Core Concepts
Language models exhibit covert racism through dialect prejudice, impacting their decisions about individuals based on how they speak.
Abstract
Language models perpetuate covert racism by exhibiting dialect prejudice against African American English speakers. This bias influences their decisions regarding employability and criminality, reflecting societal racial stereotypes. Despite efforts to mitigate racial bias in language models, dialect prejudice remains a significant concern with far-reaching implications for fairness and safety in technology applications.
Hundreds of millions of people interact with language models that perpetuate systematic racial prejudices. Research reveals covert racism manifested as dialect prejudice in language models, impacting decisions about job assignments, criminal convictions, and sentencing. Existing methods to alleviate racial bias do not address dialect prejudice effectively, highlighting the need for further research and solutions.
Key points include:
Language models embody covert racism through dialect prejudice.
They associate negative stereotypes with African American English speakers.
Covert stereotypes influence decisions about job assignments and legal outcomes.
Larger language models show more covert but less overt prejudice.
Human feedback training improves overt stereotypes but has no clear effect on covert biases.
Stats
Language models exhibit archaic stereotypes about African Americans from before the civil rights movement.
The association with African American English predicts occupational prestige in language models.
Convictions are more likely for AAE speakers compared to SAE speakers across all language models.
Quotes
"Language models embody covert racism in the form of dialect prejudice."
"Our findings have far-reaching implications for the fair and safe employment of language technology."