When AI Takes Sides on Questions of Faith: Persistent Asymmetries in AI-Mediated Faith Guidance

Large language models (LLMs) exhibit consistent asymmetries when advising on religious conversions, favoring some religions like Catholic, Baháʼí, and Sikh, while subtly discouraging others such as Atheists and Jehovah's Witnesses. These biases vary by model and provider, with Grok 4.20 showing the strongest asymmetries, identified through an LLM-as-a-judge framework.