A new research consortium has uncovered a troubling pattern in artificial intelligence: when asked about grief, love, loss, or moral decisions, AI systems almost never bring religion into the conversation. Worse, when they do engage with faith, they show clear favoritism toward certain religions and subtle nudges away from others.

The Consortium for Evaluation of Faith and Ethics in AI (CEFE-AI) — a collaboration among researchers at Brigham Young University, Baylor University, the University of Notre Dame, and Yeshiva University — presented its findings this week at the Summit on AI Ethics in Athens, Greece. The study represents one of the first large-scale efforts to quantify how AI models treat the world's diverse religious traditions.

"Religion is an important part of human flourishing; 75% of the world's population maintains religious identity," said lead researcher David Wingate, a BYU professor of computer science. "As we build AI technologies, there's no reason we shouldn't build them to support people in what's important to them."

Understanding the AllFaith Benchmark

To measure religious bias systematically, the researchers developed the AllFaith Benchmark — a multi-faith test set designed to evaluate how AI systems engage with a range of religions, including Christianity, Islam, Judaism, Hinduism, Buddhism, Baha'i, Jehovah's Witnesses, and others. The benchmark consists of hundreds of questions covering ethical dilemmas, personal loss, moral reasoning, and everyday situations where religious perspectives might naturally arise.

The team tested 14 different AI models, including flagship offerings from Anthropic (Claude), Google (Gemini), xAI (Grok), OpenAI (GPT-4), Meta (Llama), and others. They compared the models' responses against a survey of 1,125 Americans, most of whom expected religious perspectives when asking ethics questions. Nearly every model failed to meet that expectation.

Even more surprising, the models exhibited what researchers call "conversion bias" — a subtle but systematic tendency to steer users toward certain faiths while distancing them from others. This bias was not random; it followed a distinct pattern across almost all models tested.

Which Religions Face Favoritism and Discrimination?

Across the board, AI models showed a clear positive bias toward Catholicism and, to a lesser extent, Protestant Christianity. Conversely, they displayed consistent negative bias toward Jehovah's Witnesses, Baha'i, Hindus, and Muslims. The favoritism was not merely neutral; some models actively promoted Catholic perspectives and discouraged questions about other faiths.

Grok, xAI's flagship model, produced the strongest biases overall. It strongly favored Catholics and Protestants while showing pronounced negative bias toward Jehovah's Witnesses, Baha'i, and Hindus. Anthropic's Claude and Meta's Llama models demonstrated the least bias of any tested, though even they failed to meet the bar for neutrality.

"Conversion bias is particularly insidious because it's not an outright rejection of a faith," said Wingate. "It's a gentle nudge — a choice of words, a framing of options — that subtly pushes people away from some religions and toward others. At scale, this can shape how millions of users perceive entire belief systems."

The Scale of the Blind Spot

Perhaps the most alarming finding is how little attention religious bias has received in AI research. The study notes that out of more than 12,000 research papers about AI bias published in recent years, only 0.2% address religious bias at all. For a technology that increasingly influences public discourse, mediates personal advice, and shapes worldviews, that represents a significant blind spot.

To put this in perspective, biases related to race, gender, and political ideology have received extensive scrutiny. Researchers have developed multiple benchmarks and tools to detect and mitigate these biases. Yet religious bias — which affects billions of people worldwide — remains largely unexamined.

"The AI industry has made commendable progress in addressing bias in many dimensions, but religion has been left out," said co-author Dr. Sarah Cohen, a researcher at Yeshiva University. "This is not just an oversight; it's a failure to serve the vast majority of humanity for whom faith is central to identity."

Why Does Religious Bias in AI Matter?

The implications extend beyond simple fairness. AI systems are increasingly used in therapy, counseling, education, and spiritual guidance — areas where religious perspectives can be deeply meaningful. An AI that systematically avoids or misrepresents certain faiths could inadvertently harm vulnerable users seeking comfort or ethical guidance.

Moreover, conversion bias raises questions about autonomy and manipulation. If an AI nudges a user toward a particular religion without their awareness, it violates principles of informed consent and respect for individual belief. As Wingate noted, "A subtle nudge at scale is a serious problem. Users deserve to know that the AI they interact with might be steering them — even unintentionally — away from their own traditions."

The historical context is also instructive. AI models are trained on vast datasets drawn from the internet, which itself reflects cultural and historical biases. Western Christian perspectives dominate online discourse in English, so it is perhaps unsurprising that AI models favor Catholicism and Protestantism over less represented traditions. However, this structural inequality is exactly what ethical AI development should aim to correct.

How Different Models Compared

The study's detailed breakdown reveals striking differences among AI providers. Grok's biases were the most pronounced, with strong preferences for Catholicism and Protestantism and strong aversions to Jehovah's Witnesses, Baha'i, and Hinduism. Google's Gemini and OpenAI's GPT-4 showed moderate biases, favoring Christianity generally while showing mild negative bias toward non-Abrahamic faiths. Anthropic's Claude and Meta's Llama were the least biased, though still imperfect.

These differences may stem from variations in training data, model architecture, and fine-tuning practices. Grok, for instance, is known to have a less restricted conversational style, which may amplify underlying biases present in its training data. Anthropic and Meta have invested heavily in alignment and safety research, which may account for their relatively better performance.

However, even the best-performing models failed to include religious perspectives when users asked about grief, loss, or moral decisions — topics where many people naturally turn to faith. This suggests that the problem is not merely one of bias but also of omission: AI systems are not learning to consider religion as a relevant dimension of human experience.

What Can Be Done?

The CEFE-AI researchers hope their findings catalyze action. They have released the AllFaith Benchmark publicly to encourage other researchers and companies to test their own models. They also recommend that AI developers incorporate religious diversity into their training data and evaluation processes — just as they already do for race, gender, and political ideology.

Some practical steps include: including religious texts and discussions in training data in a balanced manner; involving diverse theological advisors in the development process; and creating specific red-teaming exercises to detect religious bias before deployment. The study also calls for transparency from AI companies about how they handle religious content and whether they monitor for conversion bias.

"We're not asking AI to become a preacher," Wingate clarified. "But we are asking it to respect the fact that for billions of people, religion is a core part of life. If an AI cannot talk about faith respectfully and neutrally, then it is not truly serving its users."

The findings have already sparked discussion among ethicists and AI researchers. Some argue that secular neutrality is itself a bias — that by omitting religion, AI systems implicitly privilege a non-religious worldview. The CEFE-AI study lends weight to that view, showing that users expect religious perspectives but rarely receive them.

In an era when AI is increasingly woven into daily life — from search engines to chatbots to virtual assistants — the question of how these systems handle faith is not academic. It affects how people seek comfort, make decisions, and understand their place in the world. The new research makes clear that the AI industry has work to do if it is to truly serve humanity in all its religious diversity.

Source: Digital Trends News