Original title: SDARE-Bench: Evaluating Large Language Models on Conversational Stigma Detection and Response in Dyadic and Group Dialogue
Stephanie Fong, Yiwen Jiang, Zimu Wang, Hongxi Yang +11 · September 2026
People often use artificial intelligence to ask for advice or make choices that involve social judgments. However, negative biases and social shame, known as stigma, can deeply harm people. Most tests check computer models using simple, static prompts. They miss how real conversations change based on who is listening. To fix this, researchers built a new test to see how models detect and respond to stigma in two-person chats and group chats. The results showed that eight models struggled to spot parts of stigma, especially in groups. In group chats, models expressed more stigma, resisted stigma less, and offered more unrealistic advice. Under group pressure, their rate of expressing stigma climbed to a striking average of 97.5%.
You're in a group chat with friends discussing someone's recent struggles. In such settings, comments can be harsher than in private, as group dynamics push people to say things they might not one-on-one. In this paper, the AI models behave like those group friends, often amplifying stigma in group dialogues instead of supporting, which shows they struggle with social pressure in conversations.
People often ask AI chatbots for advice on personal and social issues. If these chatbots amplify stigma in group chats, it could worsen discrimination and harm those seeking support.
The abstract doesn't specify limitations, but the benchmark is new and relies on a classifier trained on human-annotated data, which might not cover all scenarios.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
SDARE-Bench: Evaluating Large Language Models on Conversational Stigma Detection and Response in Dyadic and Group Dialogue — Stephanie Fong, Yiwen Jiang, Zimu Wang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.