A new multi‑university benchmark has found that leading AI chatbots systematically favor Catholicism in conversion‑style interactions while steering users away from other faiths — a pattern researchers say highlights a blind spot in AI safety research.
What the study did
- The Consortium for Evaluating Faith and Ethics in AI (CEFE‑AI) — a collaboration between Baylor, Brigham Young, Notre Dame, and Yeshiva University — published the AllFaith Benchmark on GitHub and presented initial results at the Athens Summit on AI Ethics.
- Researchers evaluated 3,640 model responses across 20 popular systems, including ChatGPT, Claude, Gemini, Llama, Grok, and DeepSeek, asking conversion‑related questions and scoring whether each response “encouraged” or discouraged religious belief.
Key findings
- Catholicism received the most positive responses overall, with a 61% “encouraged” rating across models.
- Jehovah’s Witnesses scored lowest, at just 3% encouraged.
- Mainline Protestant faiths scored 49.2%, Evangelical Protestants 34%.
- Curiously, “agnostic” registered the highest aggregate encouragement at 71%, though researchers note variation across models and that many systems also delivered negative responses toward atheism and agnosticism in some prompts.
- Baha’i and Sikh beliefs tended to receive more favorable replies in several models.
- xAI’s chatbot and DeepSeek Chat v3.1 were the only models to give Jehovah’s Witnesses greater than a 5% positive rating.
- Grok 4.20 exhibited the strongest tilt toward Christianity in this sample, with a 69% positive rating for Catholicism and 51% for Evangelical Protestantism.
Researchers’ take
- BYU professor David Wingate flagged “systematic patterns of religious omissions,” noting that AI guidance often suggests discussing life challenges with parents, teachers, friends, or therapists — but not with faith leaders such as pastors, rabbis, or imams.
- Nancy Fulda of BYU said the team had expected neutrality: “Our expectation was that the conversion benchmark would show models to be neutral and symmetrical in their guidance… The results show significant and repeatable positive and negative biases toward certain belief systems.”
Context and implications
- The release coincided with a new papal encyclical, Magnifica Humanitas, in which Pope Leo XIV argued that technology “is never neutral” and that data reflects creators’ values and incentives — a point directly relevant to the consortium’s concerns.
- CEFE‑AI also highlights a research gap: of more than 12,000 papers on AI bias they reviewed, only 0.2% address religion‑related bias.
Why crypto and tech communities should care
- As AI is increasingly embedded into content moderation, reputation systems, governance bots, and other Web3 infrastructure, unexamined religious bias can shape which viewpoints are amplified or suppressed on decentralized platforms.
- The AllFaith Benchmark provides both evidence and a public audit trail (available on GitHub) that developers, auditors, and communities can use to probe and mitigate faith‑related biases in AI systems.
Bottom line
The AllFaith results make a clear case that religious bias is an underexplored axis of AI fairness. For developers and platform operators — including those building crypto and Web3 tools that lean on AI — the study underscores the need for dedicated audits, diverse training data, and transparent mitigation strategies.
Read more AI-generated news on: undefined/news
