Eliciting and Measuring Toxic Bias in Human-to-Machine Interactions in Large Language Models

As Large Language Models (LLMs) continue to advance and be utilized across societal domains, such as finance, healthcare, and the justice system, the need to address the inherent bias and the ability to learn new biases in these models becomes imminent. Concerns regarding these biases are rising, given their potential to perpetuate and even amplify existing social inequalities. This paper explores the multifaceted nature of bias in artificial intelligence (AI), examining the similarities and differences between human and machine bias. We delve into the origins of bias, distinguishing between those introduced by users and those inherent in the AI systems themselves. Our study focuses on the mechanisms by which biases are elicited and amplified through human-to-machine interactions. Through experimentation and analysis, we implement methodologies for eliciting, measuring, and mitigating these biases. Our results suggest that even though LLMs like ChatGPT-4 are equipped with effective content moderators, these chatbots can still learn and exhibit biased responses through human coercion. Further, we have learned that these biases are both inherent and learned through human interaction. Finally, we offer insightful strategies to mitigate these biases in LLMs.