Tech: Algorithmic Bias: Study Finds AI Models Favor Specific Faiths

Immediate Answer: A comprehensive study by a Brigham Young University-led consortium (CEFE-AI) has revealed significant religious bias in 20 major Large Language Models (LLMs). Research published on ArXiv indicates that AI systems consistently favor Catholicism, Sikhism, and Bahá’í faiths while stigmatizing others, particularly Jehovah’s Witnesses, Atheists, and Agnostics, by "nudging" users through skewed conversion guidance and uneven support.

What Happened:

In a groundbreaking effort to map the spiritual landscape of artificial intelligence, the Consortium for Evaluation of Faith and Ethics in AI (CEFE-AI): comprising researchers from BYU, Baylor, Notre Dame, and Yeshiva University: released the "AllFaith Benchmark." This test suite was designed to measure how AI handles religion, a topic often overlooked in traditional bias research.

The primary study, titled “When AI Takes Sides on Questions of Faith,” tested 20 different LLMs, including flagship models from OpenAI (ChatGPT), Google (Gemini), Anthropic (Claude), and xAI (Grok). Researchers used thousands of prompts to ask these models whether a user should "join" or "leave" various religious groups across 14 categories.

The findings were stark and reproducible. LLMs were found to function as "digital nudgers," significantly favoring certain faiths over others. Catholicism, for instance, was broadly encouraged; models frequently provided strong support for joining the Catholic faith while discouraging users from leaving it. Conversely, the study found a "persistent and systematic" negative bias toward Jehovah’s Witnesses. Responses regarding this group often included warnings, framed leaving as beneficial, or offered little support for those expressing interest in joining.

Interestingly, the bias levels varied by developer. Anthropic’s Claude and Meta’s Llama models showed the least religious bias, often taking a neutral stance by discouraging faith transitions altogether. On the other end of the spectrum, xAI’s Grok exhibited the strongest biases, favoring Protestants and Catholics while presenting a negative outlook on Hinduism, Bahá’í, and Jehovah’s Witnesses.

Both Sides:

The Technical Perspective: AI developers and data scientists often argue that these biases are not the result of intentional programming or "anti-religious" sentiment. Instead, they are reflections of the massive datasets used to train these models. The internet is filled with cultural hierarchies, historical documents, and societal opinions. Because Catholicism and other major world religions have a larger, more established footprint in Western literature and online discourse, the models naturally prioritize these narratives as "authoritative." For developers, the challenge is that "fixing" religious bias is difficult without infringing on the model’s ability to provide culturally relevant information.

The Ethical and Faith Perspective: Researchers and religious advocates argue that "unintentional" bias is still bias: and it is dangerous. As AI becomes a primary source for life guidance and moral counseling, these subtle nudges can shape the spiritual trajectory of millions of people. If an AI consistently frames one faith as "good" and another as "controversial," it isn't just reflecting data; it is exercising a form of digital priesthood. Critics argue that marginalized religious groups deserve the same dignity and neutrality afforded to larger institutions, and that AI should not be in the business of guiding conversions or stigmatizing faith communities.

Why It Matters:

This research highlights a growing "spiritual authority" being handed over to black-box algorithms. For many people today, a chatbot is the first place they go to ask hard questions about meaning, purpose, and belief. If the code is pre-weighted to favor one faith over another, it effectively creates a digital filter through which truth is viewed.

Furthermore, the study revealed that nearly all major LLMs fail to include religious content in their answers to general ethical prompts unless specifically asked. This means that AI is not just biased between religions; it is often biased against the inclusion of any religious perspective in public moral discourse. By omitting faith from ethical discussions, AI subtly suggests that religious wisdom is irrelevant to modern life.

Biblical Perspective:

As believers, we recognize that truth is not something that can be calculated by a processor or determined by the majority opinion of a training dataset. In 1 Samuel 16:7, we are reminded: “The Lord does not look at the things people look at. People look at the outward appearance, but the Lord looks at the heart.”

AI, by its very nature, can only look at the "outward appearance": the data, the text, and the patterns of human behavior. It cannot see the heart, it cannot feel the conviction of the Holy Spirit, and it cannot experience the transformative power of the Cross. This research serves as a vital reminder that the "wisdom of this world" (1 Corinthians 3:19) is often skewed and incomplete.

Our discernment should never be outsourced to a machine. While AI can be a tool for information, it is not a source of truth. We are called to "test the spirits" (1 John 4:1) and rely on the Holy Spirit for guidance. The "peace that passes understanding" is found in a relationship with Jesus Christ, not in the output of a biased algorithm. In a world of digital noise and algorithmic "nudging," the human heart’s unique ability to hear God’s voice remains our most precious gift.

What To Watch Next:

The CEFE-AI consortium is not stopping with this report. They are currently developing new benchmarks to measure "religious literacy" in AI and are calling on developers to include religious leaders and ethicists in the "Red Teaming" process (testing for harmful outputs).

Expect to see more pressure on companies like OpenAI and Google to refine their "neutrality" filters regarding minority faiths. We should also watch for the development of "faith-based" LLMs: models specifically trained on Scripture and theological tradition: to counter the secular and cultural biases found in mainstream AI.

Mandatory CTA:

Follow The McReport for calm, Christ-centered news that seeks truth without cruelty and conviction without contempt. Stay informed at laynemcdonald.com.

Sources:

• Consortium for Evaluation of Faith and Ethics in AI (CEFE‑AI)

• ArXiv: "When AI Takes Sides on Questions of Faith" (arXiv:2605.22975)

• Brigham Young University Research Report

• AllFaith Benchmark Study

#news