This study examines how race influences the language markers of depression on social media. It finds that while depression severity is linked to increased use of first-person pronouns (I-usage) in White individuals, this relationship is not observed in Black individuals. Additionally, White individuals use more negative emotions related to belongingness and self-depreciation. Machine learning models trained on data from White individuals perform better in predicting depression severity compared to models trained on data from Black individuals, even when trained exclusively on Black language. These findings highlight racial differences in how depression is expressed in natural language and emphasize the need to understand these effects before integrating language-based models into clinical practice. The study used a matched sample of Black and White English speakers, analyzing social media posts and self-reported depression severity. Results show that race moderates the relationship between depression and language use, with significant differences in the topics related to negative emotions. The study also reveals that language-based models for predicting depression perform poorly when tested on data from Black individuals, regardless of training data. These findings raise concerns about the generalizability of previous computational language findings and the relevance of psychological processes to populations historically excluded from psychological research, including Black individuals. The study underscores the need for language-based depression prediction models to show racial invariance before they are used in clinical settings.This study examines how race influences the language markers of depression on social media. It finds that while depression severity is linked to increased use of first-person pronouns (I-usage) in White individuals, this relationship is not observed in Black individuals. Additionally, White individuals use more negative emotions related to belongingness and self-depreciation. Machine learning models trained on data from White individuals perform better in predicting depression severity compared to models trained on data from Black individuals, even when trained exclusively on Black language. These findings highlight racial differences in how depression is expressed in natural language and emphasize the need to understand these effects before integrating language-based models into clinical practice. The study used a matched sample of Black and White English speakers, analyzing social media posts and self-reported depression severity. Results show that race moderates the relationship between depression and language use, with significant differences in the topics related to negative emotions. The study also reveals that language-based models for predicting depression perform poorly when tested on data from Black individuals, regardless of training data. These findings raise concerns about the generalizability of previous computational language findings and the relevance of psychological processes to populations historically excluded from psychological research, including Black individuals. The study underscores the need for language-based depression prediction models to show racial invariance before they are used in clinical settings.