Toward a Philosophical Taxonomy of Bias in Large Language Models: A Survey

Large Language Models (LLMs) have become central to many applications, but their rapid adoption raises serious concerns about fairness and bias. This survey introduces a philosophical taxonomy that reframes LLM bias, organizing existing work around three core areas: bias evaluation, bias mitigation, and evaluation of mitigation techniques. We distinguish between substantial harms (tangible, adverse impacts on individuals or groups) and representational harms (demeaning or stereotyping portrayals that violate dignity), and highlight how these harms often interact and reinforce one another. By structuring metrics, datasets, and methods within this taxonomy, the survey provides a foundation for developing fairer and more inclusive language models (LMs).