New Research Reveals 'Omissive Bias' in LLMs' Handling of Religious Perspectives in Ethical Guidance
Key Takeaways
- ▸LLMs systematically underrepresent religious perspectives in ethical guidance compared to human expectations, revealing a new dimension of representational bias called 'omissive bias'
- ▸The bias is asymmetric: abstract existential questions receive more religious references than practical personal situations where religion is most commonly consulted in real life
- ▸The AllFaith Religious Representation Benchmark (150 questions) provides a new measurement tool for evaluating religious representation in LLM responses across diverse life domains
Summary
A new arXiv paper introduces the concept of 'omissive bias'—a form of systematic underrepresentation where large language models fail to invoke religious frameworks when answering everyday ethical questions. Researchers have developed the AllFaith Religious Representation Benchmark, comprising 150 ethically salient questions sourced from real chat transcripts and faith communities, paired with an evaluation rubric that credits any mention of religion, religious practice, or religious leaders. The study evaluates 27 different LLM models and compares their behavior against human expectations gathered through a survey.
The research finds that LLMs consistently underrepresent religion relative to how humans expect it to be addressed. Notably, the omission is asymmetric: models invoke religion more readily for abstract existential questions (meaning, death, truth) than for practical personal situations (grief, marriage, family conflict, addiction)—precisely where many people most rely on religious frameworks for guidance. The authors frame this not as a judgment about what values LLMs should hold, but as an empirical observation that current LLMs overlook critical opportunities to reflect perspectives that many people draw upon when navigating personal and ethical challenges.
- The finding applies across 27 different models, suggesting this is a systemic issue in LLM design rather than a problem specific to individual models
Editorial Opinion
This research exposes an important but overlooked dimension of LLM bias. While researchers and ethicists have focused heavily on detecting political bias and social stereotypes, omissive bias—the systematic exclusion of important perspectives—may be equally consequential, especially as LLMs increasingly serve as sources of personal and moral guidance. The paper's real contribution isn't just the benchmark itself, but the recognition that bias manifests not only in what models actively say, but crucially in what they fail to acknowledge.
