a sample report, or get your first report free and run the five checks on it yourself.

Comments

Loading comments…

Leave a comment

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

Name (required)Write a comment (max 1,000 characters)

Submit comment