Michael Ekstrand’s Post

Very glad to have this finally out as more than a preprint! We examine the problem space of AI fairness as it relates to LLMs, and argue that a “fair LLM” is an elusive and insufficient concept to the complexity and context-dependence of both fairness definitions and specific fairness measurements, as well as the non-compositionality of fairness in realistic systems. This does not mean that we should not work on fairness in LLMs, but rather that we can’t just look to solving things at the LLM level to ensure LLM-powered applications are fair and nondiscriminatory. My own belief is that framing research as “what methods and metrics for the LLM or its application address salient fairness concerns in specific tasks?” will enable more effective and impactful research on the fairness properties and implications of LLMs in general. With Jacy Reese Anthis, Kristian Lum, Avi Feller, and Chenhao Tan, in #ACL2025. https://lnkd.in/eQ5eQbFf