How we use AI

A person preaches the sermon. A model helps turn it into things you can read. A person decides what goes up.

That is the whole of it, and the rest of this page is the detail — because a one-line assurance is what everybody writes and nobody believes.

What a person does

The sermon. Every week, a real person prepares it and preaches it live, and everything on this site comes out of that. There is no message here that a model wrote and nobody preached.

We are a virtual ministry — there is no building and no room. People watch.

What a model drafts

On each sermon page, the companion material: the summary, the key idea, the takeaways, the questions people actually ask and their answers, the outline, the discussion questions, and the suggestion for the week.

Beyond the site, the same system drafts the weekly email, the daily devotional, the article, and social posts.

Not the sermon. Not the Scripture — that is the NIV, quoted under Biblica's standing permission, with the notice on every page. Not the key truths or the walkthrough on a sermon page — those are taken from the preacher's own manuscript, not written for it.

What has to be true before anything is published

Five checks run automatically on every drafted asset. A failure stops it. It is not a warning somebody can wave through:

The sermon's required details are present — including who preached it. Three people teach here, and an asset attributed to the wrong one is a lie about a real person.
There is actually content.
No placeholder text survives into anything published.
No unanswered question. When the drafting model is unsure, it is required to raise the question rather than invent an answer. Any question it raises must be answered by a person before the asset can move.
Every quoted verse matches the actual text. The quotation is fetched and compared. A misquotation stops the asset.

That last one is not hypothetical. It is why the check exists: a real drafting run produced a plausible, wrong version of Jeremiah 47:3, and it read perfectly well.

Reviewers advise. They do not decide.

Drafts are also read by independent model reviewers with different jobs — one asks whether the piece represents the sermon honestly, another whether it sounds like the person who preached it. More reviewers exist for an editor to run: whether the claim is drawn from the text, whether it says one thing, how it lands, how it sounds read aloud, and where it sits against Christian commentary.

None of them can approve or block anything. Their verdicts are recorded as evidence for the person reviewing the asset. We do not think a model should be the last word on whether something about Scripture is fit to publish, so it is not given that job.

A person approves every asset

Individually. Not a batch, not a setting, not a default. The approval is recorded with who gave it and which version they were looking at.

That rule is enforced in the database rather than in the software, which sounds like a technicality and is not: it means the rule holds even for a future part of this system that forgets to ask. Nothing can approve an asset while a check is failing, and an asset already approved is pulled back if a check fails later.

What we will not do

This is the part worth reading.

We do not add commentary the preacher did not reference. If he did not bring a scholar, a historical note, or an interpretation into the sermon, it does not appear in anything derived from it — including material that is public domain and free to use. The objection is not copyright. It is that putting a thought into a teacher's mouth is a lie about what he taught, and the fact that the thought is old and free does not make it his.
We do not resolve what the sermon left open. Search engines reward confident, complete answers. Sermons are sometimes not confident and not complete. Where those conflict, the sermon wins and the gap is marked.
We do not write in the first person as someone who did not say it. The devotional is written in the preacher's voice, from his sermon, and never puts an experience, an opinion or a memory in his mouth that he did not tell.
We do not post automatically. Nothing reaches social media or an inbox without a person sending it. Email arrives as a draft.
We do not tell you what you are feeling. The teaching may diagnose broadly; it does not tell an individual reader what is true of them.
We do not imply you were in a room with us. There is no room.

Where this is honest about itself

This site is a prototype. It carries real sermon data and real cross-references, the video is not wired up yet, and the system described above is being built in the open. Where this page describes something that runs, it runs today. It does not describe anything we intend to build.

If we got something wrong

Tell us and we will fix it — that goes for a misquoted verse, a misattributed line, or anything that misrepresents what was preached.

corrections@bb4j.com

We would rather hear it from you than leave it up.

Last reviewed: 26 August 2026.