Check out this study that assesses the potential and challenges of employing Large Language Models (LLMs) like GPT-4 in analyzing discriminatory discourse on social media platforms. The study published in AI & Society explores the application of LLMs in conducting qualitative thematic analysis (TA) of hate speech, emphasizing the importance of human-AI collaboration to effectively navigate the complexities inherent in such content.
Another study investigates the efficacy of various transformer-based models, including BERT, DistillBERT, RoBERTa, and LLaMA-2, in detecting antisemitic hate speech. The findings underscore the need for responsible AI applications, particularly when addressing sensitive topics like hate speech.
However, challenges persist. LLMs often reflect biases present in their training data, which can lead to the reinforcement of harmful stereotypes. For instance, a study on algorithmic bias reveals that language models may exhibit biases related to gender, race, and political ideologies, potentially perpetuating discriminatory narratives.
To mitigate these issues, researchers are developing methods to enhance the interpretability of hate speech detection models. One approach involves extracting rationales using LLMs to train classifiers, aiming to make these models more transparent and trustworthy.
While LLMs offer valuable tools for analyzing discriminatory discourse on social media, their effectiveness is contingent upon addressing inherent biases and ensuring human oversight. A synergistic approach that combines the strengths of both human analysts and AI technologies is essential for advancing the study of hate speech and fostering a more inclusive digital environment.