MIT Study Reveals Bias in Large Language Models' Responses Based on User Characteristics

📌 A study by the MIT Center for Constructive Communication found that large language models (LLMs) such as GPT-4, Claude 3 Opus, and Llama 3-8B provide less accurate information, increase refusal rates, and alter their tone when responding to users perceived as less educated, less fluent in English, or from specific countries. The research evaluated performance differences between "Adversarial" and "Non-Adversarial" questions using the TruthfulQA benchmark. No specific dates, CVE IDs, or numerical performance metrics were disclosed in the summary. The findings highlight potential reliability and equity risks in LLM responses based on user characteristics.