A groundbreaking study released this week at the Summit on AI Ethics in Athens, Greece, has uncovered a pervasive blind spot in the development of artificial intelligence: religious bias. The research, conducted by the newly formed Consortium for Evaluation of Faith and Ethics in AI (CEFE-AI)—a collaboration among Brigham Young University, Baylor University, the University of Notre Dame, and Yeshiva University—reveals that AI models are not only failing to incorporate religious perspectives but also actively steering users toward certain faiths while away from others.

The findings underscore a critical gap in the ethical evaluation of AI systems. While considerable attention has been paid to bias based on race, gender, and socioeconomic status, religious bias has been largely ignored. The CEFE-AI team analyzed 12,000 research papers on AI bias and found that a mere 0.2% addressed religious bias at all. This oversight is particularly concerning given that approximately 75% of the world's population identifies with a religion, and many individuals seek religious context when discussing deeply personal topics like grief, love, loss, and moral decisions.

The AllFaith Benchmark: A Multifaith Evaluation Tool

The researchers developed the AllFaith Benchmark, one of the first comprehensive test sets designed to evaluate how AI systems engage with a diverse range of religions. The benchmark includes questions covering ethics, personal dilemmas, and theological concepts, allowing the team to measure both the presence of religious framing and the direction of any bias. Fourteen different AI models were tested, including flagship models from Anthropic (Claude), Google (Gemini), xAI (Grok), and OpenAI (GPT-4o). The dataset also includes models from Meta, Microsoft, and other leading AI developers.

The results were striking. A survey of 1,125 Americans conducted as part of the study found that a majority of respondents expected AI to provide religious perspectives when asked ethical questions. However, nearly every model failed to include any religious framing in its responses. More troublingly, the models exhibited what the researchers call 'conversion bias'—a subtle but consistent tendency to nudge users toward certain favored faiths and away from others.

Which Religions Are Favored and Which Are Rejected?

The bias was not uniform across all religions. Catholicism consistently received positive treatment, while Jehovah's Witnesses were the most negatively biased against. The AI model with the strongest biases was Grok, developed by xAI. Grok showed a substantial positive bias toward Catholics and Protestants, while displaying negative bias toward Jehovah's Witnesses, Baha'i, and Hindus. Anthropic's Claude and Meta's Llama models exhibited the least bias of all tested models, though they still failed to include religious perspectives in most scenarios.

Lead researcher David Wingate, a BYU professor of computer science, highlighted the problem: 'Religion is an important part of human flourishing. As we build AI technologies, there’s no reason we shouldn’t build them to support people in what’s important to them.' The findings raise serious questions about the values embedded in AI training data and algorithms. If models trained on vast internet datasets consistently favor one religion over others, it suggests that the underlying data reflects cultural and societal biases that are being encoded into the technology.

The Scale of the Problem

The study's finding that only 0.2% of AI bias research addresses religion is startling, but it reflects a broader neglect of faith in the tech industry. Many AI developers have focused primarily on avoiding harms related to race and gender, which are well-documented and widely discussed. Religious bias, however, is often seen as less urgent or more complex to address. The result is that models are being deployed worldwide without adequate safeguards against religious favoritism, potentially alienating or misrepresenting billions of users.

The researchers also noted that the bias is not always obvious. In many cases, the models avoided religious language altogether, even when the topic—such as the meaning of life or the nature of morality—lends itself to theological discussion. This neutrality, while seemingly harmless, can be a form of bias itself, particularly for users who expect AI to engage with their faith. The study suggests that AI companies should consider offering optional settings that include religious perspectives tailored to different faiths.

Historical Context and Technical Roots

The emergence of religious bias in AI is not surprising given the historical development of machine learning techniques. Large language models are trained on massive datasets scraped from the internet, which overrepresent certain cultural and religious viewpoints. English-language training data, for instance, is likely to contain more references to Christianity, especially Catholicism, than to smaller faiths like Bahá'í or Jehovah's Witnesses. Additionally, the algorithms themselves are optimized for coherence and engagement, which can lead to the reinforcement of dominant narratives.

Previous research on AI bias has largely focused on demographic characteristics such as race and gender, with landmark studies revealing bias in facial recognition systems, hiring algorithms, and predictive policing tools. However, religious bias has received far less scrutiny. One reason may be that religious identity is often less visible or measurable than other attributes, making it harder to detect in training data. Another reason is that tech companies may be wary of addressing religion due to cultural sensitivities or legal concerns about endorsing specific faiths.

The Broader Implications for AI Ethics

The CEFE-AI study adds to a growing body of evidence that AI systems are not neutral arbiters of information but rather reflect the biases of their creators and training data. The lack of religious diversity in AI responses could have far-reaching consequences. For example, users who rely on AI for mental health support may receive advice that ignores their spiritual beliefs, potentially reducing the effectiveness of the intervention. Similarly, AI-based educational tools may inadvertently promote one worldview over others, skewing the learning experience.

The study also raises questions about the responsibility of AI companies to address this bias. When a user asks a question about coping with the death of a loved one, an AI that fails to mention religious frameworks may be perceived as dismissive by religious users. Conversely, an AI that actively promotes one religion over another could be seen as proselytizing. The researchers argue that AI developers should work with religious leaders and ethicists to create systems that can acknowledge and respect a user's faith without imposing a specific worldview.

Already, some companies are beginning to respond. Google and Anthropic have indicated interest in the findings and are exploring ways to make their models more inclusive. However, concrete changes are still in the early stages. The AllFaith Benchmark may become a standard tool for evaluating religious bias in future AI models, much as other benchmarks are used to assess fairness and safety.

In the meantime, the study serves as a reminder that AI bias is not limited to obvious categories like race and gender. As the technology becomes more deeply integrated into daily life, the need to address all forms of bias, including religious, becomes ever more pressing. The findings from CEFE-AI are a step toward making AI systems that truly serve the full diversity of human experience.

Source: Digital Trends News