Beyond Buzzwords: Practical Approaches to Mitigating Biases in the Age of LLMs

Overview

Prior work in human computation and crowdsourcing has shown how human annotators fall prey to their own opinions and implicit biases while completing annotation tasks, leading to systematic biases creeping into the resulting data collections and trained ML models. Checklists have been proposed for either combating or reporting potential cognitive biases that may have emerged during annotation processes. Recent research has also revealed discriminatory stances and stereotypical biases present in LLMs. This webinar will share insights on how we can collect bias-aware annotations from humans and discuss approaches towards building robust Human-Centered AI systems by leveraging LLMs.

Download our presentation