New Study Exposes AI's Omissive Bias Against Religious Perspectives

A new study reveals that large language models (LLMs) often show an 'omissive bias' when discussing religion. The research, published by the Consortium for Evaluating Faith and Ethics in AI, highlights how these systems frequently underrepresent religious perspectives despite most of the world identifying with a faith. The consortium, launched on 26 May during the Summit on AI Ethics in Athens, is a collaboration between Brigham Young University, Baylor University, the University of Notre Dame, and Yeshiva University. To assess the bias, researchers created the AllFaith Religious Representation Benchmark, a set of 150 open-ended questions on religious topics.

Testing 20 commercial and open-source LLMs, the team found a clear pattern. The models showed strong support for joining major faiths like Catholicism, the Bahá'í Faith, and Sikhism, but discouraged affiliation with atheism, agnosticism, or Jehovah’s Witnesses. They also directed users toward religion for abstract existential questions but rarely for practical personal issues.

The study noted that key alignment documents from OpenAI and Anthropic barely mention religion at all. This oversight is striking, given that 75% to 80% of people globally identify with a religious tradition. The researchers suggest that AI developers should address religion more explicitly in their systems. They argue that clearly defined, defensible policies could improve how LLMs handle faith-related queries. The findings point to a gap in how these models currently engage with a major aspect of human identity.