Improving hate speech detection with large language models

Abstract Efforts to curb online hate speech depend on our ability to reliably detect it at scale. Previous studies have highlighted the strong zero-shot classification performance of large language models (LLMs), offering a potential tool to efficiently identify harmful content. Yet for complex and ambivalent tasks like hate speech detection, pre-trained LLMs can be insufficient and carry systemic biases. Domain-specific models fine-tuned for the given task and empirical context could help address these issues, but, as we demonstrate, the quality of data used for fine-tuning decisively matters. In this study, we fine-tuned GPT-4o-mini using a unique corpus of online comments annotated by diverse groups of coders with varying annotation quality: research assistants, activists, two kinds of crowd workers, and citizen scientists. We find that only annotations from those groups of annotators that are better than zero-shot GPT-4o-mini in recognizing hate speech improve the classification performance of the fine-tuned LLM. Specifically, fine-tuning using the highest-quality annotator group – trained research assistants – boosts classification performance by increasing the model’s precision without notably sacrificing the good recall of zero-shot GPT-4o-mini. In contrast, lower-quality annotations do not improve and may even decrease the ability to identify hate speech. By examining tasks reliant on human judgment and context, we offer insights that go beyond hate speech detection.