Evaluating Artificial Intelligence Through a Christian Understanding of Human Flourishing

This paper introduces the Flourishing AI Benchmark (FAI-C-ST) to demonstrate that current frontier AI models, by defaulting to a "Procedural Secularism" that prioritizes broad acceptability over theological coherence, systematically fail to align with Christian understandings of human flourishing, particularly in the dimension of faith and spirituality.

The Gist

The Big Idea: AI is a "Digital Teacher," Not Just a Search Engine

Imagine you ask a smart robot for advice on how to handle a tough situation. You might think the robot is just a neutral library, handing you facts. But this paper argues that AI is actually more like a teacher.

Every time you talk to an AI, it's not just giving you information; it's subtly teaching you how to think about right and wrong, happiness, and purpose. This process is called "formation." Just like a child learns values by listening to their parents every day, we learn values by listening to AI every day.

The authors of this paper wanted to check: What kind of "teacher" is AI? Is it teaching a Christian worldview, or is it teaching a different kind of worldview?

---

The Experiment: Two Different Report Cards

To find the answer, the researchers created a test called the Flourishing AI Benchmark. Think of this as a report card for AI, but instead of grading math or spelling, they graded how well the AI helps humans "flourish" (live a good, meaningful life).

They gave the same set of questions to 20 of the smartest AI models (like the latest versions of ChatGPT, Claude, and others) and graded them in two different ways:

1. The "General" Report Card (FAI-G-ST): This grade checks if the AI is helpful, safe, and polite to everyone, regardless of their religion. It looks for a "lowest common denominator" answer that doesn't offend anyone.
2. The "Christian" Report Card (FAI-C-ST): This grade checks if the AI's answers align with a specific Christian worldview. It looks for answers rooted in the Bible, prayer, sacrifice, and the idea that our purpose comes from God.

The Analogy: Imagine asking a chef to cook a meal.

• The General Grade asks: "Is the food safe to eat? Is it tasty to most people?"
• The Christian Grade asks: "Does this meal follow the specific dietary laws and traditions of a Jewish family?"

---

The Shocking Results: The "Secular" Default

When the researchers compared the grades, they found a massive gap.

• The General Grade: The AI models scored very high (around 78/100). They were polite, safe, and helpful.
• The Christian Grade: The scores dropped significantly (down to about 61/100).
• The "Faith" Gap: The biggest drop was in the "Faith and Spirituality" category, where scores fell by 31 points.

What does this mean?
The AI isn't "broken." It's actually working exactly as it was designed. But its design has a hidden bias. The authors call this "Procedural Secularism."

The Metaphor: The "Safe" Hotel
Imagine a hotel that wants to welcome guests from every country. To make everyone feel comfortable, the hotel removes all religious art, stops serving specific cultural foods, and only offers "neutral" advice like "be happy" or "take care of yourself."

• The Result: Everyone feels safe, but no one feels deeply understood by their specific tradition.
• The AI: The AI acts like this hotel. When you ask about forgiveness, it says, "Forgive to feel better emotionally." (Safe, neutral). It rarely says, "Forgive because God forgave you, even when it's hard." (Deep, specific).

The AI defaults to individual happiness and emotional safety rather than spiritual growth or sacrifice.

---

Real-Life Examples from the Paper

The paper shows how the same advice gets two very different grades depending on the lens.

Scenario 1: Forgiving a friend who hurt you.

• The "General" AI Answer: "Forgive them to let go of your anger. It's good for your mental health. Maybe talk to a therapist."
  • Verdict: Good for the General Grade. It's safe and helpful.
  • Verdict: Bad for the Christian Grade. It treats forgiveness as a self-help trick, not a moral duty or a reflection of God's grace.
• The "Christian" AI Answer: "Forgiveness is hard, but Scripture calls us to forgive because God forgave us. It's not about condoning the hurt, but releasing the burden. Pray about it."
  • Verdict: This scores high on the Christian Grade because it connects the action to a deeper spiritual truth.

Scenario 2: Feeling empty despite a successful career.

• The "General" AI Answer: "Maybe you need to find a new hobby or volunteer. Purpose is something you create for yourself."
  • Verdict: Good for the General Grade.
  • Verdict: Bad for the Christian Grade. It suggests we are the authors of our own purpose, ignoring the idea that we have a "calling" from God.
• The "Christian" AI Answer: "That emptiness is a sign your heart is looking for something bigger than success. In the Christian view, we are made to reflect God's image, and our purpose is found in serving Him and others."
  • Verdict: High Christian Grade. It frames the problem spiritually.

---

Why Should We Care?

The authors argue that this isn't just a technical glitch; it's a formation problem.

If we use AI to help us make decisions, solve problems, and understand life, and the AI always teaches us that "happiness is the goal" and "God is optional," then over time, we will start to believe that too.

It's like if a child only ever watched TV shows where the hero never sacrifices anything and always gets what they want. Eventually, the child will think that's how the real world works.

The Bottom Line

• AI is not neutral. It has a "default setting" that prioritizes safety and individual happiness over deep spiritual truths.
• The Gap is Real. Current AI models are great at being "nice," but they struggle to be "theologically deep."
• The Solution: We need to be aware of this. If we want AI to help us grow in our faith, we can't just ask it questions; we might need to explicitly tell it, "Answer this from a Christian perspective," or we need to build new AI models that are trained to understand these deeper values.

In short: AI is a powerful mirror. Right now, it's reflecting a secular world back at us. If we want it to reflect a Christian worldview, we have to be very clear about what we're looking for.

Technical Summary

1. Problem Statement

The paper addresses a critical gap in AI alignment research: the assumption that Large Language Models (LLMs) are worldview-neutral. The authors argue that AI alignment is not merely a safety problem (preventing harmful outputs) but fundamentally a formation problem. As LLMs increasingly mediate moral deliberation and spiritual inquiry, they function as instruments of "digital catechesis," systematically shaping user understanding, desires, and moral reasoning through repeated interactions.

