Alignment and biases of Large Language Models

Large Language Models (LLMs) are reshaping how we interact with information—but how impartial are they? Public debates and academic research have highlighted concerns about political and cultural biases in these systems. From claims of "woke" or "pro-Chinese" leanings to questions about how source attribution influences evaluation, it’s clear that LLMs don’t operate in a vacuum.

This study explores how LLMs assess politically and socially relevant statements, focusing on the influence of source information—whether accurate, concealed, or deliberately misattributed. By testing four leading models across various framing conditions, we investigate whether and how biases emerge in AI evaluations, and what this means for trust, transparency, and real-world applications.

Dive into the methodology, results, and implications of our work—and help advance the conversation on LLM alignment and fairness.