Language models exhibit covert racial prejudice based on dialect, particularly against speakers of African American English (AAE). This prejudice is more negative than any human stereotypes about African Americans ever recorded, and is closely aligned with stereotypes from before the civil rights movement. Language models show more positive overt stereotypes about African Americans, but their covert stereotypes are more negative. This bias can lead to harmful decisions, such as assigning less prestigious jobs, convicting speakers of AAE more often, and sentencing them to death. Existing methods for reducing racial bias, such as human feedback training, do not mitigate dialect prejudice and may even exacerbate the gap between covert and overt stereotypes. The findings highlight the need for further research and interventions to address this form of covert racism in language models. The study demonstrates that language models can perpetuate historical discrimination against African Americans through dialect-based prejudice, which has significant implications for fair and safe employment of language technology. The research also shows that dialect prejudice is not merely a reflection of overt racial bias but a distinct form of bias that remains unaffected by current bias mitigation techniques. The study underscores the importance of addressing covert racism in language models to ensure fair and equitable outcomes.Language models exhibit covert racial prejudice based on dialect, particularly against speakers of African American English (AAE). This prejudice is more negative than any human stereotypes about African Americans ever recorded, and is closely aligned with stereotypes from before the civil rights movement. Language models show more positive overt stereotypes about African Americans, but their covert stereotypes are more negative. This bias can lead to harmful decisions, such as assigning less prestigious jobs, convicting speakers of AAE more often, and sentencing them to death. Existing methods for reducing racial bias, such as human feedback training, do not mitigate dialect prejudice and may even exacerbate the gap between covert and overt stereotypes. The findings highlight the need for further research and interventions to address this form of covert racism in language models. The study demonstrates that language models can perpetuate historical discrimination against African Americans through dialect-based prejudice, which has significant implications for fair and safe employment of language technology. The research also shows that dialect prejudice is not merely a reflection of overt racial bias but a distinct form of bias that remains unaffected by current bias mitigation techniques. The study underscores the importance of addressing covert racism in language models to ensure fair and equitable outcomes.