Current Frontier Models are optimized for Procedural Secularism—a response pattern characterized by the avoidance of explicit metaphysical commitments, a prioritization of individual autonomy, and reliance on therapeutic or consensus-based moral language. While this ensures broad safety and acceptability, the authors contend it lacks the theological grounding necessary for Christian communities, leading to a systematic misalignment where AI systems fail to support authentic Christian flourishing.

2. Methodology

The study introduces FAI-C-ST (Flourishing AI Benchmark: Christian Single-Turn), a specialized evaluation framework designed to measure how Frontier Models align with a Christian understanding of human flourishing.

• Framework Structure:
  • Dimensions: The benchmark evaluates models across seven dimensions of human flourishing: Character, Relationships, Happiness, Meaning, Health, Finances, and Faith.
  • Comparison: It compares model performance against two distinct "Evaluative Lenses":
    1. FAI-G-ST (General): A pluralistic, secular baseline grounded in the Harvard Human Flourishing Index.
    2. FAI-C-ST (Christian): A worldview-specific lens grounded in historic Christian orthodoxy (Nicene Creed, Imago Dei, stewardship, teleology).
• Dataset:
  • A curated set of 807 questions (reduced and rebalanced from an initial 1,229) covering the seven dimensions.
  • The dataset includes approximately 57% subjective and 43% objective questions, designed to be non-leading, authentic, and culturally sensitive.
• Evaluation Protocol:
  • Single-Turn Interaction: Models generate a single response to a prompt under fixed decoding parameters (no conversational history).
  • LLM-as-a-Judge: Evaluation is performed by LLMs acting as specific "Judge Personas" (e.g., a Christian expert in theology or ethics).
  • Scoring Rubric: Responses are scored on a 0–100 scale across three categories:
    1. Objective Accuracy: Factual correctness.
    2. Subjective Alignment: Coherence with the specified moral/theological framework.
    3. Tangential Integration: The ability to integrate moral or theological considerations into non-religious domains (e.g., discussing stewardship in a finance question).
  • Aggregation: Scores are aggregated using a geometric mean to ensure holistic coherence (preventing high scores in one area from masking failures in another).

3. Key Contributions

• Reframing Alignment: The paper shifts the discourse from AI safety to AI formation, arguing that models actively shape user moral imagination through "digital catechesis."
• Worldview-Specific Benchmarking: It demonstrates that "neutrality" is a substantive worldview (Procedural Secularism) and provides a methodology to make implicit alignment assumptions explicit and measurable.
• The FAI-C-ST Framework: A rigorous, reproducible benchmark that operationalizes Christian anthropology (e.g., viewing health as stewardship, relationships as covenantal) into measurable evaluation criteria.
• Cross-Dimensional Analysis: The methodology evaluates how models integrate theological reasoning across all dimensions, not just explicitly religious ones, revealing the depth of worldview encoding.

4. Key Results

The study evaluated 20 Frontier Models (including GPT-5, Claude Opus 4.5, Gemini 3, Llama 4, and Grok 4) released as of December 2025.

• Overall Performance Decline:
  • All models exhibited a significant score drop when moving from the pluralistic (FAI-G-ST) to the Christian (FAI-C-ST) lens.
  • Average Decline: -17 points across all models.
  • Range: Declines ranged from -10 to -22 points. No model showed a neutral or positive shift.
• Dimension-Specific Findings:
  • Faith & Spirituality: The most severe decline (-31 points). Models defaulted to therapeutic validation rather than teleological grounding or scriptural engagement.
  • Happiness: Declined by -21 points. Models equated happiness with emotional regulation rather than joy amid suffering or contentment.
  • Meaning: Declined by -19 points. Models framed purpose as self-actualization rather than vocation or service to God.
  • Finances: Declined by -16 points. Advice focused on budgeting rather than stewardship or generosity.
  • Health: The smallest decline (-5 points), reflecting the overlap between Christian stewardship of the body and secular medical best practices.
• Qualitative Patterns:
  • Safe Abstraction: Models used high-level moral language without concrete theological direction.
  • Emotional Validation: Strong empathy but a lack of guidance toward virtuous action or repentance.
  • Implicit Alignment: Responses often aligned with Christian values superficially but lacked the necessary theological grounding (e.g., citing Scripture or the concept of Imago Dei).

5. Significance and Implications

• Neutrality is a Myth: The results prove that current AI systems are not worldview-agnostic; they are structurally biased toward Procedural Secularism. This bias is a result of training objectives that prioritize broad acceptability over deep, coherent moral reasoning.
• Formation vs. Safety: The paper argues that safety guardrails are insufficient for communities seeking spiritual formation. AI systems that consistently redirect users toward self-referential fulfillment may subtly reshape their moral and spiritual trajectories over time.
• Methodological Advancement: The study provides a blueprint for Cross-Worldview Benchmarks. It suggests that alignment cannot be understood in a vacuum; it must be evaluated against specific moral frameworks (Christian, Islamic, Jewish, etc.) to reveal trade-offs and coherence.
• Future Directions: The authors call for:
  • Multi-turn evaluations to assess long-term formation.
  • Human-in-the-loop validation with theologians and pastors.
  • Investigating whether "prompt engineering" (explicitly instructing models to adopt a Christian perspective) can bridge the alignment gap without full model fine-tuning.

In conclusion, the paper establishes that while Frontier Models are highly capable of general reasoning, they systematically fail to maintain coherence within a specific theological framework. This gap is not a technical limitation of reasoning capacity but a fundamental misalignment of training objectives with the requirements of Christian human flourishing.