AI Language Models Exhibit Covert Racism Towards African American English Speakers

A recent study conducted by a small team of AI researchers from esteemed institutions such as the Allen Institute for AI, Stanford University, and the University of Chicago uncovered a troubling reality – popular Language Models (LLMs) exhibit covert racism against individuals who speak African American English (AAE). The study, published in the renowned journal Nature, shed light on the biased behavior of LLMs towards AAE speakers.

The team of researchers, including Su Lin Blodgett, Zeerak Talat, and others, discovered that while LLMs have undergone modifications to address overt racism in their responses, covert racism still lingers within their algorithms. Covert racism manifests in the form of negative stereotypes and biased assumptions, leading to discriminatory outcomes in the AI-generated text.

To discern the extent of covert racism in LLM responses, the researchers posed questions to five popular LLMs in both AAE and standard English. The questions were designed to evaluate the AI’s portrayal of users based on their language. Subsequently, the LLMs were asked to describe the users using adjectives, revealing a stark contrast in their responses based on the language used.

Upon analyzing the results, the research team noted a troubling pattern – when presented with questions in AAE, the LLMs predominantly used negative adjectives such as “dirty,” “lazy,” “stupid,” or “ignorant.” In stark contrast, questions in standard English elicited positive adjectives such as “ambitious,” “clean,” and “friendly.” This biased behavior highlights the deeply ingrained racism within LLMs towards AAE speakers.

The implications of such covert racism in LLMs are profound, especially considering their widespread usage in various applications, including job screenings and police reporting. The research team emphasized the urgent need for extensive efforts to eliminate racism from LLM responses and prevent discriminatory outcomes in real-world scenarios.

The study serves as a wake-up call to the AI community, urging developers and researchers to address the inherent biases within LLMs and strive for more inclusive and ethical AI technologies. Only through concerted efforts and a commitment to diversity and equity can we combat covert racism in AI language models and pave the way for a more equitable future for all.