The Bias and Fairness of Language Models: Reviewing Bias and Fairness in the Context of Language Models

Abstract

Language models have rapidly emerged within the field AI and are able to power a wide variety of downstream applications, with the most prominent example being ChatGPT. These models are often trained on billions of words, enabling the modelling of natural language in all its facets, demonstrating human-like capabilities in relation to language understanding as well as natural language generation (NLG). Language models have however been shown to inhabit and exhibit social biases, similarly to those of people. Also in this regard, are these models able to imitate humans. Despite the growing prevalence of language models and their diffusion into various applications, the understanding of language model biases remains fragmented and complex. Concerns as to the fairness of these models are increasingly being raised, with reference to the potential harmfulness of the inhabited biases. This review delves into the research literature to explore the diverse uses and interpretations of the concepts of bias and fairness in relation to language models. The review reveals that contrasting uses of these concepts are employed in the research literature, and points to a need for more consistent terminology and clearer distinctions between the concepts of bias and fairness. This review advocates for a separation of bias and fairness, with the suggestion that fairness should primarily be associated with equal downstream user performance, while bias should be separated from performance concerns. Moreover, it is proposed that future research on language model bias and fairness should define potential use cases and user groups and acknowledge the task-dependent nature of fairness evaluation. Furthermore, the importance of explicitly articulating desired model behavior and potential harms caused by models is stressed. This thesis advocates for a more nuanced understanding of the concepts of language model bias and fairness and offers key insights for future research and practice.