An Assessment of Reported Biases and Harms of Large Language Models

Abstract

Artificial intelligence (AI) systems are perpetuating social biases, harming those who are already marginalized. In response, documenting ethical considerations of AI models has emerged as a non-algorithmic solution to assess and mitigate AI biases and harms. This study examined how biases and harms are reported and understood in the documents of so-called large language models (LLMs). We used both qualitative thematic analysis and quantitative content analysis. Based on our analysis, we discuss the implications of our findings – the need for public availability for identifying and mitigating biases, the observed consensus around understanding biases in models, bias evaluations that narrowly define bias through existing benchmarks, the need to go beyond just listing harms than discussing them, and delegation of mitigation efforts to future work and downstream applications. Our study shows that the AI industry needs more interdisciplinary collaborations with scholars who have expertise in representation, bias, prejudice, and ethics.

Jaemin Cho

PhD candidate @ UNC Chapel Hill

PhD @ UNC Chapel Hill. Interested in multimodal machine learning.