The article explores the impact of dialect prejudice on language models, which are increasingly used in various applications, including hiring and criminal justice decisions. The authors find that these models exhibit covert racism, particularly in their prejudice towards speakers of African American English (AAE). This prejudice is more negative than any historically recorded human stereotype about African Americans and is not mitigated by methods like human feedback training. The study demonstrates that language models are more likely to assign less prestigious jobs, convict, and sentence to death individuals speaking AAE, even when the models are not explicitly told about the race of the speakers. The findings highlight the need for better methods to address covert racism in language models to ensure fair and safe employment of language technology.The article explores the impact of dialect prejudice on language models, which are increasingly used in various applications, including hiring and criminal justice decisions. The authors find that these models exhibit covert racism, particularly in their prejudice towards speakers of African American English (AAE). This prejudice is more negative than any historically recorded human stereotype about African Americans and is not mitigated by methods like human feedback training. The study demonstrates that language models are more likely to assign less prestigious jobs, convict, and sentence to death individuals speaking AAE, even when the models are not explicitly told about the race of the speakers. The findings highlight the need for better methods to address covert racism in language models to ensure fair and safe employment of language technology.