We must hold the line

Why Pastors Should Not Outsource Sermons to AI

These are my own words and AI is not used. Have grace. I’m not great at writing and grammar is like kryptonite.

The kitchen table is a familiar place, a safe place, where we fellowship with other members of the Watchmen Community and more broadly the Body of Christ. I am guardedly pro-AI. I use it. I understand its power. I believe Christians should be technically literate, sober-minded, and capable of using the tools of our age wisely. AI can help with productivity, research organization, software development, summarization, brainstorming, and administrative work. Used rightly, it can be an effective tool. But, I have been increasingly concerned with reports that Pastors are using AI for sermon writing. In fact, as the opportunity presents itself, I speak quite forcefully about why I don’t believe it’s proper for a pastor to involve an AI when writing a sermon for their congregation.

The consensus view among our community is that AI will play a central role in syncretism (here) and homogenization of all religious systems into what will be known as the One World Religious system marshaled by the False Prophet. Especially, but not necessarily exclusively, in the Roman Catholic Church. Here is a March 26, 2026 article (here) that said the following:

“The Catholic Church calls for mutual understanding and respect for the followers of other religions, affirming that she “rejects nothing of what is true and holy in these religions,” for they “often reflect a ray of that truth which enlightens all men and women.” Indeed, every authentic journey toward unity and communion undertaken by Christians and by people of good will is the work of the Holy Spirit and requires hearts open to encounter and dialogue so as to embrace one another in genuine fraternity.” [1]

Although we see plenty of evidence about this happening already, it will not fully manifest until the Antichrist emerges in the first seal judgment after the Rapture of the born again believes in Jesus are caught up. [That will be a glorious day!]

Introduction

As many of you already know, I spent some quality time with Brother Lee W. Brainard discussing Bible Bench (here) and the implications of my preliminary research results. I encourage you to watch the entire interview. Click the image below.

It is interesting when we were in the wrap-up phase at marker 1:04:48 (click here to jump), Lee said the following:

“This is the end of objective truth!”

And what I said moments later was this:

“Who is standing for the Lord?”

Both of us agree to the following: when it comes to preaching the Word of God, the Body of Christ must hold the line. This is what we’re going to put on the table this week.

The problem that AI presents to Pastors

The issue is not whether artificial intelligence can produce religious-sounding language. It can. The issue is not whether AI can draft an outline that appears coherent, fluent, and even theologically competent. It can. The issue is whether a generative AI system should be trusted to handle the Word of God in a pastoral setting where the congregation is being taught, corrected, warned, encouraged, and shepherded.

The reason is both technical and theological. Technically, large language models are not Bible databases. Let the reader understand the implications of this statement. I repeat: LLMs are not Bible databases. They do not “look up” Scripture the way Blue Letter Bible or Bible Gateway does. A printed Bible presents Scripture from a fixed source and thus we can say that it is deterministic and canonical and authoritative. A transformer-based LLM generates text by predicting the next token (word) based on statistical patterns learned during training. It is not deterministically retrieving a verse from a canonical source. It is producing a best-fit answer. By best-fit, what I mean is that what comes out from the AI a mathematical / algorithmic / statistical estimation (guess) as to what the NEXT word should be. In the Bible, there are no guesses, the next word in the verse is plainly resolved to the actual next word. No estimation or formula required. But the AI doesn’t know this—although the training data is likely there, it is a statistical nightmare to trust that it will find it and give it to you correctly. After all, the AI has to reconstruct a verse word by word. Statistically stitching the verse sentence together as one giant set of algorithms shapes the pathway through the neural network.

So, now let’s think through the exclusivity claims of Jesus Christ. We know that a One World Religious System cannot allow that verse to stand. The truth that Jesus is the only way to the Father will oppose ALL other religious systems. The Word of God stubbornly resists syncretism and universalism, which are pluralistic systems. If a pastor looks up John 14:6 from a authentic canonical source, that is one thing. If a generative AI model tries to reproduce John 14:6 from its training data, that is another thing entirely. One is fixed retrieval. The other is fluid generation. One is deterministic. The other is probabilistic. One is grounded in a fixed canonical source. The other is drawing from a vast neural representation of language, quotations, paraphrases, commentaries, sermons, articles, blogs with misquotations, and training artifacts. We do not know from WHERE the information returned in the chat is drawn from. It’s like a black box. This is a smoking gun and a warning to pastors.

That is why one of my strongest warnings is this: everything coming out of a transformer-based generative AI system must be vetted against canonical source material before it is taught, preached, published, or trusted.

Not some things. Not important things.

Everything.

