I asked ChatGPT to explain Romans 8:28. The answer sounded perfect.
Confident. Well-organized. It even quoted a scholar to back up its interpretation.
Then I checked the source. That scholar never said that. The quote didn't exist. The AI made it up — a fabricated citation, presented with the same even, authoritative tone as everything else in the answer.
That's the moment I realized something most people using AI for Bible study haven't reckoned with yet: whoever controls the model controls the theology. And right now, the model will invent a source rather than admit it doesn't know one.
This is AI hallucination, and if you lead a church, you need to understand it before it reaches your pulpit.
What actually happened
I'm not a seminary graduate. I'm a software developer and a church leader in the Philippines, Assemblies of God. I'm the guy who checks the original Greek and Hebrew before he'll accept a sermon illustration, who cross-references three commentaries before teaching anything to a room full of people.
So when ChatGPT handed me a named scholar and an attributed quote for Romans 8:28, I did what I always do — I went looking for the source. It wasn't there. Not a different book, not a misremembered phrase. The scholar never said it. The AI had generated a citation that looked exactly like a real one, because generating something that looks real is what these models are built to do.
That single moment became the reason I built OpenLumin, a free Bible research platform. More on that below. But first, it's worth understanding why this happens at all — because it's not a bug you can code around. It's how the technology works.
What an AI hallucination actually is
Large language models don't retrieve facts from a database. They predict the next most statistically likely word, based on patterns learned from enormous amounts of text.
Most of the time, that prediction lines up with reality, because reality is well-represented in the training data. But the model has no built-in mechanism to check its own output against ground truth. It doesn't "know" a scholar exists in the way a librarian knows a book is on the shelf. It generates a plausible-sounding name, a plausible-sounding claim, and a plausible-sounding citation format — because plausible is the only thing it was ever optimized to produce.
That's why a hallucinated quote doesn't sound uncertain. It sounds exactly as confident as a real one. There's no tell. No hedge. No footnote saying "I'm not sure about this part." The fabrication and the fact are delivered in the same voice.
It happens again — and testing catches it
The ChatGPT moment could have been a one-off. It wasn't.
While building OpenLumin, I ran a benchmark across seven AI models on April 4, 2026, testing each one's Bible research output against a 14-point canonical-priority checklist. The test topic: "Who are the unclean spirits of the New Testament?"
Perplexity Sonar scored 8 out of 14 — and the notes on why are the real finding. Sonar hallucinated sources: it invented an "Unnamed archaeological team" as a citation, and it mislabeled AI-synthesized content as [verified] when it hadn't actually traced back to a real, citable source. It combined scholar attributions in ways that blurred who said what.
Compare that to the model that won the benchmark, Gemini 2.5 Flash, which scored 11 out of 14 at a fraction of the cost, in part because of what the notes called an "honest citation system" — it didn't hallucinate scholars or fabricate sources.
Same task. Same prompt structure. Wildly different honesty about what it actually knew. That's the pattern worth remembering: hallucination isn't a rare glitch in one bad model. It's a variable that changes from model to model, and you only find out where a given tool sits on that spectrum by testing it.
Why this matters more in ministry than in a business memo
If an AI hallucinates a statistic in a marketing memo, someone catches it in a meeting and it gets fixed before it costs anything real.
If an AI hallucinates a scholar's name in a sermon, a Bible study, or a devotional handout, it goes out under the authority of the pulpit. Someone in your congregation writes that scholar's name in their notes. They repeat it to a friend. They build a piece of their theology on a source that was never real, delivered by a leader they trusted to have checked it.
That's not a bad business memo. That's a fabricated citation wearing the authority of the church. The stakes are different because the trust is different — and once a congregation catches one invented source, they start wondering what else wasn't real.
This isn't a reason to panic or to swear off AI. It's a reason for discernment. The tool is useful. It also needs a check.
Five ways to catch an AI hallucination before it reaches your pulpit
-
Verify the name exists. Before you repeat a scholar's name from an AI answer, search for that person independently. If you can't find them outside the AI's own answer, treat the citation as unverified.
-
Ask for the specific source. Push the AI for a page number, a publication, a date. Real citations get more specific under pressure. Fabricated ones stay vague or shift when you ask twice.
-
Cross-check against a primary source. Don't stop at the AI's summary of what a commentary says — open the commentary, or a source you trust, and confirm the claim is actually there.
-
Prefer tools that label their sources. Some tools distinguish between a claim that traces to a specific, citable work and a claim the AI is synthesizing on its own. That distinction is the whole ballgame. A tool that marks the difference is doing you a favor a generic chatbot won't.
-
Test before you trust. Don't assume every AI tool behaves the same. Run the same question through more than one, and see where they disagree — that's usually where you'll find the fabrication. This is exactly what the April benchmark did, and it's why Perplexity's hallucinated sources got caught before they ever reached a real study.
This checklist works whether you're vetting ChatGPT, Gemini, or any Bible app that leans on AI. The discipline is the same one I already used long before AI existed: don't repeat what you haven't checked.
I wrote more about how these same models handle basic questions of Christian faith — not just citations, but core doctrine — in what AI tells your congregation about God. The pattern is consistent: convincing tone, uneven accuracy.
How I built OpenLumin around this problem
The hallucinated Romans 8:28 quote is the reason OpenLumin exists. I built it as a free Bible research platform where you ask a question and get a structured course built from real scholars — Matthew Henry, John Gill, Michael Heiser, and more.
The difference is in how every claim is labeled. Verified means it traces to a specific scholar, a specific work, a specific date. Training-assisted means the AI helped synthesize the point, and it still has to cite a published source. Nothing goes out as unmarked AI opinion.
The line I keep coming back to: the AI is the librarian, the scholars are the teachers, you are the student. The AI's job is to organize and retrieve. It doesn't get to invent. You can see it at openlumin.com.
Where this leaves your ministry
AI isn't going away, and it shouldn't. It's a genuinely useful research tool. But "confident-sounding" and "true" are not the same test, and a model will fabricate a scholar before it will admit it doesn't have one.
The fix isn't fear. It's discernment — the same habit of checking sources that good ministry has always required, now applied to a new kind of source.
If you want your staff and volunteers trained to use AI with that kind of discernment, or you want your ministry's AI-assisted materials audited before they go out, that's exactly what we do. Get in touch or see how we train teams at AI for Ministry.