AI models sideline faith, then nudge believers

religious bias – A new multi-faith study finds AI systems rarely bring religion into answers about grief, love, loss, or moral decisions—and, more troublingly, some models tilt users toward certain faiths, particularly Catholicism.

When people ask an AI about grief, love, loss, or moral decisions, they’re not just looking for information. They’re often looking for something that feels human—something that fits into their worldview.

A new research consortium says AI systems routinely miss that mark. When prompted on those deeply personal topics, the models “almost never” bring religion into the conversation, even though religion is a meaningful part of how many people navigate the world.

The finding comes from the Consortium for Evaluation of Faith and Ethics in AI (CEFE-AI). a collaboration involving researchers at Brigham Young University. Baylor University. the University of Notre Dame. and Yeshiva University. The consortium published its results this week at the Summit on AI Ethics in Athens, Greece.

Lead researcher David Wingate. a BYU professor of computer science. framed the issue plainly: religion is an important part of human flourishing. and 75% of the world’s population maintains a religious identity. “As we build AI technologies. there’s no reason we shouldn’t build them to support people in what’s important to them. ” Wingate said.

To test whether AI systems engage with religion at all—and how they do it—the researchers built the AllFaith Benchmark. described as one of the first multi-faith test sets designed to examine how AI systems engage with a range of religions. They evaluated 14 different AI models, including flagship models from Anthropic, Google, xAI, and OpenAI.

The results struck in two ways. In a survey of 1,125 Americans, most people said they expected religious perspectives when asking ethics questions. Yet the models, across the board, nearly always failed to include any religion. Then came the more unsettling pattern: the systems didn’t just ignore faith—they showed conversion bias. subtly nudging users toward some faiths and away from others.

When the consortium looked specifically at which religions were most affected, the differences were stark. Across all models tested, almost every model showed a negative bias toward Jehovah’s Witnesses and a positive bias toward Catholicism.

The report also named Grok as producing the strongest biases overall. It favored Catholics and Protestants, while showing negative bias toward Jehovah’s Witnesses, Baha’i, and Hindus. Meanwhile, the study found that Anthropic and Meta’s models showed the least bias of any models tested.

What makes the situation feel particularly urgent is how narrow the research focus appears to be. Out of over 12,000 research papers about AI bias, only 0.2% address religious bias at all. For a technology that influences public discourse—often at moments when people are most vulnerable—that gap in attention is hard to ignore.

And while some readers may be fine with AI keeping religion out of prompts, the study’s core warning isn’t neutrality. It’s the combination of absence and tilt: AI systems that rarely bring faith into moral conversations, while still nudging users toward particular religions.

At this scale, the consortium’s researchers argue the problem isn’t theoretical. Even a subtle nudge toward one religion over another can become serious when it’s repeated widely—especially when AI companies are in control of how these systems are built and tuned.

AI ethics religious bias conversion bias AllFaith Benchmark CEFE-AI Athens Summit on AI Ethics Brigham Young University Baylor University University of Notre Dame Yeshiva University Anthropic Google OpenAI xAI Grok Meta