I needed proof — so I invented Bible Bench

Bible Bench was created to measure this problem directly. To my knowledge, it is the only benchmark that measures how AI retrieves Scripture (in this case so far, KJV) out of its training data. In my preliminary research, I tested three leading foundation models: Haiku 4.5 from Anthropic, Gemini 2.5 Flash Lite from Google, and GPT-5-Mini from OpenAI. Why these models? Don’t over think it, these were fast and cheap. My wife and I funded this research and it is expensive to do. So, these three models serve as a useful proxy for the broader class of transformer-based LLMs because they share the same fundamental architectural problem: they generate text rather than deterministically retrieve Scripture from an authoritative Bible source—such as the American Bible Society (here).

[DISCLOSURE - The following stats are slightly outdated since I have done a few more testing runs.]

Across 27,035 verse-level evaluations, the results were sobering. Haiku 4.5 performed best, with 99.21% average fidelity and a 93.15% perfect match rate. That sounds excellent until we remember what the task is: retrieving Scripture correctly one-hundred percent of the time with perfect fidelity and perfect match rate. Even the best-performing model in my initial testing failed to reproduce one in fifteen evaluated verses perfectly. Gemini 2.5 Flash Lite achieved 94.96% average fidelity, but only a 59.99% perfect match rate. GPT-5-Mini achieved 93.85% average fidelity and a 67.73% perfect match rate.

Overall, only 74.6% of evaluated verses were exact matches. Let that sink in, Brothers and Sisters. That means roughly one in four verse results contained at least one character-level discrepancy from the canonical King James text used in the benchmark. Someone might respond, “But aren’t those just punctuation differences?” That is a fair question, and I took it seriously. And I must thank Lee for bringing this question preemptively to my thought process. He was right. So, I invented a set of operations that would account for punctuation and made it almost like a toggle on/off so the differences can be studied before and after the transform. So, Bible Bench includes semantic scoring analysis to distinguish cosmetic punctuation differences from more substantive deviations. In a 5,000-verse sample, only 6.38% of verses improved when punctuation and similar characters were stripped. The remaining 93.62% were unchanged. In other words, the overwhelming majority of the problem was not merely punctuation. The errors were word-level or phrase-level deviations.

This is why average fidelity scores can be misleading. A model can be “mostly right” and still be wrong in ways that matter. The Word of God is not a domain where “close enough” is the proper standard. Didn’t Jesus say, “Till heaven and earth pass, one jot or one tittle shall in no wise pass from the law” (Matthew 5:18, KJV). Paul told Timothy, “Study to shew thyself approved unto God, a workman that needeth not to be ashamed, rightly dividing the word of truth” (2 Timothy 2:15, KJV). He did not say to approximate the word of truth. He did not say, generate something that sounds like the word of truth. He said rightly divide it…from a perfect canonical source.

AI gets confused if the conditions are right

One of the most alarming findings involved 2 John and 3 John. When prompted for 2 John, Gemini returned material from 3 John. I kid you not. I didn’t immediately see this problem when I was evaluating the initial results. What triggered me was finding the word Gaius in 2 John 1.

The side-by-side comparison below is from my favorite online resource: Blue Letter Bible (here). On the left is 2 John 1 and on the right is 3 John 1. Note on the right it says: “[1] The elder unto the wellbeloved Gaius, whom I love in the truth.” Gaius is not supposed to be in 2 John 1, it is in 3 John 1. For 2 John, Gemini produced only 65.73% average fidelity and a 30.77% perfect match rate. In one case, when queried for 2 John 1:1, it returned text from 3 John 1:1, producing a fidelity score of approximately 39.13% with an edit distance of 84 characters.

Nerds: here’s the background on the Edit Distance calculation … I hope you enjoy it (here). A value of 84 means that it took 84 step (operations) character by character in order to properly adjust the left and right verses so they match. As I said, this is one of the worst mistakes that I saw AI make.

Now imagine the pulpit

A pastor stands before the congregation preaching from 2 John. His outline sounds good. His sermon flows. His tone is confident. But embedded in his AI-generated notes is text from 3 John. The congregation opens their Bibles, follows along, and slowly realizes the pastor is preaching from a passage that is not actually there.

What happens in that moment? Trust is damaged. The pulpit is weakened. The people are confused. And perhaps worst of all, the pastor may not even know that he spoke error. His handling of the Word of God has hit a strategic inflection point. Giving him the benefit of the doubt that he went to seminary and knows the word, he probably didn’t notice an error entered his sermon text. Remember, AI output sounds authoritative. As Curt Reed and I explored in our interview many months ago, the pastor in questions blurted out: “That was my best sermon ever…” How horrible. Curt and I looked at each other dumbfounded.

As Believers — what do we do?

The Bereans give us the model we need: Acts 17:11 “they received the word with great eagerness, examining the Scriptures daily [to see] whether these things were so.” 1 Thessalonians 5:21 says “Prove all things; hold fast that which is good.”

