The article explores socio-psychological harms posed by language models, particularly focusing on omission, subordination, and stereotyping based on gender, sexual orientation, and race. By employing a fine-tuned GPT-3.5-turbo model, the authors automate the extraction of gender references and names from a dataset of 500,000 generated stories. Their methodology aims to quantify and understand these harms, ultimately contributing to discussions around ethical AI usage and its implications for marginalized identities.
In analyzing the impacts of generative language models, we focus on three primary forms of socio-psychological harm: omission, subordination, and stereotyping, examining how these harms manifest across diverse identities.
Our methodology utilizes a fine-tuned GPT-3.5-turbo model for high-precision extraction of gender and racial indicators from a substantial dataset of synthetic stories generated by various language models.
Collection
[
|
...
]