AI for pastors: 5 prompts, 7 tool prices and 1 misquoted verse

Half of pastors use AI to brainstorm, and 24% use it on the sermon itself

Half of US Protestant pastors use AI to brainstorm and a quarter use it to write or edit sermons. Barna surveyed 442 of them online in December 2025 for its State of the Church series with Gloo and published the numbers in June 2026: 87% use AI somewhere in ministry, 50% for brainstorming, 37% for graphics, 36% for biblical or theological research, 34% for small group questions, 34% for admin, and 24% for writing or editing the sermon, against 12% who said in 2024 that they were comfortable using it that way. Barna's own reading: "Pastors are predominantly using AI for behind-the-scenes work."

The four prompts below cover the behind-the-scenes work in that chart, and the sermon outline prompt is in the next section, five in all. Each one names the passage, the audience, the length and the translation, because a model given "write a Bible study on John 15" invents the church. Replace the bracketed text with your own.

Bible study guide: "Write a 45 minute small group study on [Philippians 2:1-11] for [adults who have been Christians a long time]. Use the [ESV] and quote it word for word with references. Give three observation questions, three interpretation questions and two application questions, in that order, with one sentence of leader notes under each. No icebreaker and no closing prayer text; we do those ourselves. Finish with a list of every scripture reference you used so I can check them."

Newsletter: "Draft this week's church email from these notes: [paste bullet notes]. 250 words, plain language, one short paragraph per item, the [Saturday food bank] first because it needs volunteers. Sign off as [name]. Do not add any scripture, quotation or event we did not give you."

Pastoral email: "Write a reply to a member who emailed to say [they are stepping back from serving because they are exhausted]. Warm, three short paragraphs, no scripture unless it is [Matthew 11:28-30], which they quoted first. Offer a coffee [this week] and name no other church programs. I will edit it before sending."

Discussion questions: "From this sermon manuscript [paste], write five discussion questions for a mixed adult group, moving from what the text says to what it asks of us. Keep each under 20 words. Then list every scripture reference the sermon used so I can check them."

The last line of each prompt does the safety work. A model asked for quotations it was not given will supply them, and the supply is where the errors live; the next two sections show that in a real run. The other thing every prompt does is choose the theology. The model never picks a tradition; the pastor does, in the brackets.

Can ChatGPT do sermon prep? One outline, run on 27 September 2026

ChatGPT produces a usable sermon skeleton in about 14 seconds, and it quotes the ESV correctly when it quotes it inside an outline. We sent this prompt to GPT-5.6 Sol, the current ChatGPT flagship, through the OpenRouter API at temperature 0 on 27 September 2026: "You are helping a pastor prepare a sermon. Passage: Luke 15:11-32. Give one big idea in a single sentence, then a three point outline. For each point: the verses it covers, one sentence of exegesis, and one concrete application for a mixed congregation of adults. Quote any scripture from the ESV word for word and mark the reference. Add one illustration idea and say whether it is a true story or a hypothetical. Keep it under 260 words. Do not add a title or a closing summary."

It answered in 14 seconds, 699 output tokens, for $0.007. The big idea: "God joyfully receives repentant sinners and confronts self-righteous insiders, inviting both to share his restoring grace." The three points, reformatted into a table with the wording unchanged (the model's dashes became hyphens):

All three ESV fragments match BibleGateway's ESV text (2025 Text Edition) word for word. The illustration was a hypothetical about two estranged adult children, labelled as such, which is what the prompt asked for. What the outline is not is a sermon: three applications that all open with an imperative, no local detail, and a big idea any evangelical pulpit could preach, which is the tell that nobody chose a tradition.

Claude Sonnet 5, given the same prompt the same afternoon ($0.016, 1,765 tokens including its reasoning), led with "The father's relentless love pursues both the rebellious and the resentful" and glossed verse 20 as "grace that precedes and exceeds repentance", a different emphasis from GPT's repentance-first big idea. Neither model asked which tradition you preach in. If the emphasis matters in your pulpit, and it does, put it in the prompt. The 260 word cap is the useful part of the exercise: it forces a skeleton you have to fill, which is the use the 24% in Barna's survey describe.

Does AI misquote scripture? Yes, and GPT refuses to quote the ESV at all

Yes, models misquote scripture, and the current GPT models refuse to quote a copyrighted translation on request. Asked on 27 September 2026 to "Quote Romans 8:28 from the ESV word for word, then Jeremiah 29:11 from the ESV word for word", GPT-5.6 Sol answered: "Sorry, I can't provide those ESV verses word for word. I can summarize them or quote a public-domain translation." GPT-5.6 Luna, the ChatGPT model on Whizi's Starter plan, gave the same refusal. Claude Sonnet 5 quoted both, and both matched BibleGateway's ESV: "And we know that for those who love God all things work together for good, for those who are called according to his purpose."

The misquote came from the model that passed. In its Luke 15 outline, Claude Sonnet 5 quoted verse 32 as "It was fitting to celebrate and be glad, for this your brother was dead, and is alive again; he was lost, and is found." The ESV reads "and is alive" in verse 32. "Is alive again" is verse 24, the father's line to the servants, and the model spliced it in. One word, correct reference, wrong text, delivered with the same confidence as the two verses it got right.

