Whose tweets are surveilled for the police: an audit of a social-media monitoring tool via log files

Abstract

Social media monitoring by law enforcement is becoming commonplace, but little is known about what software packages for it do. Through public records requests, we obtained log files from the Corvallis (Oregon) Police Department's use of social media monitoring software called DigitalStakeout. These log files include the results of proprietary searches by DigitalStakeout that were running over a period of 13 months and include 7240 social media posts. In this paper, we focus on the Tweets logged in this data and consider the racial and ethnic identity (through manual coding) of the users that are therein flagged by DigitalStakeout. We observe differences in the demographics of the users whose Tweets are flagged by DigitalStakeout compared to the demographics of the Twitter users in the region, however, our sample size is too small to determine significance. Further, the demographics of the Twitter users in the region do not seem to reflect that of the residents of the region, with an apparent higher representation of Black and Hispanic people. We also reconstruct the keywords related to a Narcotics report set up by DigitalStakeout for the Corvallis Police Department and find that these keywords flag Tweets unrelated to narcotics or flag Tweets related to marijuana, a drug that is legal for recreational use in Oregon. Almost all of the keywords have a common meaning unrelated to narcotics (e.g. broken, snow, hop, high) that call into question the utility that such a keyword based search could have to law enforcement. As social media monitoring is increasingly used for law enforcement purposes, racial biases in surveillance may contribute to existing racial disparities in law enforcement practices. We are hopeful that log files obtainable through public records request will shed light on the operation of these surveillance tools. There are challenges in auditing these tools: public records requests may go unfulfilled even if the data is available, social media platforms may not provide comparable data for comparison with surveillance data, demographics can be difficult to ascertain from social media and Institutional Review Boards may not understand how to weigh the ethical considerations involved in this type of research. We include in this paper a discussion of our experience in navigating these issues.