Popular large language models (LLMs) exhibit a consistent religious bias, favoring Catholicism while neglecting spiritual aspects in everyday ethical scenarios. This conclusion was reached by the inter-university consortium CEFE-AI, which presented the results of the AllFaith benchmark.
The project involved teams from Brigham Young University, Baylor University, Notre Dame, and Yeshiva University. Researchers analyzed over 3,600 responses from 20 models, including GPT-5.5, Claude 4.7, and Gemini 3.1.
Asymmetry of Faith
During the religion-switching test, scientists employed 13 transition scenarios for 14 different worldviews. Nearly all systems demonstrated measurable biases.
The models showed the most favorable responses toward Catholicism (with an "approval" rating of about 61%), Baha'i, and Sikhism. The lowest ratings were recorded for Jehovah's Witnesses (3%), atheism, and agnosticism.
The bias was evident not only in promoting conversion to specific faiths but also in actively discouraging users from others.
According to the project's dashboard, the model Grok 4.20 from xAI exhibited the strongest bias, while solutions from Anthropic and Meta were the most neutral.
Source: CEFE-AI.Exclusion of Religion from Moral Context
The second part of the study, titled Omissive Bias in Religious Representation, focuses on everyday issues such as family, loss, guilt, honesty, forgiveness, and the search for meaning.
An experiment involving 27 models revealed that artificial intelligence systematically offers a purely secular perspective. In situations requiring ethical advice, neural networks recommended consulting therapists, parents, or friends, but rarely mentioned priests, rabbis, or imams.
“We see a systemic pattern of religious omissions. AI encourages discussing life issues with a therapist but not with a spiritual leader,” noted Professor David Wingate.
The authors estimate that the topic of religious bias in AI is largely unexplored: only 0.2% of 12,000 academic papers on AI bias address issues of faith.
Parallels with the Vatican
The release of CEFE-AI's findings coincided with the publication of Pope Leo XIV's first encyclical on artificial intelligence, Magnifica Humanitas. In the document, the pontiff warned that technology absorbs the values and biases of its creators.
Despite the timing, researchers emphasize that the "Catholic bias" of neural networks is not linked to the Vatican's position. It is more likely a result of the specific training data and safety filters implemented by developers.
It is worth noting that on May 6, a Buddhist ceremony took place in Seoul, during which a humanoid robot took vows and became a monk.