When AI Gets the Bible Wrong: Fake Verses, Confident Errors, and How to Protect Your Sermon

The most-downloaded Bible app in the world has decided not to hand you a chatbot.

YouVersion's founder, Bobby Gruenewald, explained the reasoning in a March 2026 interview: the best-performing model, working with the most heavily indexed English translations, still misquotes Scripture at least 15% of the time. Depending on which model you use, that number climbs toward 60%. So they haven't shipped a public-facing tool that answers Bible questions, because the accuracy isn't there yet.

Sit with that number. It isn't "occasionally misses a nuance." It's misquotes—wrong words, wrong wording, sometimes wrong verse entirely. And it's coming from the people whose whole job is indexing the text.

Barna found last December that 87% of pastors are using AI in some form, about a quarter of them for writing or editing sermons. Which means a lot of Scripture is passing through a system that gets Bible verses wrong at a rate no pastor would tolerate from a study Bible.

This is fixable. But only if you understand what's actually happening under the hood.

Why AI hallucination happens in the first place

Here's the thing most pastors were never told: a large language model isn't looking anything up.

When you type a reference into Logos or YouVersion, the software goes and fetches that verse from a database. There's a real record, and it either exists or it doesn't. When you type the same reference into ChatGPT, nothing gets fetched. The model predicts what text most plausibly comes next, one fragment at a time, based on patterns absorbed from an enormous pile of writing.

Prediction, not retrieval. That's the entire difference, and it explains everything downstream.

For famous verses this works remarkably well—John 3:16 has appeared in the training data ten million times, so the prediction is essentially memorization. But ask for something obscure, ask for a list of cross-references on a narrow theme, ask for the Hebrew behind a rare word, and the model doesn't hit a wall and stop. It keeps predicting. It produces something verse-shaped: a plausible book, a plausible chapter number, cadence that sounds like the translation you asked for. Confidently. Because confidence is also a pattern it learned.

Researchers at OpenAI published a paper in September 2025 arguing that this isn't a bug so much as an incentive problem: standard training and evaluation reward a model for guessing over saying "I don't know." A guess might score points. An admission of uncertainty never does. So the models bluff—not out of malice, but because bluffing is what we rewarded.

You've met this personality. It's the confident seminary classmate who has never once said "I'm not sure."

A ChatGPT fake Bible verse doesn't look fake

Fabricated citations aren't a Bible-specific quirk. The clearest documented case comes from a courtroom.

In 2023, a New York attorney used ChatGPT for legal research and filed a brief built on six court decisions that did not exist. Not garbled—invented, complete with plausible case names, judicial reasoning, and internal quotations from other fake cases. Opposing counsel couldn't find them. The judge couldn't find them. Judge P. Kevin Castel sanctioned the lawyers $5,000 in June 2023, and the case became the standard warning across the legal profession.

Read that and substitute your Thursday night. A trained professional, working in his own field of expertise, could not tell fabricated citations from real ones by looking at them. That's the whole danger. Fabrications don't arrive marked. They arrive formatted correctly, in the house style, sounding exactly like the genuine article.

Now add the part that's specific to us: your congregation can't check you. When you say "as Paul writes in Second Corinthians," four hundred people take it on trust. The verification burden sits entirely on the preacher, and there's no opposing counsel in the third row.

The three things that break most often

The failures cluster into three kinds, and they don't carry the same risk.

Verse references. Sometimes the reference is invented outright. More often—and this is the sneaky one—the reference is real but the quoted text is subtly off: a word swapped, a clause tightened, a translation blended. The citation checks out. The words in your mouth don't match the words on the page.

Quotes from people. This is the highest-risk category, because AI inherits errors already circulating in print. Every preacher has heard "Preach the gospel at all times; if necessary, use words," attributed to Francis of Assisi. He didn't say it. It appears nowhere in his writings or in any early biography of him—The Gospel Coalition ran a fact-check on it years ago—and the sentiment runs against a man who sometimes preached in five villages a day. Ask a chatbot for a Francis quote about preaching and you may well get it back, confidently sourced, because the misattribution is all over the internet it learned from.

Numbers and history. Statistics drift. A study's finding gets rounded, then re-rounded, then attributed to the wrong organization and the wrong year. Historical claims about church fathers, revivals, and word origins are especially prone to it—that "the Chinese word for crisis means danger plus opportunity" genre of illustration that feels too good to check.

The five-minute fact-check

You don't need a system. You need five minutes and four habits, run on anything AI touched before it reaches the pulpit.

1. Every verse reference goes into a real Bible. Not back into the chatbot—asking the model to check itself just generates a second prediction. Open Logos, YouVersion, Blue Letter Bible, the paper one on your desk. Confirm the reference exists and that the wording matches your translation exactly.
2. Every quote goes to a primary source. If you can't find the sentence in something the person actually wrote, you have two options: cut it, or preach the idea without the name attached. "Someone once said" is honest. A false attribution is not, even by accident.
3. Every statistic goes back to the study. Find the organization, the year, the actual number. If you can't locate it in ninety seconds, it doesn't go in the sermon.
4. Every historical claim gets two independent sources. Not two blog posts quoting each other. Two.

Five minutes. Fewer, once it's habit. And notice what it costs you relative to standing up on Sunday and attributing something to Jesus that Jesus never said.

So can you trust AI with Scripture?

Honest answer: don't hand it Scripture in the first place.

That's not a dodge—it's a design principle, and it's the one thing in this article that actually removes the risk instead of managing it. There's a structural difference between a tool that produces text for your sermon and a tool that asks you questions about your sermon. A tool that drafts will inevitably generate verse references, and every one is a fabrication risk. A tool that asks you what your text says, why you landed on that reading, and whether your second point is really in the passage or just in your head, cannot put a fake verse in your mouth. It has no mouth. The exegesis stayed with you the whole time.

We should say plainly where we sit: we build an AI sermon coach, so we have an interest here. But this is precisely why we drew the line where we did. The coach questions your handling of the text; it never supplies the text. If you're weighing a general chatbot or a sermon generator instead, both can be used responsibly—just know that responsibility means you personally verifying every reference they hand you, every week, forever.

One step this week: take last Sunday's manuscript, find every verse reference and every quotation, and check them against a real Bible and a primary source. Whether or not AI touched it. Most preachers find one thing—usually a quote they've carried for years and never traced. Fix that one, and you'll have the habit that makes the rest of this article unnecessary.