Date created: | Last Updated:

Category: Project

Description: Natural language contains and communicates a variety of social biases, and this fact increasingly impacts human life, from education to corporate communication and the development of artificial intelligence. Large language models, for example, when naively trained on text data harvested online, demonstrate a variety of unintended negative traits such as stereotypical or derogatory associations with social features like gender, race and disability. The causes of social biases within human language are diverse, including not just the beliefs, prejudices and goals of individuals and social groups, but also historically-ingrained patterns of use for particular words and phrases. Here, however, we provide evidence for a previously-undocumented psychological cause of biased language, in which it arises as a consequence of the automatised mechanisms by which humans retrieve words from memory in order to produce sentences.