Measuring and Mitigating Bias in AI-Chatbots

The use of artificial intelligence (AI) to train conversational chatbots in the nuances of human interactions raises the concern of whether chatbots will demonstrate prejudice similar to that of humans, and thus require bias training. Ideally, a chatbot is void of racism, sexism, or any other offensive speech, however several well-known public instances indicate otherwise (e.g., Microsoft Taybot). In this paper, we explore the mechanisms of how open source conversational chatbots can learn bias, and we investigate potential solutions to mitigate this learned bias. To this end, we developed the Chatbot Bias Assessment Framework to measure bias in conversational chatbots, and then we devised an approach based on counter-stereotypic imagining to reduce this bias. This approach is non-intrusive to the chatbot, since it does not require altering any AI code or deleting any data from the original training dataset.