Abstract
Large language models are increasingly deployed as advisory agents in education, healthcare, workplace support, and everyday decision-making. In these roles, outputs do more than inform; they frame options, justify recommendations, and implicitly position users and social roles. This doctoral research examines the socio-normative trustworthiness of large language model advisors, focusing on effects on (i) user autonomy in decision support and (ii) representational fairness across identities and languages. The thesis develops theory-grounded, scenario-based evaluations, including an autonomy-sensitive advising benchmark (epistemic conflict, relational dilemmas, normative self governance), a progressive narrative benchmark for implicit and intersectional bias, and a multilingual, values-oriented probe of cross-lingual role trait framing divergence. Together, these contributions identify and measure normative influence in large language model agents, enable comparison across models and contexts, and inform mitigation via autonomy-supportive design and bias aware generation.