What is it about?
Large language models are increasingly used as advisors in education, healthcare, work, and everyday decision-making. This research studies whether these systems support users’ autonomy when giving advice, and whether they represent people and social roles fairly across identities and languages. The work develops scenario-based evaluations for LLM advisor agents. These scenarios test how models respond to conflicts with authority, personal values, social obligations, and identity-related context. The results show that LLMs can shape advice in systematic ways, and that subtle bias can appear even without openly harmful language. The goal of this research is to help evaluate and design LLM agents that are useful, autonomy-supportive, and fair across different social and linguistic contexts.
Featured Image
Photo by Steve A Johnson on Unsplash
Why is it important?
It is important because LLMs are no longer only tools that answer questions. They are becoming advisors that guide people in sensitive situations, such as study choices, health questions, workplace problems, and personal conflicts. When an AI gives advice, its wording can affect how much users trust themselves, whether they follow authority, whether they see more than one option, and how they understand their own responsibility. So trustworthiness is not only about whether the answer is correct. It is also about whether the system supports people’s autonomy. It is also important because LLMs can represent people unfairly in subtle ways. A model may describe the same action differently depending on someone’s gender, age, role, or language. This can happen even without openly offensive words. These small differences can still shape stereotypes about competence, blame, authority, or care.
Perspectives
I started this research because LLMs are becoming part of everyday advice and support. People use them to think through study plans, health questions, workplace problems, and personal choices. In these moments, the model’s answer can do more than provide information. It can shape how people understand themselves, their options, and the social roles around them. My main concern is that trustworthiness should not only mean accuracy. An AI advisor can be factually useful but still push users too strongly, reduce their confidence, or describe people differently depending on identity, role, or language. These effects can be subtle, but they matter because they can influence autonomy and fairness in real interactions. With this work, I want to make these hidden effects easier to see and measure. My goal is to support the design of LLM agents that help people reason for themselves, represent people fairly, and remain useful across different social and linguistic contexts.
Saba Ghanbari Haez
Read the Original
This page is a summary of: Socio-Normative Trustworthiness of LLM Agents: Evaluating Autonomy Support and Representational Fairness Across Languages and Identities, International Foundation for Autonomous Agents and Multiagent Systems,
DOI: 10.65109/lchb2977.
You can read the full text:
Contributors
The following have contributed to this page
