AI generates covertly racist decisions about people based on their dialect.

Area of Science:

• Artificial Intelligence
• Sociolinguistics
• Computer Science

Background:

• Language models (LMs) are widely used but known to perpetuate racial biases.
• Previous research focused on overt racism, neglecting subtler forms like covert racism.
• Covert racism, particularly post-civil rights, is a developing area of social science study.

Purpose of the Study:

• To investigate if covert racism, specifically dialect prejudice, manifests in large language models.
• To compare LMs' covert and overt stereotypes regarding African Americans and speakers of African American English (AAE).
• To assess the impact of current bias mitigation strategies on these stereotypes.

Main Methods:

• Analysis of stereotypes embedded within large language models.
• Comparison of LM-generated stereotypes with experimentally recorded human stereotypes.
• Evaluation of bias mitigation techniques like human preference alignment.

Main Results:

• LMs exhibit significant dialect prejudice against speakers of African American English (AAE).
• LM stereotypes for AAE speakers are more negative than recorded human stereotypes.
• Overt stereotypes for African Americans in LMs are paradoxically more positive.
• Bias mitigation methods like human preference alignment amplify the disparity between covert and overt stereotypes.

Conclusions:

• Language models embody covert racism through dialect prejudice, with potentially severe societal consequences.
• Current bias mitigation practices are insufficient and may obscure deeper, systemic racism in LMs.
• Urgent reevaluation of LM development and deployment is needed for fair and safe AI applications.