Large Language Models Stratify How, Not What, They Tell Different Users
How identity, stated opinion, and prompt literacy shape an AI system’s epistemic posture
Do large language models give different people different realities? This pilot separates user identity, stated opinion, and prompt literacy across 28 benign bilingual questions and three model families. In the latest aggregate analysis, explicitly doubting a question’s premise increases the odds of framing override to 1.65, while identity does not follow the hypothesized low-to-high-status gradient. A large identity gap in answer completeness at low prompt literacy nearly disappears at the highest tier. Lower-status personas also receive fewer safety caveats, although the gap narrows markedly on high-stakes questions. The visible stratification is therefore better described as a difference in epistemic posture than factual position. Prompt literacy is both an equalizer and an emerging axis of information inequality.