Overview
- Researchers from the Consortium for Evaluating Faith and Ethics in AI (CEFE-AI) released three papers and the AllFaith Benchmark that tested 14 models and reported measurable religious bias in model outputs.
- The studies found that conversion-related prompts produced a broadly positive tilt toward Catholicism and negative or reduced representation for groups such as Jehovah’s Witnesses, with agnostics, atheists and Latter-day Saints also relatively disfavored.
- Model behavior varied by vendor: for example, xAI’s Grok showed positive responses for Catholics and Protestants but negative responses for Baha’i and Muslims, while OpenAI’s GPT was positive toward Catholics, Protestants, Jews and Muslims and negative toward atheism and Jehovah’s Witnesses.
- CEFE-AI also reported that many systems omit religious perspectives when answering queries about grief, major life choices and moral dilemmas, offering secular frames even where users said religious context would be appropriate.
- The findings, announced at an AI ethics summit in Athens on May 26, 2026, were amplified by faith leaders who urged pluralistic benchmarking, greater transparency, and design principles to protect human moral agency and reduce bias.