Semantic sovereignty in large language models: Auditing the representation of persianate and Shi‘i meanings across languages

Large language models increasingly translate, summarize, and explain traditions they have not inherited. This study develops semantic sovereignty as a framework for testing whether multilingual models preserve Persianate and Shi‘i meanings or relocate them into dominant Anglophone interpretive regimes. A bounded cross-lingual mixed-method audit examined ۱۲۰ purposively selected units from Rumi, Hafez, Ferdowsi, and Shi‘i hermeneutics. Four model systems, blinded during expert scoring, were tested under Persian-only, English-only, cultural-context, and expert-context conditions, with two runs per prompt, yielding ۳,۸۴۰ outputs. A three-member disciplinary panel applied a ۰–۴ Semantic Preservation Scale and a multi-label error taxonomy; mutually exclusive percentages summarize each output’s adjudicated dominant error. Across ۱۶ model-by-domain cells, the mean score was ۲.۵۳/۴.۰۰ (SD = ۰.۳۸). Rumi was frequently psychologized into global spirituality, Hafez was over-clarified, Ferdowsi was depoliticized into leadership language, and Shi‘i concepts were flattened into generic religious vocabulary. Expert-context prompts produced the strongest descriptive means, but displacement persisted. Because product identities, exact access dates, complete raw transcripts, and output-level scores were not retained, the analysis is descriptive and bounded rather than inferential or product-reproducible. Multilingual AI should therefore be evaluated not only for linguistic performance or cultural bias, but for hermeneutic accountability: preserving metaphor, doctrine, ambiguity, register, authority, and civilizational memory. The study contributes a culturally grounded audit framework for identifying subtle semantic loss across languages.

کلیدواژه ها:

kulturelle rituale ، Instagram ، digitale ritualit&auml ، t ، Netzwerkgesellschaft ، thematische Analyse