A recent study by the Wharton AI & Analytics Initiative examines whether race and gender influence large language models (LLMs) in hiring and evaluations. Researchers, including Prasanna (Sonny) Tambe, found that biases inherent in online data are mirrored in LLM outputs, which can produce varying evaluations based on demographic descriptors — even reversing traditional biases at times. Traditional auditing methods were inadequate; hence, new techniques like LLM-based correspondence experiments were developed to measure responses to identical candidates across race and gender. The findings revealed persistent yet subtle disparities in ratings, particularly benefiting women and racial minorities slightly over White males. This underscores the necessity for context-specific audits tailored to the application of LLMs, as biases can differ based on task or context. Ultimately, Tambe emphasizes the urgency for organizations to conduct thorough audits of LLMs before integrating them into decision-making processes to ensure ethical AI usage.
Source link
