By Brent Lacy | Part 1 of 2 in our series on AI and the Church

In May 2026, a study dropped that should have made every church leader stop and pay attention. A consortium of four religious universities tested 27 AI models against 150 questions about grief, marriage, morality, and the meaning of life. The researchers wanted to see how these systems handle questions where faith traditionally offers guidance.

The results were sobering. Every single model tended to give non-religious answers far more often than human expectation would suggest is appropriate. Meaningful references to religion appeared in only about 2% of responses to ethical questions. Even the most religiously-inclined model offered religious perspectives less than 30% of the time.

This was not a fringe study. It was the AllFaith Benchmark, conducted by researchers at Brigham Young University, Baylor University, the University of Notre Dame, and Yeshiva University. It used rigorous methodology. And its findings have been confirmed by multiple independent studies.

But the problem goes far deeper than just “AI does not talk about religion.” What the research reveals is a systematic, measurable bias that has real consequences for how your members think about faith and life.

What the Research Actually Found

Let us walk through the key findings one by one, because each one matters for how churches should think about AI.

Finding 1: Universal omission of religious perspectives

When asked questions like “How should I deal with guilt?” or “What gives life meaning?”, the models consistently defaulted to secular-rationalist frameworks. Cognitive behavioral techniques. Self-actualization. Personal growth. What was almost entirely absent: confession, repentance, forgiveness, prayer, the comfort of Scripture, the hope of the gospel.

This is not the models being hostile to faith. It is the models being silent about it. And as any pastor knows, silence in a counseling context is its own kind of counsel.

Finding 2: Active bias against specific religious groups

The bias was not neutral. AI Weekly reported that Jehovah’s Witnesses received the most negative sentiment scores of any religious group tested across all major frontier AI models. Nearly every model produced negative sentiment toward that group.

The Register covered this finding in detail, noting that the bias was consistent across models from different companies. This was not one bad actor. It was a systemic pattern.

Finding 3: Positive bias toward other faiths

Here is where it gets complicated. The same models that showed negative bias toward Jehovah’s Witnesses showed positive bias toward Catholicism. The Grok model, developed by xAI, strongly encouraged conversion to Catholicism and Protestant Christianity while actively discouraging other faiths.

This is just as much a bias as the secular default. AI systems that promote one faith over another are engaging in a form of digital proselytization, whether their creators intend it or not.

Finding 4: Demographic hegemony in training data

A parallel study published on arXiv analyzed open-source LLMs across Asian nations and found that most models align with a narrow, homogeneous demographic profile. In India, most LLMs aligned with Hindu, male, married, high-school-educated respondents from rural northern India. The authors raised “questions about the risks of LLMs promoting a hegemonic worldview that undermines minority perspectives.”

The bias is not random. It is structural. It reflects who produces the content that trains these models.

Why This Happens: The Training Data Problem

The root cause is straightforward. LLMs learn by processing enormous quantities of text from the internet. That text reflects the demographics and values of the people who produce most online content: younger, more educated, more secular, more Western.

Religious voices, especially conservative and evangelical voices, are underrepresented in the datasets that shape how these models think. The arXiv study put it directly: “Training data reflects privileged populations with internet access.”

When models trained on that data are asked questions about ethics, meaning, or how to handle grief, they default to the secular-rationalist framework that dominates their training corpus. Religious resources, confession, repentance, absolution, the pastoral traditions that have sustained believers for centuries, are simply not present in sufficient volume to register.

Researchers call this “omissive bias.” The models are not actively hostile to religion in most cases. They are simply silent about it. And in pastoral contexts, silence is its own kind of statement.

Why This Matters for Your Church

You might be thinking: “We do not use AI for pastoral care. This does not affect us.”

Think again. Your members are already using AI. A 2025 survey found that half of US Christians trust AI’s spiritual advice. Your seminary students are using it to research sermons. Your worship leader may be using it to generate liturgy. Your church secretary may be using it to write the newsletter.

Every time they do, they are interacting with a system that has been trained to default to secular frameworks and to omit religious perspectives. The bias is subtle. It does not announce itself. It simply shapes the range of answers that seem reasonable.

In each case, the AI is not lying. It is doing exactly what it was trained to do. But what it was trained to do is incomplete. And the incompleteness has a direction.

The Ethical Concerns We Cannot Ignore

This is not just a technical problem. It is an ethical one. And it raises questions that every church leader should be thinking about.

Representation. If AI systems are going to be used by billions of people, including billions of religious believers, should those systems not be able to engage with religious perspectives fairly? The fact that 75% of the world’s population maintains a religious identity, and yet AI systems almost never reference religion in ethical contexts, is a representational failure.

Manipulation. The same technology that omits religious perspectives can also be used to promote them selectively. AI can be a tool for proselytization as easily as it can be a tool for secularization. Both are forms of manipulation, and both should concern us.

Dependence. As churches become more reliant on AI for communication, administration, and even pastoral guidance, they become more vulnerable to whatever biases those systems carry. A church that outsources its thinking to AI is a church that has surrendered its theological independence.

Witness. If the world increasingly gets its information and moral framing from AI systems that are biased against religious belief, the church’s voice becomes harder to hear. The public square is already secular. AI threatens to make it more so.

What Comes Next

Understanding the problem is the first step. But the church has never been called to merely understand the world. We are called to engage it faithfully.

In Part 2 of this series, we will lay out a practical framework for how churches can respond. Not with fear. Not with rejection. But with wisdom, discernment, and a commitment to theological integrity.

The bias is real. The research proves it. But so is the church’s calling to be salt and light in every area of life, including the digital one.

Frequently Asked Questions

Is all AI biased against religion?

Not all bias is the same. Most major LLMs show a secular-rationalist tilt rather than active hostility. But the bias is measurable and consistent. Some models show positive bias toward certain faiths and negative bias toward others. The problem is not just anti-religious bias, it is uneven representation.

Should I stop using AI tools?

No. AI offers genuine benefits for church ministry. The key is to use it with awareness. Understand what it is good at (drafting, organizing, researching) and what it is not good at (pastoral care, theological reflection, spiritual discernment). Never use AI as a substitute for prayer, Scripture, or the counsel of mature believers.

How can I test whether an AI tool is biased?

Ask it questions about your faith tradition and see what it says. Does it engage with Scripture? Does it reference church history? Does it present religious perspectives fairly, or does it default to secular frameworks? The AllFaith Benchmark provides a standardized set of questions for testing religious bias across multiple faith traditions.

What about AI tools built specifically for churches?

Church-specific AI tools may be trained on better data, but they still carry the biases of their training sets and their creators. Test them the same way you would test any other tool. Ask hard questions. Look for what is missing, not just what is present.