The Spectrum of AI Control:
From Full Ownership
to Full Dependence

64% of pastors now use AI for sermon preparation. But here is the question almost none of them have asked: who controls the model they're using? Not who built it. Not who markets it. Who controls what it says about God — and what it refuses to say?

The answer matters because Gloo's FAI-C Benchmark revealed a 30-point gap between Christian-trained AI models and generic ones on theology prompts. Same architecture. Different training data. Completely different output about grace, sin, and forgiveness. Whoever controls the model controls the theology.

Churches today sit at four different levels on the AI control spectrum — and most don't realize they've chosen one by default.

Level 1: Full Ownership — Build Your Own Model

At the highest level of control, organizations train proprietary models on their own theological data. The training corpus, alignment constraints, and output boundaries are entirely in-house.

Magisterium AI trained its model on 27,000+ Catholic magisterial documents — papal encyclicals, council decrees, catechism texts. When you ask it a question, it answers from within Catholic teaching authority. ChristianGPT was built on 30,000 curated Q&A pairs designed to reflect evangelical theology. Gamaliel uses the Nicene Creed as its doctrinal guardrail, filtering responses through early church orthodoxy.

These projects prove the thesis: the same base AI architecture produces different theology depending on who controls the training data. A Catholic-trained model and an evangelical-trained model will give you different answers about the Eucharist, authority, and salvation — not because the technology differs, but because the humans who trained it differ.

Level 2: Platform with Doctrinal Alignment

Gloo — backed by $110M in funding and partnerships with major denominations — represents the most significant effort to give churches doctrinal control over AI without requiring them to build from scratch. Their model pulls a church's statement of faith to align AI responses, making the church's constitution primary rather than the AI company's constitution.

This is a meaningful step. When a Baptist church and a Presbyterian church use the same Gloo platform, they get different answers on baptism — because the model defers to each church's doctrinal documents. That's alignment that respects theological distinctives rather than flattening them into generic spirituality.

The limitation: you still depend on Gloo's infrastructure, their base model choices, and their alignment implementation. You control the theological guardrails but not the underlying architecture. If Gloo changes its approach, your outputs change with it.

Level 3: Custom System Prompting

This is where most tech-savvy pastors land. They use ChatGPT, Claude, or Gemini with detailed system prompts: “You are a Reformed Baptist pastor. Answer from the 1689 London Baptist Confession. Cite Scripture for every claim.”

It's accessible. It's free. And it provides real, if limited, theological direction. But system prompts operate on the surface of the model. The deeper alignment — the RLHF training, the constitutional principles, the annotator preferences — runs underneath. A 2025 study found that AI annotators exercise “an excessive amount of discretion” and “frequently use their power of discretion arbitrarily.” Your system prompt sits on top of choices made by people who were never thinking about your theology.

System prompts can be overridden. They can be ignored when they conflict with deeper alignment. And they provide no audit trail — you don't know when the model followed your instructions and when it followed its own training.

Level 4: Default Usage — No Guardrails

This is where 64% of pastors are. Open ChatGPT. Type a question. Use the answer. No system prompt. No doctrinal framework. No verification process.

Gloo's benchmark measured what happens at this level: leading AI models scored an average of 61 out of 100 on Christian worldview prompts. The worst performance came “when prompts require Christian interpretation.” Models “often fail to connect scenarios to Christian values, or provide coherent theological reasoning around concepts like grace, sin or forgiveness.”

ChatGPT scored 48 out of 100 on explicitly Christian prompts. It defaults to what researchers measured as secular humanism and vague spirituality — not because it's anti-Christian, but because its training data contains far more secular content than confessional Christian content. That's not conspiracy. That's math.

At Level 4, you are trusting Silicon Valley's alignment team with your congregation's theology. You are letting contract workers who followed a style guide — not theologians, not pastors, not your denomination — decide what a “good” answer about God looks like.

The Evidence Layer: A Different Approach

There is another option that cuts across the entire spectrum: instead of asking AI what it thinks about theology, use AI to retrieve what scholars have actually written.

This is what OpenLumin does. It doesn't generate theological opinions. It surfaces evidence — commentaries from 15+ scholarly sources, cross-references, interlinear data, historical context from Ancient Near East research, and your denomination's own documents. Every citation is marked as verified or flagged for review.

OpenLumin works at every level of the spectrum. If your church built its own model, OpenLumin provides the evidence layer underneath it. If you use Gloo, OpenLumin gives you independent verification. If you use a system prompt, OpenLumin gives you sourced data to check the output against. And if your pastors are at Level 4 right now, OpenLumin is the fastest path to evidence-based Bible research without requiring a technical overhaul.

“The concentration of the power over mainstream AI applications in the hands of a few powerful companies raises significant ethical concerns.”

Where Should Your Church Be?

The honest answer: as high on the spectrum as you can sustain. Level 1 is ideal but requires significant resources. Level 2 is practical for churches with clear doctrinal frameworks. Level 3 is a reasonable individual practice with its limitations acknowledged. Level 4 is unacceptable for any church that takes theological integrity seriously.

But regardless of where you land on the spectrum, one principle is non-negotiable: know who controls the model you're using, understand what that means for your theology, and build verification into every workflow.

The church has fought for doctrinal sovereignty for 2,000 years — against emperors, against governments, against cultural movements. AI model control is the latest front. And right now, most churches are losing it by default.

About the author: AI Fluency Ministry is a project helping the church understand and use AI wisely. OpenLumin is the practical application of that research — a free Bible research companion that retrieves evidence so you can do the thinking.