The Quote Owen Never Wrote
Tim Challies recently published an honest account of his experiments with AI research tools. He caught two fabricated citations: a composite attributed to John Owen that Owen never wrote, and an invented Ryle reference that Ryle never made. He caught them because he knew to ask. He prompted the AI to verify its own sources, and when pressed, the tool admitted it had confabulated. His conclusion is sound: prompt AI to never fabricate, and always verify against source documents. This is good counsel, and his practices represent the responsible ceiling of general AI use for theological research. But that ceiling is not high enough for the pulpit, and the reason has nothing to do with prompting technique.
A pastor who cites a fabricated Owen quote from the pulpit is not simply making a factual error. He is lending the authority of one of the most rigorous theological minds in church history to words Owen never wrote. The congregation does not fact-check sermons. They receive the quote as Owen's, carry it home, repeat it in small groups, post it to social media, and cite it in their own conversations. The misattribution propagates. It becomes part of the ambient theological culture of that church, attributed to a man who never said it.
This is not a hypothetical concern. Spurgeon, Luther, and Ryle are each burdened with dozens of quotes they never wrote, circulating widely because they sound right and carry the weight of a trusted name. AI does not create this problem from nothing; it accelerates a problem that already exists, at a scale and speed that prior generations of misattribution could not achieve.
Understanding why this happens matters. A language model trained on vast quantities of theological text absorbs thousands of genuine Owen passages alongside thousands of secondary descriptions of Owen's thought, paraphrases of Owen, summaries of Owen, and sentences that sound like Owen but were never penned by him. When asked to produce an Owen quote, the model does not retrieve from a verified document. It generates, from the full distribution of Owen-adjacent language it has encountered, something that is statistically plausible as an Owen sentence.
The result can be elegant. It can be theologically coherent. It can even be consistent with Owen's actual positions. And it will be false. Challies caught his errors after the fact, by knowing to look. That discipline is exactly right for a writer publishing to an audience of careful readers, many of whom will notice and flag problems. But the pastoral situation is structurally different, and the difference is not trivial.
Richard Baxter's charge to pastors in The Reformed Pastor to take heed to themselves before taking heed to the flock was not merely a call to personal holiness. It was a call to the kind of integrity that refuses to handle the things of God carelessly.
That standard applies to sources. A pastor who would never knowingly misquote Owen should be equally unwilling to unknowingly misquote him, which means the tool he uses for research cannot be one that generates plausible Owen and presents it as actual Owen. The stakes here are not primarily about reputation or embarrassment, though those matter. They are about the integrity of the teaching office. When a pastor stands in the pulpit, he is not offering his own opinions dressed in borrowed authority. He is handling a tradition, a body of received testimony about what the Scriptures teach, built up over centuries by men who labored carefully and at great cost. To cite Owen is to invoke that tradition. To cite a fabricated Owen is to corrupt it, however unintentionally, and to do so in a context where the corruption will go uncorrected.
The solution Challies recommends, better prompting and post-hoc verification, is genuinely useful for a writer working in a context where errors can be caught and corrected. But it is not a structural solution. It is a discipline layered on top of an architecture that remains capable of fabrication. The prompt "never fabricate" does not change what the model can do; it only changes what it is instructed to attempt. Under pressure, under ambiguous queries, or simply in the ordinary conditions of a busy pastor researching at ten o'clock on a Tuesday night, that instruction will sometimes fail. And the failure will not announce itself.
There is a better architecture for this use case, and it works differently at the level of retrieval rather than generation. Consider a concrete scenario. You are preaching on Thursday night and want to verify an Owen quote you half-remember from a commentary you read two years ago. A retrieval-based system, one that can only surface language appearing in a verified document, tells you: "I found this passage in Owen's Mortification of Sin." A generation-based system tells you: "Here is an Owen-sounding passage about mortification." One answer lets you check the source. The other leaves you guessing, with no way to know whether what you are holding is Owen or a plausible approximation of Owen.
When a retrieval-based system cannot locate what you are looking for, it says so. That is an honest answer. A system that says "I cannot locate that in Owen" is telling the truth. Failure to find is a trustworthy response. Fabrication is not. Pastoral ministry requires the former and cannot afford the latter. This is why Theostack's research assistants are built on retrieval from a curated library rather than open-ended generation: when the source is not in the library, the system says so, rather than constructing something that sounds like it might be.
The accumulated weight of a ministry built on verified sources is a different thing entirely from one built on plausible approximations. The question for every pastor is not whether AI can help with research, because it can, but whether the tool he chooses makes it easier or harder to handle the tradition with fidelity. Congregations deserve the real Owen. They deserve the real Ryle. And the pastor who loves them will settle for nothing less.
The calling to handle the word of truth with precision does not end at the text of Scripture. It extends to the tradition of those who have labored over that text before us. To cite them faithfully is to honor their work and to serve the people who trust us to tell them the truth. That is worth whatever extra care it requires.