StereoTales: Multilingual Open-Ended Stereotype Discovery in LLMs (research.giskard.ai)

Researchers have unveiled StereoTales, a groundbreaking multilingual dataset and evaluation framework aimed at uncovering social biases in open-ended story generation by large language models (LLMs). By analyzing over 650,000 stories created by 23 prominent LLMs across ten languages, the study identified more than 1,500 socio-demographic associations, revealing a troubling prevalence of harmful stereotypes. Notably, every evaluated model exhibited systemic issues regardless of size or provider, shedding light on the pervasive nature of bias in AI-generated content. The research highlights important gaps in the current understanding of bias, emphasizing that stereotypes are not only present but also culturally specific, adapting to the linguistic context of the prompt. Additionally, while models and human assessments typically align on identifying harmful associations, the study found LLMs often underestimate harm related to socio-economic status. This new framework challenges existing bias detection methodologies, moving beyond template-based tasks and reinforcing the necessity for holistic bias evaluation across different languages and contexts. The released resources include a comprehensive dataset on Hugging Face and a GitHub repository for further exploration, aiming to empower the AI/ML community to better address the challenges of bias in language models.