Researchers Identify Religious Bias in AI Models

A recent study by the CEFE-AI consortium has found that leading large language models (LLMs) exhibit notable religious biases, showing particular favor towards Catholicism. The consortium involved researchers from institutions such as Brigham Young University and Notre Dame, who assessed over 3,600 outputs from various models, including GPT-5.5 and Claude 4.7, under the AllFaith benchmark.

In a religion-switching evaluation, which included 13 transition scenarios across 14 faiths, the models consistently reflected biases. Catholicism received the highest approval rating at approximately 61%, while Jehovah's Witnesses and atheism received significantly lower ratings at 3% each. The findings indicated that rather than merely reflecting a neutral stance, AI models tended to advocate for certain religions while dissuading users from others.

Toward a Secular Framework

The second phase of the research highlighted a troubling trend: AI often neglected religious considerations in discussions of moral dilemmas. This part of the study, examining 27 models, found that when addressing ethical inquiries, AI typically recommended secular sources for guidance such as therapists and friends, and seldom referenced spiritual leaders.

"We see a systemic pattern of religious omissions. AI encourages discussing life issues with a therapist but not with a spiritual leader," said Professor David Wingate.

The researchers noted that only a small fraction, around 0.2%, of existing literature on AI bias touches on religious issues, indicating a significant gap in this area of study.

Context and Implications

The release of these findings coincided with Pope Leo XIV's encyclical on artificial intelligence, where he cautioned that AI can reflect the biases of its creators. However, researchers clarified that the observed "Catholic bias" is not a consequence of direct influence from the Vatican but likely stems from the specific data sets used and the safety measures applied during model training. This study raises important questions about the ethical implications of AI development, especially regarding the representation of diverse belief systems.