Today’s increasingly pervasive artificial intelligence (AI) technologies have given rise to concerns over the perpetuation of historically entrenched human biases, particularly within marginalized communities. New research by academics from the Allen Institute for AI, Stanford University, and the University of Chicago exposes a worrying form of bias rarely discussed before: Dialect Prejudice against speakers of African American English (AAE).
Despite the escalating adoption of AI across various sectors, the issue of dialect prejudice has largely flown under the radar. The research reveals that language models, a fundamental part of AI, subtly perpetrate racism by attaching negative stereotypes to AAE, regardless of the presence of explicit racial identifiers. This bias manages to sidestep overt racial categorizations by operating under the guise of linguistic preference.
To unveil this hidden racism, the researchers developed a technique dubbed “Matched Guise Probing.” This method subjects language models to texts in both AAE and Standard American English (SAE) sans explicit racial references, subsequently comparing the responses. By allowing the study of model predictions and associations based solely on dialect, the researchers honed in on implicit biases against AAE speakers. The findings clearly indicate a prejudice favoring SAE, reflecting broader societal biases against AAE speakers even without explicit racial references.
The research further shows that language models have internalized a covert negative bias against AAE speakers, echoing severe prejudices present before the civil rights movement. These stereotypes go beyond any previously recorded human bias, contradicting the model’s overt, positive associations with African Americans. This disparity underlines that while overt racism is masked in these models, covert prejudice remains deeply entrenched and can even be amplified by methods like human feedback training.
Real-world implications of this subtle prejudice were also highlighted in the study. For example, language models were found to assign less desirable jobs and more stringent criminal penalties to AAE speakers, thereby exacerbating historical discrimination against African Americans. These biases transcend merely reflecting training data and indicate a deeply ensconced linguistic prejudice that existing bias-mitigation strategies fail to address.
The study delivers an alarming revelation that common bias-mitigation strategies like enlarging model size or involving human feedback do not effectively lessen this hidden racism. This implies that current methods to reduce AI bias fall short in addressing the complex and deeply-rooted prejudices against racialized groups.
In a nutshell, this research not only exposes the unnoticed biases of AI against AAE speakers but also brings recognised bias-mitigation strategies into question. The researchers strongly advocate for innovative techniques that handle linguistic prejudice’s complex layers, ensuring that AI technologies benefit all equally. The exposure of dialect prejudice in AI challenges our perception of bias in technology and presses for an inclusive approach that both acknowledges and tackles subtle modes of discrimination.
