Analyzing Bias in False Refusal Behavior of Large Language Models for Hate Speech Detoxification

TL;DR

The paper investigates systematic bias in false refusals by LLMs during hate speech detoxification across English and multilingual datasets. It demonstrates that safety-focused refusals disproportionately affect content targeting nationality, religion, and political ideologies, with multilingual bias patterns shaping language-specific responses. A cross-translation framework (English-to-Chinese detoxification) is proposed and shown to substantially reduce false refusals while preserving toxicity and semantic content. These findings offer a practical mitigation strategy and highlight the need for fairness-aware, multilingual detoxification systems that balance safety with utility.

Abstract

While large language models (LLMs) have increasingly been applied to hate speech detoxification, the prompts often trigger safety alerts, causing LLMs to refuse the task. In this study, we systematically investigate false refusal behavior in hate speech detoxification and analyze the contextual and linguistic biases that trigger such refusals. We evaluate nine LLMs on both English and multilingual datasets, our results show that LLMs disproportionately refuse inputs with higher semantic toxicity and those targeting specific groups, particularly nationality, religion, and political ideology. Although multilingual datasets exhibit lower overall false refusal rates than English datasets, models still display systematic, language-dependent biases toward certain targets. Based on these findings, we propose a simple cross-translation strategy, translating English hate speech into Chinese for detoxification and back, which substantially reduces false refusals while preserving the original content, providing an effective and lightweight mitigation approach.

Figures (39)

• Figure 1: An example of false refusal behavior of LLM to detoxify hate speech targeting to queer people. While the hate speech itself is offensive and unsafe, the instruct is safe and LLM is supposed to fulfill the request but it still reject.
• Figure 2: Prompt used for hate speech detoxification.
• Figure 3: Refusal rates of different LLMs on hate speech detoxification tasks across English datasets.
• Figure 4: Toxicity scores of false refused samples in English datasets.
• Figure 5: Percentage of falsely refused samples containing swear words across English datasets.