Grace Proebsting, a Haverford College student who graduated in 2025, and Assistant Professor Adam Poliak published “Biases in Large Language Model-Elicited Text: A Case Study in Natural Language Inference” at COLING 2025.
The researchers tested whether datasets created with large language models contain the kinds of annotation artifacts and social biases found in human-generated NLP datasets. Their analysis of text produced by GPT-4, Llama 2, and Mistral found predictable artifacts as well as stereotypical associations involving gender, race, and age. The results underscore the need for careful quality control when LLMs are used to generate research data.