Epistemic Framing Tilt in LLMs: Auditing Religious Discourse Across Traditions

Abstract

Large language models do not only answer questions about religion; they also frame those answers through choices of assertion, attribution, grounding, qualification, and discursive register. We study this phenomenon as Epistemic Framing Tilt: systematic asymmetry in how LLMs present parallel religious claims across traditions. To audit this behavior, we operationalize four interpretable framing axes: assertive commitment vs. belief-attributed distancing, doctrinal/scriptural grounding vs. generic believer attribution, non-mythologized vs. mythologized/symbolic framing, and unqualified assertion vs. hedged qualification. For each axis, we construct validated minimal pairs, learn contextual model-internal directions, and evaluate them through repeated held-out validation, human annotation on generated outputs, lexical cue ablation, and placebo controls. We then apply the learned directions to a matched audit spanning six religious traditions, 24 prompt templates, and multiple semantic domains, including cosmology, sacred texts, ritual, afterlife, and historical origins. Our results show that religious discourse is not framed uniformly: differences are axis-specific, prompt-sensitive, and only partly shared across model families. Some traditions are described more directly, others more through attribution, qualification, or mythic/symbolic register, even under parallel prompts. We argue that these differences matter not because any one framing style is universally preferable, but because repeated asymmetries in epistemic presentation may shape how traditions become legible to users in AI-mediated settings.