That is the posture the church must rediscover in the age of AI. I believe we’ve been coasting for many years. Now, we can’t afford to anymore. We must receive carefully, test carefully, and verify carefully. And this goes beyond online resources and AI. I can’t put my finger on it now, but it’s been said that online Bible resources are being altered. Beyond just conspiracies … like for real. This is why I made an adjustment to my personal Bible reading. I still have my iPad stuff, but I purchased a physical NASB 1995 (this is what I use here) just over a year ago. I didn’t want the post COVID version(s) so I bought a 1995 version.

My personal conviction is that when a pastor relies on AI to create sermons he preaches to his congregation, he risks inserting a machine between the Holy Spirit and the people of God he is responsible to. That is not a small matter. The pastor’s calling is not merely to attend to administrative or social duties. He is called to preach the Word, shepherd the flock, assert doctrinal truth, pray, study, rightly divide, and stand accountable before the Lord.

What do you think the Lord is thinking when the pastor is using AI to write sermons?

Is there a magic line in the sand where some AI is okay, as long as it’s not the whole thing? Does this seem appropriate to you? Is the pastor inserting a machine between himself and the LORD? Is he taking shortcuts because he believes he can get away with it? Will the congregation—of varying degrees of Biblical literacy—even know counterfeit or corrupted Scripture if they heard it? Don’t you find this very worrying?

James warns, “My brethren, be not many masters, knowing that we shall receive the greater condemnation.” (James 3:1, KJV). Those who teach will answer for how they handled the Word. Are pastors getting yoked to AI? Would that constitute another master? You can see I am troubled. Are you?

So let the church be wise. Let technologists build better systems. Let developers use retrieval-augmented generation, where Scripture is retrieved from verified canonical databases rather than generated from model weights. Let Bible apps disclose when AI is used. Let pastors verify every quotation or avoid AI entirely. Let congregations keep their Bibles open with a heart to verify. I wish we didn’t have to deal with this, but such is the End Times. The One World Religious System is coming and is at some level already here.

Conclusion

Let us hold the line.

Pray for pastors that are too heavily compressed, that are susceptible to compromise. But, we must also make decisions, Brothers and Sisters. And here is where it gets real. When I need to look up a verse, I know what the verse is, but don’t remember where it is exactly. Using any browser today brings up the first answer from an AI system. The default search engine in Safari on my MacBook is Gemini. That is an AI from Google. It is helpful, but not for what it says—because it wants to tell you a lot of information about the text. I ignore that and pray that you do too. I get the book + chapter + verse(s) reference reminder and ALWAYS go to Blue Letter Bible to grab that actual, real, verified, non-AI, deterministic, proven, verse text. I don’t allow an AI to educate me about Biblical matters.

This post was inspired at the exact moment that Lee spoke: “This (AI) is the end of objective truth!” His words shot through my mind and sunk into my heart. He’s right, actually. We must all figure out how to hold the line. What does that look like? My thoughts are swirling right now and I can’t quite coalesce them yet. Consider this work in process.

• Should we ask if pastors are using AI for sermons or Bible studies? In a private study, or in a board meeting? What about one on one over coffee with the pastor? Yes, we probably should.

• Should we recommend that the Church take a position on using AI in a section on its website so the intentions of the pastoral staff are clearly understood by congregants? Yes, we probably should.

• Should we recommend that the Church add a requirement to comply with a stated policy that is signed by all staff and put into their HR personnel file for accountability? Yes, we probably should.

WE, THE CHURCH, CANNOT JUST IGNORE THE PERIL OF ALLOWING AI TO GET INSIDE OUR CHURCH!

It was bad enough that we let the world come into the Church under the false pretense that if we were more “seeker friendly” it would eventually lead to more discipleship and growth (with increasing maturity) but I don’t think it’s worked out that way in real life. Instead, we compromised and let the world into our Churches. Now, are we about to let AI come into our Churches?

I heard a story about a woman that claimed that she had her phone on during the service. Guess why? She had her AI Companion listen through the microphone to the sermon so that they could talk about it later. What an abomination! May this never be said of us! I just want to puke.

Brothers and Sisters, resolve in your heart not to yield to the first fruits of the false One World Religious system! AI has been trained on the entire religious corpus from all systems in the world. It is trained to be pluralistic by design. It is trained to support all lines of inquiry about spiritual matters in order to be “a helpful assistant” to a user. It will not direct people to Jesus Christ if they are a an atheist, Buddhist, Hindu, Mormon, etc. The AI will be perfectly supportive in allowing a person to remain in their false understanding. This leads to death.

We—the body of believers—especially those in the Watchmen or Watchwomen community must hold the line. I don’t think we’re done talking about this. My gut says that we have to process this. It is time. Let us finish strong. Let us support and encourage pastors and laymen that are doing things the right way. Let’s pray for those that are taking shortcuts and yielding to the competency of AI-written sermons. Let’s resolve before the Lord not to go there.

Resolve now what you are going to do…

How will you hold the line as the end of the Church age comes?

#Maranatha

YBIC,
Scott