Informational Justice in AI-Assisted Fact-Checking

Abstract

Faced with the scale of misinformation, fact-checking organizations are turning to algorithms to efficiently triage claims in need of verification. However, there is uncertainty regarding the appropriate "ground truth" for training and evaluating these algorithms given the varied factors that influence how claims are prioritized for checking. For instance, numerous fact-checking organizations prioritize checking claims that are most likely to impact the "general public," while others prioritize claims that harm vulnerable demographic groups. To better understand the implications of these and other algorithmic design choices, we first extend and then apply the theoretical lens of informational justice to elucidate the often-competing interests of representation, participation, credibility, and the distribution of benefits and burdens among stakeholders affected by algorithms designed to assist with fact-checking. From our examination of an original dataset, we show that different definitions of claim prioritization lead to certain topics being systematically prioritized over others. Moreover, even when using the same definition, we show that data labelers interpret and apply them differently based on their perspectives and demographics. We conclude with a discussion on the theoretical and practical implications of these findings for fact-checking organizations, highlighting the risks of "off-the-shelf" algorithms and opportunities for technical approaches to informational justice.

Keywords: Algorithmic Fairness, Fact-Checking, Misinformation Detection, Algorithmic Bias, Applied Justice