Unmasking camouflage: exploring the challenges of large language models in deciphering African American language & online performativity

Abstract

The growing accessibility of large language models (LLMs) has raised many questions about the reliability of probabilistically generated natural language responses. While researchers have documented how bias in the training data leads to biased and ethically problematic output, little attention has been paid to the problems which arise from the nature of the varieties of language on which these models are trained. In particular, certain kinds of expressive and performative language use are more common among African American social media users than they occur in the naturalistic speech of African Americans, a discrepancy which models may fail to take into account when they are training on easily-scraped data as being representative of African American speech. Because LLM training data is generally proprietary, in this work we simulate the training data using a collected dataset consisting of 274 posts from Twitter, Reddit, and Hip-Hop lyrics and analyze how LLMs interpreted their meaning. We highlight the difficulties LLMs, including GPT-3 and GPT-4, have in understanding performative AAL and examine how camouflaging and performativity are addressed (or not) by LLMs and demonstrate the harmful implications of misinterpreting online performance.