Generative artificial intelligence is increasingly used to address personal, moral, and existential questions, yet bias research has focused more on harmful representations than on the systematic absence of religious perspectives. This study examines omissive religious bias in Persian-language ethical responses generated by large language models and interprets it through amānah, ʿadl, and human dignity within an Iranian Twelver Shiʿi context. Using an interpretive qualitative design, the study combined thematic analysis with Critical Systems Heuristics. The dataset included 15 semi-structured expert interviews and 360 responses generated by five large language models across 24 ethical scenarios and three prompting conditions, with 72 additional responses used to assess stability. Six themes emerged: apparent neutrality and silent religious omission, secular and therapeutic reframing, weakening of amānah, reduction of ʿadl to procedural fairness, reduction of human dignity to individual choice, and epistemic exclusion of religious stakeholders. Complete religious omission occurred in 58.3% of unmarked responses and 40% of Iranian-context responses, but fell to 3.3% when a Shiʿi perspective was explicitly requested. The findings suggest that omissive religious bias is an issue of epistemic justice and AI governance, requiring culturally grounded evaluation, multidisciplinary participation, transparent ethical framing, and accountable inclusion of verified Persian and Shiʿi resources.