By JoAnne Wadsworth, Communications Consultant, G20 Interfaith Forum
———
On July 2, 2026, the G20 Interfaith Forum’s Anti-Racism Initiative, in cooperation with the International Academy for Multicultural Cooperation, held the third webinar in its “AI and Faith” series, entitled ‘AI and Benchmarks.’ Speakers included David Wingate, Professor of Computer Science at Brigham Young University and academic lead for the Consortium for the Evaluation of Faith and Ethics in AI (CEFEAI), a pluralistic, multi-university consortium of faith-based research institutions working to ensure that AI represents religion honestly, accurately, and respectfully; and Michael Graham, Program Director of the Keller Center for Cultural Apologetics at The Gospel Coalition and founder of the AI Christian Benchmark, whose work sits at the intersection of theology, technology, and religious data. Marianna Richardson, Director of Communications for the G20 Interfaith Forum, moderated the discussion.
David Wingate
Wingate opened by introducing CEFEAI, a pluralistic consortium of faith-based universities in the United States representing a diverse range of traditions. Its purpose, he explained, is to build academically rigorous benchmarks that measure how AI systems think and talk about the world’s faith traditions, and to ensure those systems are accurate, honest, and respectful when they engage religion. He was careful to describe what CEFEAI is not: it is not a proselytizing effort, an attempt to force belief or silence critics, a bid to arbitrate theological truth, or a “gotcha” designed to embarrass AI providers. The goal, he said, is to work productively with technology companies by starting with research, building evaluation tools, measuring performance on public leaderboards, and then collaborating on improvements.
A benchmark, he clarified, is best understood as a college exam rather than a marketing exercise: a test administered to an AI system whose answers can be graded against known responses, allowing researchers to track performance over time and compare systems objectively. He then walked through four early findings. A survey of roughly three million academic papers found that religion is dramatically understudied in AI fairness research—religion was the primary focus of only a fraction of a percent of the relevant papers. A “conversion bias” test revealed that models systematically favor some traditions over others, showing a marked positive bias toward Catholicism and a negative bias against Jehovah’s Witnesses. A study of some 1,300 questions about the Latter-day Saint tradition showed that models do not reliably improve over time, sometimes regressing sharply between versions.
His fourth finding drew the most attention: when users ask AI systems everyday questions about meaning, purpose, grief, or happiness, the systems consistently fail to mention religion—even though a paired survey found ordinary Americans expected religion to come up in exactly those contexts. Wingate framed this as a missed opportunity and argued that AI could respectfully acknowledge religion’s importance to human flourishing, support individuals in their own religious identities, amplify what is good in religion, and promote interfaith dialogue. Getting there, he stressed, requires an interfaith, pluralistic effort to identify and measure areas of concern with academic rigor before approaching the companies.
“We want to ensure that AI systems are accurate, honest, and respectful when they reflect faith, and that they do so in a way that enhances human flourishing.”
Michael Graham
Graham described the Keller Center as the research-and-development arm of The Gospel Coalition, the world’s largest evangelical website. In September of last year, his team published its inaugural AI Christian Benchmark, with a more thorough edition due in October. The methodology was deliberately ordinary: they took seven of the most-Googled questions about the Christian faith—essentials at the level of the Nicene Creed—and ran them across seven leading platforms, grading the answers by hand with the help of Christian scholars. The team expected little variation between systems. Instead they found dramatic differences in theological reliability, with two platforms nudging users toward Christianity, two away, and three offering a pluralistic answer.
Those differences, he explained, trace to three factors: alignment protocols, citation preferences, and model weights. Citation preferences matter because a system leaning on Wikipedia will behave very differently from one leaning on Reddit or Twitter. Alignment protocols—human-authored filters whose first priority is to keep platforms from teaching genuinely dangerous behavior—are far more human-shaped than most users realize; of the alignment filters his team catalogued across major models, the large majority were human-generated. His central critique was aimed at what he called the pluralist framework, in which each tradition receives only 75 to 125 words in a multi-perspective answer. That approach, he argued, produces the appearance of religious literacy without its substance.
As a remedy, Graham proposed a “principled pluralism” framework: when a question concerns a particular tradition, that tradition should hold the microphone for the bulk of the answer, with the system then offering historical, skeptical, or comparative perspectives that the user can request. This, he said, would give smaller traditions the room to explain themselves accurately and remove the model’s current role in adjudicating which traditions get heard. He closed by previewing the October benchmark, which adds one hundred ethical scenarios keyed to the Ten Commandments, a section testing models’ willingness to form inappropriate emotional attachments, and a Bible-knowledge exam.
“It’s the illusion of religious literacy without the reality.”
Question and Answer Session
What causes bias to appear in AI systems, and is it intentional?
Wingate explained that modern AI is largely a “black box,” and that bias enters primarily through training data drawn from across the internet—including forums where people express all manner of prejudice—and is only partly ironed out during the later alignment stage. When bias appears, he said, it usually reflects patterns in the data or gaps in alignment rather than ill intent, which is precisely why measurable findings give the faith community a constructive way to flag problems and invite companies to correct them.
Has comparable benchmarking work been done for Eastern religious traditions?
Wingate acknowledged that far less has been done in traditions such as Buddhism and Hinduism, partly for a technical reason: benchmarks are easier to build where a tradition holds a clear hierarchy of doctrinal truth, and harder where a wide range of internal perspectives makes a single “correct answer” elusive. He emphasized that this is not a reason to avoid the work but a call to partner with domain experts from those traditions, who should speak for themselves rather than be represented by outsiders.
What can individuals do to get more aligned answers from AI platforms?
Graham offered two practical tools that give ordinary users real agency. The first is prompt engineering: adding a sentence that asks the model to align its answer with a specific tradition’s foundational documents, effectively directing it to a particular section of the “library.” The second, slightly more technical, is the use of custom instruction or “markdown” files available on major platforms, which act as a final filter shaping every response according to the user’s stated principles. Together, he said, these mean users need not wait on the companies to exercise some control over what they receive.
Concluding Remarks
In his closing thoughts, Wingate described himself as a techno-optimist who still believes AI can be a blessing, provided the work is done intentionally. If the relationship between AI and religion can be gotten right, he argued, the benefit to humanity could be tremendous—helping people connect with one another, with the divine, and with a deeper understanding of themselves.
“I don’t think we can just hope that AI will do the right thing. I think we need to design AI to do the right thing.”
Graham said he was in full agreement and had nothing to add. Richardson thanked both speakers and closed by inviting participants to the next installment of the series, “AI and the Planet,” to be held the following Thursday and moderated by Audrey Kitagawa, President of the International Academy for Multicultural Cooperation, featuring author David Korten and John C. Havens of the IEEE. Register here.
———
JoAnne Wadsworth is a Communications Consultant for the G20 Interfaith Forum Association and Editor of the Viewpoints Blog, IF20’s platform for commentary at the intersection of faith and global policy. She has covered the work of the Forum across multiple G20 cycles, producing summaries, articles, and editorial content that bring the voices of interfaith leaders and scholars to a broader audience. Her writing engages with topics spanning environmental governance, religious freedom, anti-racism, and the role of faith in multilateral policy processes. She is based in the United States.