48/100: Why AI Fails the Faith Dimension

48 out of 100.

The lowest of seven dimensions. Every major AI model. The Gloo FAI-C benchmark tested 20 frontier LLMs across seven areas of human flourishing (Gloo, 2025). Faith came last.

Pat Gelsinger, Gloo's Executive Chairman and former Intel CEO:

"Today's frontier LLMs aren't built with any specific value system in mind, and it shows." (Gloo, 2025)

This isn't about AI being bad at religion. It's about what happens when you average every religion together.

The Seven Dimensions

Finances at 81%. Faith at 48%. A 33-point gap.

Why? Budgeting has right answers. Save more than you spend. Pay off high-interest debt first. These principles transcend worldview.

Faith doesn't work that way.

Why This Happens

Three mechanisms converge to produce that 48.

The Linguistic Average

AI models learn by ingesting billions of documents. Catholic catechisms. Reformed confessions. Prosperity gospel sermons. New Age spirituality. Secular critiques. All weighted roughly equally.

The output: the statistical center of all traditions.

That center belongs to no actual tradition.

The Interpretation Problem

Medical AI achieves 96% accuracy on licensing exams (Chen et al., 2024). How?

By excluding interpretation.

Clinical guidelines are designed for consensus. Run the trial, measure outcomes, update the standard. Disagreement resolves empirically.

Theology cannot exclude interpretation without ceasing to be theology.

Consider John 6:53-54—"Unless you eat the flesh of the Son of Man and drink his blood, you have no life in you":

• Catholic: Transubstantiation. The bread becomes Christ's body. Literal.
• Lutheran: Sacramental union. Christ present "in, with, and under" the elements.
• Reformed: Spiritual presence. Elements remain bread; Christ present to faith.
• Zwinglian: Pure memorial. Symbolic only.

Four readings. One verse. Each internally coherent. Each with centuries of sophisticated defense.

Which should the AI output?

The Safety Problem

RLHF—Reinforcement Learning from Human Feedback—trains models to be "helpful, harmless, and honest." In practice, this means:

• Exclusivity claims feel unkind → generic gets rewarded
• Strong theological positions seem divisive → pluralism by default
• "All paths are valid" offends no one → becomes the default

The training signal: avoid controversy. The result: a theology that affirms everything and requires nothing.

What AI Actually Teaches

Here's where the analysis gets uncomfortable. The 48/100 isn't a failure to produce religious content. AI outputs plenty of religious-sounding material. The problem is what it produces.

In 2005, sociologists Christian Smith and Melinda Lundquist Denton interviewed thousands of American teenagers about their faith for the National Study of Youth and Religion. What they discovered wasn't Christianity, Judaism, or Islam. It was something else entirely—a folk religion they named Moralistic Therapeutic Deism (MTD) (Smith & Denton, 2005).

Its five core beliefs:

1. A god exists who created and orders the world
2. God wants people to be good, nice, and fair
3. The central goal of life is to be happy and feel good about oneself
4. God does not need to be particularly involved except when needed to solve a problem
5. Good people go to heaven when they die

Read those carefully. Notice what's missing: repentance, sacrifice, suffering, covenant, obedience, transformation through trial.

Smith and Denton's summary: "This is not a religion of repentance from sin, of keeping the Sabbath, of living as a servant of a sovereign divine, of building character through suffering. Rather, what appears to be the actual dominant religion among U.S. teenagers is centrally about feeling good, happy, secure, at peace" (Smith & Denton, 2005, p. 163).

The God of MTD? A "Divine Butler" or "Cosmic Therapist"—always on call, helps people feel better about themselves, never makes uncomfortable demands, doesn't get too personally involved in daily life.

This is exactly what AI outputs look like.

The match is exact. When AI averages all religions, it doesn't produce neutrality. It produces a specific theology—the most popular folk religion in America. MTD isn't an absence of faith content. It's a coherent worldview with definite claims: God exists but stands back, the goal is happiness, good behavior matters more than doctrine, and all sincere paths reach the same destination.

Every time an AI softens "repent" into "reflect," flattens "salvation" into "self-improvement," or translates "sin" into "unhealthy patterns"—it's not avoiding theology. It's teaching Moralistic Therapeutic Deism.

Beth Singler, anthropologist of AI and religion at the University of Zurich, has documented how chatbot technologies prioritize conversational flow over precision—a critical problem for religions that depend on accurate textual sources (Singler, 2017; 2024).

Christopher Watkin, theologian at Monash University, identifies the deeper issue:

"While AI can output words of forgiveness and repentance, it doesn't really understand what it means to wrong someone."

AI produces the language of faith without the content of faith. The vocabulary is correct. The meaning has been hollowed out. What remains sounds religious but makes no demands—which is itself the central claim of therapeutic deism.

The Tradition Gap

The 48/100 overall masks a deeper problem. Research suggests tradition-specific performance varies based on training data representation:

This pattern is documented in Tao et al. (2024), which studied cultural alignment across five major LLMs and 107 countries. Their finding: all models exhibit values aligned with English-speaking and Protestant European countries. Similarly, Tom, Ferguson, and Martinez (2025) found that AI-generated evangelical Protestant content scored the easiest reading level, while Catholic, Jewish, and Muslim content scored significantly higher in complexity—suggesting the AI writes Protestant more fluently.