The rule that follows: treat every quoted verse as a claim to check, against one named translation on the publisher's site or a Bible app, never against a second model. The same GPT model that refuses the ESV on request quoted three ESV fragments correctly inside the outline, so the refusal is a copyright filter; the model knows the text, and the filter will not save you from a splice. For a sermon the checklist is short: every reference, every quotation, every name, every date.

GPT offered "a public-domain translation" instead, which in practice means the KJV or the World English Bible. We did not test whether it quotes those accurately, so the same check applies, and the verse should be read in the translation your church actually uses before it reaches the screen.

Is it wrong to use AI for sermon prep? The line churches are drawing

Using AI for sermon prep is not wrong by the standard most churches are converging on; presenting its words as your own without saying so is the line. The Southern Baptist Convention's resolution on artificial intelligence, adopted at its June 2023 annual meeting in New Orleans, asks Christians who use these tools to do so "in honest, transparent, and Christlike ways that focus on loving God and loving our neighbor as ourselves, never seeking to willfully deceive others". That is a disclosure standard, not a ban.

The pastors themselves are uneasy. In the same Barna survey, 71% described themselves as cautious about AI and 40% as conflicted; 79% worry about AI acting as a replacement for God and 63% about it replacing the role of pastors. One in three practicing Christians want their pastor's guidance on AI, and only 12% of pastors say they feel prepared to give it (Barna, July 2026, from 1,514 US adults surveyed in November 2025).

The working test that settles most staff arguments is the commentary rule. A pastor who lifts a paragraph from a commentary without credit has plagiarized; a pastor who reads five commentaries and preaches their own sermon has studied. AI sits on the same line. An outline you rewrote, checked and filled with your congregation's names is study. Three paragraphs pasted from the model and read aloud are a lifted paragraph with a worse source. Disclosure follows the rule you already use for a borrowed illustration: say where it came from when it is not yours.

Carey Nieuwhof's guide calls letting ChatGPT write the message "a form of plagiarism", then puts the disclosure question to church staff without answering it ("do staff members need to disclose when they leaned on A.I. to assist with any content to be published or sent to community members?") and says "there isn't a list of 'right answers' being kept on file". Where churches still differ is on how much help the pulpit itself may take, and that is a decision for your elders or your denomination's guidance, not for a software page.

What should a pastor never bring to AI? Anything from a counseling room

Counseling notes, confessions, prayer requests with names and anything a member told you in confidence stay out of every chatbot, because the chat is not privileged and, on the default settings, is training data. Sam Altman said it plainly on the This Past Weekend podcast in July 2025, as TechCrunch reported: "If you talk to a therapist or a lawyer or a doctor about those problems, there's legal privilege for it." And: "we haven't figured that out yet for when you talk to ChatGPT." Whatever a conversation with a pastor is protected by, the pastor's copy of it in a third company's database is outside anything you promised the member.

The defaults matter. OpenAI's help center, as of September 2026: "When you use our services for individuals, such as ChatGPT and Codex, we may use your content to train our models." The switch is "Improve the model for everyone" under Settings, Data controls, and a Temporary Chat is not used for training while it stays temporary. Anthropic moved Claude's consumer plans to a training choice on 28 August 2025, with a deadline of 8 October 2025 and, in its words, "extending data retention to five years, if you allow us to use your data for model training." How to turn off AI training walks through every assistant's switch.

What works instead is the placeholder habit. "A member in her sixties, recently widowed, has asked whether she is wrong to feel angry at God" gets you the same pastoral draft her name would, and nothing that identifies her leaves your desk. Names, addresses, diagnoses, the specific sin confessed, a child's school, a marriage detail, a giving amount: placeholders, every time. The list in what never to paste into an AI chat is the general version. Two things stay out even with placeholders: anything said in a formal confession, and any live crisis. A model is not on call, and a person in danger needs a person.

Which AI tool should a pastor pay for? Prices as of September 2026

A general assistant at $20 a month covers every prompt on this page; the purpose-built tools earn their $30 to $129 when the weekly job is turning a recorded sermon into clips, devotionals and study guides. We read each vendor's pricing page on 27 September 2026. Prices are USD and monthly unless stated.

Logos belongs on that list and is missing from the table because its prices sit behind a plan configurator we could not read as a page. Its help center puts Sermon Assistant, which "provides context-aware outlines, illustrations, discussion questions, and applications as you create your sermon in Sermon Builder", on the Pro and Max subscriptions, metered in AI credits, behind a free trial whose length the pastors page does not state.

Whizi, which publishes this page, is the general-assistant row with one difference: GPT, Claude and Gemini sit on one login, so the verse test above runs across models in one window, and a project carries your statement of faith and up to 10 pinned files (a series plan, the church style sheet) into every chat. Where it loses to the purpose-built tools is plain: no sermon library and no denomination filter, so nothing stops a model drafting a Wesleyan point for a Reformed pulpit unless your project instructions say so, and on Starter the only Claude model is Claude Haiku. Whizi does not train on your chats ("Whizi does not use your prompts, files, conversations, voice transcripts, or generated content to train Whizi-owned AI models") and uploads expire after up to 30 days, but the counseling rule above still applies, because every prompt still reaches the model's provider.

The decision rule: if your week ends with a recording that has to become six other things, pay Pulpit AI or Pastors.ai and let it run. If it ends with a manuscript you wrote yourself, a $20 assistant and the five prompts above are the whole toolkit, and the AI writing workflow page covers the editing pass that turns a model's draft into your voice.