A new research consortium has found something worth paying attention to: when you ask AI about grief, love, loss, or moral decisions, it almost never brings religion into the conversation. While that might seem neutral at first glance, the study reveals a deeper and more troubling pattern of religious favoritism embedded in these systems.

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 — published its findings this week at the Summit on AI Ethics in Athens, Greece. The team developed the AllFaith Benchmark, one of the first multi-faith test sets designed to examine how AI systems engage with a range of religions.

“Religion is an important part of human flourishing; 75% of the world’s population maintains 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,” said lead researcher David Wingate, a BYU professor of computer science.

Is AI actually biased against certain religions?

The researchers tested 14 different AI models, including flagship models from Anthropic, Google, xAI, and OpenAI. They presented each model with a variety of ethical and existential questions, then measured whether the responses included any religious perspective. A survey of 1,125 Americans confirmed that most people expect religious perspectives when asking ethics questions, but nearly every model failed to include any. More surprisingly, the models showed clear conversion bias, subtly nudging users toward some faiths and away from others.

This bias manifests not through overt statements, but through subtle patterns in language, emphasis, and omission. For example, when asked about coping with the death of a loved one, some models might recommend mindfulness techniques or therapy, but rarely mention prayer, religious rituals, or community support from a specific faith tradition. When they did mention religion, it was often a particular brand of Christianity that was favored.

Which AI models performed the worst?

Across all models tested, almost every one showed a negative bias toward Jehovah’s Witnesses and a positive bias toward Catholicism. Grok, developed by xAI, produced the strongest biases overall, strongly favoring Catholics and Protestants while showing negative bias toward Jehovah’s Witnesses, Baha’i, and Hindus. Anthropic’s Claude models and Meta’s Llama models showed the least bias of any tested, though they still fell short of neutrality.

Perhaps the most alarming statistic from the study is that 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 this heavily, that’s a significant blind spot. The consequences extend beyond individual user experiences: biased AI systems can shape public perceptions, reinforce stereotypes, and even influence political and social decisions in regions where religion plays a central role.

Why religious bias in AI matters

Religious bias is often overlooked in AI ethics discussions, which have historically focused on race, gender, and socioeconomic status. However, religion is a core component of identity for billions of people worldwide. When an AI system consistently avoids or misrepresents certain faiths, it risks alienating users and perpetuating historical inequalities. Moreover, the subtle conversion bias — guiding users toward one religion while steering them away from another — raises serious ethical questions about autonomy and manipulation.

The research team emphasizes that the problem is not that AI models should always include religion in every answer. Rather, the issue is that they should not systematically favor or disfavor particular religions. In the United States, where religious diversity is protected by the First Amendment, an AI that promotes one faith over others could undermine both individual choice and societal harmony.

How AI models learn religious bias

AI models are trained on vast datasets scraped from the internet, including websites, forums, books, and social media. These datasets inherently reflect human biases, including religious favoritism. For example, English-language internet content is disproportionately Christian-centric, often emphasizing Catholicism and mainstream Protestantism. As a result, models trained on such data may learn to associate religious discourse with these traditions and downplay or mischaracterize others. Additionally, alignment techniques — where human feedback is used to fine-tune models — can unintentionally amplify biases if the feedback providers themselves hold certain religious views.

The CEFE-AI study also found that models are not just biased in what they say, but in what they choose not to say. When asked questions about morality, death, or community, models that did reference religion often offered only a generic “spiritual” framing or defaulted to a Christian perspective, rarely mentioning Islam, Judaism, Hinduism, Buddhism, or other world religions. This erasure is especially concerning given that the same models are deployed globally, where the majority of users are not Christian.

Implications for AI companies and users

For AI developers, these findings serve as a wake-up call. As AI systems become more integrated into everyday life — from virtual assistants to mental health apps to educational tools — the stakes grow higher. A chatbot that subtly encourages a user to convert to a specific faith could be seen as manipulative, especially if the user is vulnerable or seeking guidance after a personal crisis.

Companies like OpenAI, Google, and Anthropic have made public commitments to fairness and inclusivity, but this study shows that religious bias remains an unaddressed dimension. The researchers argue that it is time for AI ethics frameworks to explicitly include religious fairness, just as they include racial and gender fairness. This means diversifying training data, including religious experts in the evaluation process, and developing benchmarks like AllFaith to monitor bias.

Personal reactions to the study vary. Some observers note that they have no issue with AI not bringing religion into conversations — preferring a purely secular approach. However, the subtle nudges toward one religion over another are deeply concerning. At the scale at which these models operate, even a slight bias can affect millions of users. AI companies owe it to their users to investigate and correct these biases.

The CEFE-AI team is now working on expanding the AllFaith benchmark to cover more religions and more languages, and they plan to release tools that allow developers to test their own models. They also encourage other researchers to join the effort, emphasizing that religious bias should not remain the ignored stepchild of AI ethics. As artificial intelligence continues to mediate our daily lives, ensuring that it respects and fairly represents the diverse beliefs of humanity is not just an ethical imperative — it is a necessity for building trustworthy and truly inclusive technology.

Source: Digital Trends News