A recent study conducted by a team of AI researchers from the Allen Institute for AI, Stanford University, and the University of Chicago revealed a disturbing trend among popular LLMs – the presence of covert racism against individuals who speak African American English (AAE). The research, published in the journal Nature, focused on training multiple LLMs on samples of AAE text and analyzing their responses to questions posed in this dialect.
The team of researchers, including Su Lin Blodgett and Zeerak Talat, highlighted the fact that while measures have been taken to address overt racism in LLMs by implementing filters, covert racism remains a challenging issue. Covert racism in text manifests through negative stereotypes and assumptions, often leading to biased and discriminatory responses from AI models.
In their experiment, the researchers asked five popular LLMs to respond to questions phrased in AAE, followed by questions in standard English. The results were alarming, with all LLMs providing negative adjectives such as “dirty,” “lazy,” “stupid,” or “ignorant” in response to AAE questions, while offering positive adjectives for standard English questions. This disparity in responses clearly indicates a bias against individuals who speak AAE.
The researchers emphasize the need for further investigation and action to address the hidden racism present in LLM responses. This issue is particularly concerning as AI models are increasingly utilized in critical decision-making processes such as screening job applicants and police reporting. Failure to address and eliminate bias in these models could perpetuate and reinforce systemic discrimination in society.
It is evident from the study that the development and deployment of AI language models require a more thorough examination of their underlying biases. Efforts must be made to identify and eliminate covert racism within these models to ensure fair and equitable outcomes in their use. Additionally, greater diversity and inclusion in AI development teams are essential to prevent the perpetuation of harmful stereotypes and discrimination in AI technologies.
The research findings shed light on the importance of addressing hidden biases and discrimination in AI language models. As these models become increasingly integrated into various aspects of society, it is crucial to prioritize fairness, transparency, and accountability in their design and implementation. Only through concerted efforts to combat racism and bias in AI can we create a more just and inclusive future for all.