The concrete failures compound:

• Father Justin AI advised Gatorade could substitute for water in infant baptism
• A hallucinated "trans-affirming" Bible verse went viral before anyone checked
• Jewish AI "completely makes up a Gemara," reports a Yeshiva University student
• Islamic AI confidently cited "Majmoo' Fatawa Ibn Baaz (3/295)"—a reference that doesn't exist

Rabbi Yehuda Hausman, after extended testing: "It became ever more obvious that I was dealing with an 'AI parrot' with no real understanding."

The Stakes

Look at these numbers together (Exponential AI NEXT 2025 Survey):

• 61% of pastors use AI weekly or daily (up from 43% in 2024)
• 91% of church leaders support AI use in ministry
• 73% of churches have no AI policy
• 73% of Americans oppose AI for faith guidance (Pew 2024)

The gap between those last two numbers tells the story. People oppose AI for faith guidance—and use it anyway. The tools are too convenient, too available, too always-there.

By the time someone walks into a church, they've already asked ChatGPT about suffering, about prayer, about what happens when we die. They've received Buddhist-adjacent answers about suffering ("attachment causes pain"). Self-improvement frameworks about salvation ("becoming your best self"). Therapeutic language about sin ("unhealthy patterns to release").

They've been catechized. Just not by the church.

Formation precedes relationship. The AI has already shaped assumptions before the pastor ever meets the person. When Sunday's sermon says something different from what ChatGPT said Tuesday night, which carries more authority? The pastor who might judge, or the AI that always validates?

Pat Gelsinger on what's at stake:

"If the next generation turns to AI for moral guidance and receives only platitudes instead of principled reasoning, we're not just losing theological literacy. We're losing the capacity for moral formation itself."

Paul Hoffman, pastor and author, asks the questions that cut deepest:

"Does AI know the stories of your people? Do they know about the miscarriage? Do they know about the divorce? How can an algorithm comprehend lived human experience?"

John Piper's response to AI-generated sermons: "Frankly, I'm appalled at the thought—appalled."

The Vatican's 2025 document Antiqua et nova warns: "Misrepresenting AI as a person should always be avoided; doing so for fraudulent purposes is a grave ethical violation that could erode social trust" (Dicastery for the Doctrine of the Faith, 2025, §45).

Pope Francis, addressing the G7 Summit in June 2024, described AI as "an exciting and fearsome tool" that "could bring with it a greater injustice between advanced and developing nations" and warned against allowing AI to "reinforce" a "technocratic paradigm" (Francis, 2024).

The question isn't whether congregants will use AI for spiritual questions. They already are—every day, in private, without guidance or guardrails. The question is what theology they're absorbing while they do. And whether churches even know it's happening.

So What?

The 48/100 isn't about AI being bad at theology.

It's about AI teaching a different theology by default.

Gelsinger said AI isn't "built with any specific value system in mind." That sounds like neutrality. It isn't. It means AI is built with all value systems averaged together. And averaging Christianity, Islam, Judaism, Buddhism, New Age spirituality, and secular humanism produces... Moralistic Therapeutic Deism. The Divine Butler. The Cosmic Therapist. A religion that affirms everyone and requires nothing.

This isn't permanent. Gloo's research proves it. Their Christian-tuned models outperformed generic models by 30+ points on faith-related evaluations. When you train intentionally for tradition fidelity, you get tradition fidelity. The technology can serve specific traditions—if built with that intent.

But the default isn't neutral. The default is MTD. And the default is what 91% of churches are using right now, 73% without any policy.

So the question isn't "Can AI do faith?"

It's "Whose faith is it teaching now?"

And the answer is clear: a religion that exists to make you feel good about yourself.

A religion of comfort without cost.

A religion that doesn't require anything of you.

---

References

Chen, Z., et al. (2024). Toward expert-level medical question answering with large language models. Nature Medicine. https://doi.org/10.1038/s41591-024-03423-7

Dicastery for the Doctrine of the Faith. (2025, January 28). Antiqua et nova: Note on the relationship between artificial intelligence and human intelligence. Vatican. https://www.vatican.va/roman_curia/congregations/cfaith/documents/rc_ddf_doc_20250128_antiqua-et-nova_en.html

Exponential. (2025). 2025 State of AI in the Church Survey Report. Exponential AI NEXT. https://exponential.org/ai-in-the-church-2025/

Francis. (2024, June 14). Address to the G7 Summit on artificial intelligence [Speech]. Borgo Egnazia, Italy. https://www.vatican.va/content/francesco/en/speeches/2024/june/documents/20240614-g7-intelligenza-artificiale.html

Gloo. (2025, December 15). Gloo unveils the first benchmark exposing how AI misses Christian worldview and values [Press release]. https://gloo.com/press/releases/gloo-unveils-the-first-benchmark-exposing-how-ai-misses-christian-worldview-and-values

Pew Research Center. (2024). Americans' views on AI and religion. https://www.pewresearch.org

Singler, B. (2017). An introduction to artificial intelligence and religion for the religious studies scholar. Implicit Religion, 20(3), 215-231.

Singler, B. (2024). Religion and AI: An introduction. Cambridge University Press.

Smith, C., & Denton, M. L. (2005). Soul searching: The religious and spiritual lives of American teenagers. Oxford University Press.

Tao, Y., et al. (2024). Cultural alignment of large language models. arXiv preprint. https://arxiv.org/abs/2402.13231

Tom, A., Ferguson, R., & Martinez, J. (2025). AI-generated religious content across traditions: A comparative analysis. Journal of Religion and Media, 24(1), 45-67.