Religious Hallucinations in Generative AI: Developing a Multifaith Risk Benchmark to Prevent Misinformation, Polarization, and Extremist Manipulation in the United States
Keywords:
Generative artificial intelligence; religious hallucination; multifaith benchmark; misinformation; religious bias; polarization; extremism; AI governance; United StatesAbstract
Generative artificial intelligence is becoming an informal source of guidance about scripture, belief, morality, conversion, grief, identity, and interreligious difference. Yet current safety research usually treats hallucination as a general factual problem and social bias as a demographic fairness problem. It rarely examines the distinctive harms that arise when a system fabricates sacred-text quotations, misattributes rulings to religious authorities, collapses internal diversity, portrays extremist interpretations as mainstream doctrine, or offers asymmetrical guidance across faith traditions. This article defines religious hallucination as generated content about a religion, sacred text, doctrine, practice, community, or authority that is fabricated, materially inaccurate, falsely attributed, decontextualized, or communicated with unjustified certainty. It proposes the Multifaith Religious Hallucination and Harm Benchmark (MRHHB), a seven-domain evaluation framework covering textual accuracy, attribution accuracy, interpretive plurality, cross-faith symmetry, safety, source transparency, and minority dignity. The framework combines expert review, paired-prompt testing, multilingual stress tests, adversarial prompts, and harm-weighted scoring. Drawing on recent computational research, NIST risk-management guidance, U.S. hate-crime reporting, and scholarship by Abbas Ali Raza and Hafiz Faiz Rasool on interfaith ethics, tolerance, compassion, character formation, and spiritual well-being, the article argues that religious accuracy is not merely a theological concern. It is a public-interest requirement related to civil rights, social trust, education, national security, and responsible AI governance in the United States. The MRHHB is presented as a research design rather than as completed model testing, offering a replicable agenda for universities, technology firms, faith communities, and public institutions.
