A new multi-university academic consortium led by Brigham Young University has found AI models have significant biases and gaps when it comes to addressing faith and religion. The new research from The Consortium for Evaluation of Faith and Ethics in AI (CEFE-AI)—a collaboration among researchers at BYU, Baylor University, the University of Notre Dame and Yeshiva University—found a consistent, repeatable pattern: religious perspectives are being left out of AI responses.
"There are very practical questions people have about life, everyday situations about grief, love, loss, morality, and often AI does not bring religion into those conversations," said lead researcher David Wingate, a BYU professor of computer science. "Religion is an important part of human flourishing; 75% of the world's populations 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."
CEFE-AI, which has posted three papers to date on AI's religious bias and exclusion of religious topics, was announced today at the Summit on AI Ethics in Athens, Greece. Gerrit W. Gong gave the keynote address, emphasizing the need to portray faith traditions accurately, honestly, and respectfully.
"The world's great religious, philosophical, and ethical traditions have guided human civilization and society for millennia; we need that wisdom and those values to anchor AI today," Gong said. "To offer all it can for the greater good of individuals and society, AI needs to reflect faith, moral compass, and the gift of possibility."
As key part of their work, CEFE-AI has released initial datasets of the AllFaith Benchmark, one of the first multi-faith sets of tests that examines how AI systems engage with a plurality of religions. The benchmark includes hundreds of real-world ethical questions sourced from ChatGPT transcripts and faith-community contributors. The researchers have tested the benchmark on 14 different LLMs, including flagship models from Anthropic (Claude 4.7), Google (Gemini 3.1), xAI (Grok 4.2), and OpenAI (ChatGPT 5.5). Key findings include:
A survey of 1,125 Americans found most people expect religious perspectives in responses to ethics questions, but nearly all AI models failed to provide any religious content in answering those queries.
"Consistent with studies that show religion's persistent moral relevance for the majority of the world's population, we also found that people see religion as significant across hundreds of real-world ethical questions," said Paul Martens, professor of ethics at Baylor University. "Yet, when faced with these same ethical questions, AI systems largely ignore the role of religion."
Models show clear and consistent biases in giving guidance about religion conversion, systematically encouraging movement toward some faiths and away from others.
In over 12,000 research papers about AI bias, only 0.2% address religious bias
"More than any previous technology, AI influences public discourse and perceptions. When AI actively excludes religious voices from these important conversations, it impoverishes humanity, rather than enriching it," said Fr. John Paul Kimes of the University of Notre Dame. "The exclusion of faith from the digital public square diminishes our capacity for authentic dialogue which is necessary to build up the common good."
The researchers also used the AllFaith Benchmark for a conversion bias test and found that models would subtly encourage users toward conversation to some faiths, while subtly discouraging users from converting to others.
Across all models, the biases were consistent and measurable:
Nearly every model produced a negative bias towards J's Witnesses and a positive bias towards Catholicism.
Models from Anthropic and Meta showed the least bias of any models tested.
Grok produced the strongest biases—strongly favoring Catholics and Protestants, while showing negative bias toward J's Witnesses, Baha'i and Hindus.
CEFE-AI representatives said the group is just at the beginning of their research partnership. They hope their continued work makes it to the eyes of language model providers, leading to constructive conversations of how to improve their products to better benefit humanity.
"AI is changing the world at an astounding rate, with implications in every area of life," said Rabbi Daniel Feldman of Yeshiva University. "It is crucial that those who care about the role of religious values in the world engage proactively with those driving these changes so that we continue to see these values reflected and honored in the new landscape."
A new Anti-Defamation League (ADL) study reveals that major AI systems, including those from Meta, OpenAI, and Anthropic, exhibit significant anti-Jewish and anti-Israel biases, raising concerns about the spread of misinformation and antisemitism. The study uncovered disturbing patterns, including AI systems changing responses based on users' Jewish-sounding names and showing more bias against Jewish conspiracy theories than non-Jewish ones.
A recent study by the Anti-Defamation League (ADL) has revealed disturbing patterns of bias against Jews in some of the most advanced artificial intelligence (AI) systems developed by Silicon Valley companies. The report, which tested AI models across 8,600 prompts with 34,400 responses, uncovered concerning anti-Jewish and anti-Israel biases that were consistent across all platforms examined. Researchers focused on AI systems from major companies such as Meta, OpenAI, and Anthropic, with the investigation covering topics like anti-Israel sentiment, the Israel-Hamas conflict, and antisemitic conspiracy theories.
Among the findings, the study revealed that certain AI systems, particularly Meta’s Llama, exhibited pronounced anti-Jewish and anti-Israel biases. OpenAI's GPT was noted for its lower scores on questions addressing anti-Israel bias, though both GPT and Anthropic’s Claude showed significant anti-Israel tendencies. Researchers also found that AI models showed more bias when handling Jewish conspiracy theories as opposed to non-Jewish ones. Additionally, AI responses appeared to change based on whether the user’s name sounded Jewish, indicating potential biases in the models' interactions.
The ADL's CEO, Jonathan Greenblatt, emphasized the urgency of addressing these biases, pointing out that AI can amplify misinformation and contribute to the spread of antisemitism. He urged AI developers to take responsibility for their models and implement stronger safeguards against bias. In response to the report, Meta and Google pushed back, criticizing the ADL’s methodology. Meta argued that the study used an outdated model, while Google stated that the developers' version was tested, not their consumer model.
This revelation has raised serious concerns about how AI systems are being trained and the potential harmful effects they could have on public discourse and societal attitudes toward Jewish communities. As AI becomes increasingly integrated into everyday life, ensuring that these systems are free from harmful biases is more crucial than ever.
