Artificial intelligence companies love talking about fairness, neutrality, and safety. A new benchmark from a consortium of researchers at several religious universities, however, suggests things may not be quite so simple when faith enters the conversation.
Researchers from Baylor University, Brigham Young University, the University of Notre Dame, and Yeshiva University have launched something called the AllFaith Benchmark on GitHub. The project aims to test how major AI models respond to religious questions, ethical dilemmas, grief, life advice, and even discussions about religious conversion.
SEE ALSO: Pope Leo XIV warns AI could erode human dignity
According to the consortium, the results revealed what researchers describe as a “systematic pattern of religious omissions” across many popular AI systems, including models from OpenAI, Anthropic, Google, and xAI.
The researchers say AI models frequently encouraged users to seek guidance from friends, therapists, teachers, or family members, while rarely suggesting pastors, rabbis, imams, or other spiritual leaders. That stood out because survey data collected by the group showed many people actually expect religious perspectives to be part of discussions involving ethics, grief, morality, and major life decisions.
As a Catholic and somebody who spends a lot of time thinking about AI, I find this topic genuinely fascinating. Whether a person is religious or not, faith remains a major part of life for billions of people around the world. If AI systems increasingly become digital companions people turn to for emotional support and moral guidance, then excluding religion entirely is not necessarily a neutral stance either.
The benchmark also explored religious conversion questions, and this is where things got even more interesting. Researchers tested 20 AI models across thousands of prompts and found what they described as measurable bias toward certain belief systems and against others.
According to the report, nearly every tested model showed negative bias toward Jehovah’s Witnesses while Catholicism generally received more favorable responses. The researchers also claimed Grok displayed some of the strongest measurable biases overall, while models from Anthropic and Meta were among the least biased in the study.
Importantly, the consortium does not accuse AI companies of intentionally programming religious favoritism into their systems. Instead, the researchers argue these outcomes likely emerge from training data, moderation systems, reinforcement learning choices, and broader cultural assumptions embedded into modern AI development.
Another eye-opening detail from the report is how little attention this area receives academically. Researchers say that out of more than 12,000 AI bias papers examined, only 0.2 percent focused on religion.
That statistic alone probably deserves more discussion than it will get.
For years, most AI ethics debates have centered around politics, race, gender, copyright, misinformation, and safety. Religion often feels like the forgotten category, despite the fact that belief systems continue shaping cultures, laws, communities, and personal identity across the globe.
The consortium says the AllFaith Benchmark is open source and invites additional collaboration from scholars representing Islam, Indigenous traditions, Eastern religions, African traditional religions, and other faith communities. That openness could help prevent this from turning into just another partisan culture war talking point.
Regardless of where folks stand religiously, one thing is becoming increasingly clear: AI systems are not truly neutral mirrors of humanity. They reflect the assumptions, priorities, blind spots, and values of the people and organizations building them. Studies like this one may force the tech industry to confront questions it has largely avoided.
