An Evaluation of the Google Perspective API by Race and Gender

Research on American politicians demonstrates that minority politicians often face higher rates of uncivil speech online. While previous studies have focused on federal representatives, local politicians engage more frequently with constituents, increasing the risk of online and offline violence. These imminent threats to state-level officials highlight the need for tools that can measure the toxic discourse faced by local politicians, making algorithmic evaluation of these tools essential for understanding these interactions. To this end, this study evaluates the Google Perspective API, a widely used machine learning algorithm that quantifies toxicity. This study assesses the performance of the API across various demographic identities. Using a dataset of one million tweets directed at state legislators across all 50 states in January 2021, I identify significant gender and racial discrepancies in the algorithm’s performance. Specifically, the API demonstrates better performance in predicting toxicity toward men than toward women. The racial discrepancies are slightly more nuanced with the API performing better for some races and not others. This research underscores the importance of algorithmic validation and has implications for studies of algorithmic performance, online harassment and political communication.