Trust but Verify: The Only Defensible AI Workflow

Two failure modes for the minister using AI in sermon prep. The first is to refuse the tool entirely, on principle, and lose the hours it could return. The second is to accept everything it produces, because it sounded confident, and preach material that the model invented.

Both failures share a shape. They are decisions to skip a step. The first skips the trust step — refusing to engage with material that, used carefully, could have improved the sermon. The second skips the verification step — accepting material that, used carefully, would have been fine but used carelessly is dangerous. The defensible workflow is the one that does both halves.

What trust looks like

Trust here does not mean credulity. It means letting the tool do the work it is good at without flinching. AI is genuinely good at surfacing the lexical surface of a passage — every word with its lemma, its rough range, its appearances elsewhere. It is good at proposing candidate angles. It is good at producing a list of cross-references that might illuminate the passage. It is fast in a way that no library is fast, and the speed is real value.

The minister who refuses to use any of that on principle is paying a real cost — hours that should have gone to prayer, congregation, and judgment, spent instead on retrieval that a tool could have done. Refusal is not piety. It is a choice that has consequences for the second half of the sermon labor.

What verification looks like

Verification is the discipline that makes trust safe. Every claim the tool produced that crosses the pulpit threshold is a claim you have personally seen the source for. Greek glosses checked against a real lexicon. Historical claims traced to the cited source and that source's evidence. Cross-references opened and read in context, with their surrounding verses, in the actual translation.

This is not new work. It is the work ministers have always done with reference materials. AI changes the volume of candidates that need verifying, not the discipline that does the verifying.

The candidates that matter most for verification are the ones most likely to be wrong. Lexical claims about Greek or Hebrew, because models produce confident etymological detail that lexicons would not defend. Historical-cultural claims, because models pattern-match folklore alongside scholarship. Quotations from real people — Augustine, Calvin, Luther, Lewis — because models will produce paraphrases that sound right and are not in the source.

Where the workflow lives

The workflow lives in two places. The first is the study, where a minister is prepared to throw away material that does not survive verification. The second is the pulpit, where the minister stands under what they say. The two places are connected by the work that happens between them.

The minister who treats the study as a place where any plausible material can survive into the sermon has skipped half the workflow. The minister who treats the pulpit as a place where unverified claims are tolerable has skipped the other half. The defensible workflow is the one that treats both halves as real.

Trust the tool to do retrieval. Verify everything that crosses the threshold. Stand under what you say. The discipline is not new and not novel. It is the same discipline that produced sound preaching for centuries, applied to a tool that would otherwise compress the timeline on which preaching can fail.