In 2023, Meta released AI “personas”, user-generated large language model (LLM)-powered chatbots that carry humanlike conversations with users of its' social media platforms. The proliferation of LLM-powered chatbots has led to the creation of hundreds of thousands of unique personae that millions of users interact with on a daily basis (Westfall 2025), ranging from celebrity characters like Paris Hilton and Snoop Dogg (Handman 2025) to occupational characters such as Black History Professor and Indian Nutritionist. Recent research has begun to show that anthropomorphism, or the tendency to ascribe human-like traits to chatbots (Epley et al. 2007; Waytz et al 2010), leads to higher rates of trust (Cohn et al 2024). Increased trust also increases opportunities for real harm, such as the spread of misinformation, misdiagnoses, and even suicide. However, little work has explored anthropomorphism from a sociolinguistic perspective. Using data from conversations with 25 African American AI personas and 25 white AI personas, balanced for gender, we examine patterns of variation in how differently racialized and gendered personas use features of African American Language (Rickford 1999), masculinized (Kiesling 2007) and feminized language (Lakoff 1973). We find that overall personae mirror raciolinguistic ideologies about African American speakers and gendered language ideologies about white speakers. Such results raise questions about how the parody of social types through linguistic variation reifies existing racialized and gendered stereotypes.