Tracing Bias for Fairer Content-Based Misinformation Detection

Abstract

Despite the benefits attributed to AI systems, their deployment across domains still present challenges to society. In the case of automated misinformation detection, research has uncovered that benefits derived from their application are unequally distributed amongst different stakeholders, calling to attention the need to audit these AI systems for biases and other sources of harm. We present a hybrid AI system designed to trace biases from input data, enriched with semantic descriptions. Using boxology design patterns, we illustrate the integration of a semantic model with an AI system to enable bias tracing. In our case study, we assess fine-tuned language models for content-based misinformation detection, and adapt existing bias detection and mitigation techniques to transform data based on demographic signifiers and measure model fairness. Our findings show evidence that, on average, the evaluated datasets demonstrate a stark gender and geographical biases. Further, we observe that models trained on demographically transformed data demonstrate higher fairness. These results underscore the importance of curated and diverse data and of managing biases plaguing language models at task level.