Noise in text causes large language models to overestimate social bias
When Noise Fabricates Bias: The Fragility of LLM-as-a-Judge Bias Measurement under Noisy Text
Computation and LanguageMachine Learning
Summary
Large language models (LLMs) are often used to judge social bias in written text. This study found that when the text contains errors like typos or bad punctuation, the models are much more likely to think neutral text is biased, rather than missing bias in clearly biased text. This means that bias measurements from noisy text tend to be exaggerated. The researchers tested different levels of noise and found mild noise causes the most distortion in bias judgments.
What this means in practice
- •For content moderation teams: Improve bias detection tools by accounting for noise-induced exaggeration in social bias measurements made by language models.
- •For natural language processing engineers: Enhance bias assessment systems by adjusting inputs to reduce noise or correcting for bias overestimation caused by surface text errors.
Authors
DongHyun Ryu, Jaehyeok Lee, YeongJun Hwang, JinYeong Bak
Abstract
Large language models are increasingly used as judges to measure social bias in text, yet the passages they judge are often noisy, containing typos, informal spelling, and broken punctuation. The consequences of such surface noise for social bias measurement remain unclear. To investigate this question, we apply five realistic noise conditions at multiple intensity levels to 3,822 stereotype-related responses and compare the resulting bias judgments with those on the original text. We find that such surface noise does not degrade bias measurement symmetrically: it is far more likely to turn neutral judgments into biased ones than biased judgments into neutral ones, by up to a 120x margin. We further observe two non-obvious effects across four LLM judges: in the most fragile judge the distortion is at its purest at mild, realistic noise levels, where erasure is scarcest, and as judges grow robust it attenuates toward parity rather than reversing. Bias measured on noisy text is therefore systematically overestimated, most in the categories that matter most for fairness.