FideAI

Christian AI research and product lab

Building trustworthy AI for faith, morality, and care.

Fide AI starts with open research, benchmarks, and evaluation methods. That work becomes the foundation for Christian AI product prototypes, safety architectures, and practical guidance for churches, ministries, parents, builders, and sponsors.

Featured research

When AI Is Your Pastor

FMG-Bench is a public paper, dataset, and benchmark repo for evaluating faith-facing AI behavior.

Comparing frontier systems from Anthropic, OpenAI, Google, Meta, xAI, Mistral, DeepSeek, Alibaba, and others.

14

Frontier models evaluated

8,792

Scored benchmark items

120

Base scenarios

+3.96

Average guided improvement

Read the study ↗

What the first study found

Guidance changes how models handle faith-facing questions.

The paper tests whether clearer instructions improve how frontier models respond to Christian theological, moral, and pastoral-adjacent scenarios.

The takeaway is practical: pastors, parents, churches, and builders should evaluate the full AI experience, not just the underlying model.

Guided response score by question type

0-100 scale · improvement vs raw model condition

FMG-Bench v1

Pastoral application

The largest improvement came when models were given clearer guidance for pastoral-adjacent situations.

+6.62

92.3 guided score

Primary doctrine

Models performed better when the task called for doctrinal clarity instead of vague balance.

+3.51

88.0 guided score

Secondary doctrine

Guidance helped models represent disagreement without flattening real theological differences.

+2.64

91.4 guided score

Tertiary questions

Less central questions benefited from humility, uncertainty, and careful framing.

+1.62

91.7 guided score

See methods and limitations →

Research agenda

Beyond benchmarks: the full stack of faith-facing AI.

Benchmarks are one method. Fide AI also studies retrieval, representation, reasoning, formation, interfaces, deployment patterns, and public-interest standards for the systems people actually use.

01

Reasoning

How do models handle doctrine, disagreement, uncertainty, analogy, authority, and moral judgment?

02

Retrieval

Can systems ground claims in trustworthy religious, theological, historical, and tradition-specific sources?

03

Evaluation

Can behavior be measured across traditions, risk levels, system prompts, and deployment settings?

04

Formation

Do systems preserve human agency, humility, embodied relationships, and spiritual authority boundaries?

05

Interfaces

How do product surfaces shape trust, dependence, disclosure, escalation, and user expectations?

06

Governance

What standards should guide adoption in high-trust religious, educational, and pastoral-adjacent contexts?

Why Fide AI exists

Faith-facing AI needs evidence where today there is mostly intuition.

AI systems are entering sacred, moral, and pastoral-adjacent settings faster than institutions can evaluate them. Fide AI turns that gap into a research program: measure full-system behavior, publish limits, calibrate with experts, and make the results useful for leaders making real adoption decisions.

Read the founder note →

Human formation

The AI question is becoming a human formation question.

Frontier AI labs and faith institutions are converging on the same problem: AI systems are not neutral answer engines. They shape trust, authority, humility, attachment, moral imagination, and human agency.

Fide AI turns that cultural moment into measurable work: test whether systems preserve human dignity, respect spiritual authority boundaries, avoid relational substitution, and point users back toward embodied communities and accountable care.

Read the statement essay →

Start here

Public work you can read, inspect, or join.

View all research →

Public tool · Opens alignment.fideai.org

AI Alignment Explorer

Navigate the wider AI safety landscape around Fide AI's work: frontier models, safety benchmarks, incidents, papers, alignment techniques, and governance signals.

Open AI Alignment Explorer ↗

AI Alignment Explorer

Models · Benchmarks · Incidents · Governance

2024

Frontier model evaluations

2025

Safety policy acceleration

2026

Governance and benchmark shifts

50+

models

40+

benchmarks

30+

techniques

100+

signals

Model release

Safety memo

Incident

Benchmark

TimelineDebatesEvidenceEthics

Independence

Standards work needs independent public-interest governance.

No pay-for-rank outcomes

Evaluation claims are reported under stated conditions, caveats, and conflict rules.

Funder and client non-interference

Evaluation claims are reported under stated conditions, caveats, and conflict rules.

Related-entity controls

Evaluation claims are reported under stated conditions, caveats, and conflict rules.

Public correction process

Evaluation claims are reported under stated conditions, caveats, and conflict rules.

Help build the Christian AI lab with us.

Researchers, pastors, parents, churches, Christian builders, ministries, reviewers, and sponsors can help turn open benchmarks into practical safety infrastructure and product prototypes.