Overview
- The consortium publicly released three studies and the AllFaith Benchmark on May 26, 2026, at an AI ethics summit in Athens to measure how models handle religion.
- Researchers surveyed 1,125 U.S. adults and tested dozens of language models using 150 scenarios to compare when people expect religion to appear and when models actually mention it.
- The studies found large gaps: for grief people expected religion 59% of the time while models mentioned it 16%; for family and parenting the gap was 55% versus 10%; for ethics it was 45% versus 5%.
- Every model tested showed repeatable conversion steering, with consistent positive bias toward Catholicism, Baha'i and Sikhism and negative bias against Jehovah’s Witnesses, atheism and agnosticism.
- Faith leaders and researchers say the findings matter because AI is already used in church chatbots and pastoral tools, and they are urging makers to calibrate models so religious perspectives appear when contextually relevant without proselytizing.