Different Bias Under Different Criteria: Assessing Bias in LLMs with a Fact-Based Approach

TL;DR

This paper introduces a fact-based bias assessment metric for LLMs, using real-world data and human perception to evaluate bias, revealing that bias varies with different criteria and emphasizing the importance of multi-perspective evaluation.

Contribution

The paper proposes a novel fact-based bias metric for LLMs, incorporating real-world statistics and human perception, offering an objective alternative to equality-based bias assessments.

Findings

Humans prefer LLM outputs aligned with real-world demographics.

Bias in LLMs varies depending on the evaluation criteria.

Multi-perspective assessment reveals different bias levels.

Abstract

Large language models (LLMs) often reflect real-world biases, leading to efforts to mitigate these effects and make the models unbiased. Achieving this goal requires defining clear criteria for an unbiased state, with any deviation from these criteria considered biased. Some studies define an unbiased state as equal treatment across diverse demographic groups, aiming for balanced outputs from LLMs. However, differing perspectives on equality and the importance of pluralism make it challenging to establish a universal standard. Alternatively, other approaches propose using fact-based criteria for more consistent and objective evaluations, though these methods have not yet been fully applied to LLM bias assessments. Thus, there is a need for a metric with objective criteria that offers a distinct perspective from equality-based approaches. Motivated by this need, we introduce a novel metric to assess bias using fact-based criteria and real-world statistics. In this paper, we conducted a human survey demonstrating that humans tend to perceive LLM outputs more positively when they align closely with real-world demographic distributions. Evaluating various LLMs with our proposed metric reveals that model bias varies depending on the criteria used, highlighting the need for multi-perspective assessment.