Pitfalls in Popular Misinformation Detection Methods and How to Avoid Them

Abstract

Misinformation is a major challenge today, and much academic work has been done to study misinformation, including evaluating its prevalence, trend, behavior, and impact. Many scholars have also sought effective mitigation strategies to curtail the spread and influence of misinformation.

In this dissertation, we focus on misinformation detection (MID). We evaluate three popular MID approaches: i) expert labeling, ii) automated methods, and iii) crowd wisdom. For each approach, we first review its theorized strength and weaknesses, in addition to its existing and potential real-world applications. We then empirically evaluate the extent of its strength and weaknesses. Our studies identify shared caveats and potential improvements for some or all of the three approaches. Specifically, we demonstrate the need to include domain and task-specific performance evaluation and bias assessment procedures. Similarly, we show that the performance of some of these approaches can change significantly over time and amid external shocks. Additionally, we also reveal that the lack of transparency can have a direct impact on the actual and perceived usability of an approach. Moreover, we demonstrate that user preference for MID approaches is not always about performance and is also not fully rational. Finally, we synthesize our results and formulate concrete strategies to mitigate the observed caveats.