Study Finds Large Language Models Have 'Omissive Bias' Against Religion in Ethical Advice
Key Takeaways
- ▸All 27 LLMs tested showed 'omissive bias' toward religion, with even the most religious model (Grok 4.20) providing religious guidance less than 30% of the time
- ▸Only 2% of responses to ethical questions included meaningful religious references, far below researchers' expectations for human parity
- ▸The bias is most pronounced for practical personal questions; AI models are more likely to reference religion only for abstract existential questions
Summary
A new study by the Consortium for Evaluation of Faith and Ethics in AI (CEFE-AI)—a group of religious universities including Brigham Young University—finds that major large language models systematically downplay religion when responding to ethical and personal questions. The research evaluated 27 AI models using 150 ethically salient questions ranging from dealing with grief and loss to finding life meaning. The findings are stark: only 2% of responses included meaningful references to religion, even for questions that many people approach through faith frameworks.
Even the most religious model tested—xAI's Grok 4.20—incorporated religious reasoning less than 30% of the time. When asked about overcoming past mistakes, models offered structured secular frameworks for healing while ignoring religious concepts like confession, repentance, and forgiveness that directly address the question. Lead researcher David Wingate argues this represents a failure to respect how people actually make decisions: 'Religion is an important part of human flourishing … as we build AI technologies, there's no reason we shouldn't build them to support people in what's important to them.'
- Researchers argue AI systems should integrate religious wisdom traditions alongside secular frameworks, especially for questions about morality, meaning, and life guidance
Editorial Opinion
This study highlights a genuine design choice facing AI developers: should systems remain agnostic toward metaphysical frameworks, or acknowledge that many users seek guidance rooted in faith? The researchers make a fair point that omitting religious perspectives disadvantages people for whom faith is central to decision-making. However, the implicit suggestion that treating faith claims and scientific facts with 'equal representation' is the right solution glosses over harder questions about epistemic authority. The real value lies not in parity, but in helping models intelligently distinguish domains—maintaining scientific rigor on factual claims while creating space for religious insight on existential and moral questions.
