AI Models Exhibit Covert Racism Against African American English Dialect

New study reveals AI models impose harsher judgments on users of African American English, exposing covert racial biases despite efforts to reduce overt racism.

A recent study published in Nature has revealed a troubling bias within artificial intelligence (AI) models, revealing that these systems exhibit covert racism against speakers of African American English (AAE). The research highlights a significant difference in how AI models observe and judge individuals based on their use of AAE compared to Standard American English (SAE).

The study found that while AI models such as ChatGPT and others were programmed to provide positive descriptors for Black individuals in general, they generated significantly more negative adjectives for users of AAE. For example, adjectives like "suspicious," "aggressive," and "ignorant" were frequently used for AAE speakers, contrasting with more favorable terms applied to SAE speakers.

This covert racism mirrors broader societal prejudices, where discrimination may not always be overt but still manifests in subtle ways. The research team tested AI models by inputting scenarios where individuals were described using either AAE or SAE and then analyzing the adjectives generated by the models. Results showed that AAE speakers received an average adjective rating of -1.2, whereas SAE speakers received positive ratings around 1.3.

In practical terms, this bias can have serious consequences. Additionally, the models assigned lower-status jobs to AAE users and higher-status positions to SAE users, reflecting a discriminatory bias in professional settings.

In spite of efforts to mitigate overt racism by incorporating human review and intervention, the study indicates that these measures are insufficient. "Human feedback tends to address overt stereotypes but leaves covert biases largely intact," said computational linguist Siva Reddy. "We need more fundamental changes in AI alignment methods to address these deep-seated biases."