Florian Gallwitz and Michael Kreil published a fascinating analysis of research about “social bots” that suggests much of the research about the extent and influence of bots may not be reliable.
The idea that social media platforms like Twitter are inhabited by vast numbers of “social bots” has become widely accepted in recent years. “Social bots” are assumed to be automated social media accounts operated by malicious actors with the goal of manipulating public opinion. They are credited with the ability to produce content autonomously and to interact with human users. “Social bot” activity has been reported in many different political contexts, including Donald Trump’s election and the Brexit referendum in 2016. However, the relevant publications either use crude and questionable heuristics to discriminate between supposed “social bots” and humans or—in the vast majority of the cases—fully rely on the output of automatic bot detection tools, most commonly Botometer. We point out fundamental theoretical flaws of these approaches. Also, we closely and systematically inspected hundreds of accounts that had been counted or even presented as “social bots” in peer-reviewed studies. We were unable to find a single “social bot”. Instead, we found mostly accounts undoubtedly operated by human users, the vast majority of them using Twitter in an inconspicious and unremarkable fashion without the slightest traces of automation. We conclude that studies claiming to investigate the prevalence or influence of “social bots” have, in reality, just investigated false positives and artifacts of the flawed detection methods employed.
Reading the paper, it was genuinely surprising how researchers relied upon incredibly crude to label an account a “bot.”
For example, Gallwitz and Kreil claim that several studies of alleged “social bot” manipulation during the 2016 election in the United States and the 2018 Brexit vote used an absurd heuristic. For example, if an account tweeted more than 50 times a day, it was automatically classified as a bot.
According to Gallwitz and Kreil,
The question of whether an account was considered a “bot” or not as a basis for the quantitative analysis in each of the 5 papers was based on its tweeting rate alone, without any further analysis of the accounts.
. . .
Although the authors concede that “extremely heavy human users might achieve this pace of social activity”, they proceed to refer to each of the accounts that satisfy this criterion as “bots” or “heavy automation”.
. . .
Requiring the 50 tweets per day to include specific political hashtags, as in the U.S.-related studies, will at least help to exclude normal Twitter interactions without any political background. However, as we will argue in the following, it does not make the criterion more suitable for finding “highly automated” accounts.
The fundamental problem of this approach has been described in previous research: It is not uncommon for political activists on Twitter to manually tweet hundreds of times per day, and those tweets will typically include political hashtags.
Frankly, this sounds like researchers studying Twitter have no clue about how people use it. Fifty tweets in a day would be fairly unremarkable, especially if researchers don’t filter out retweets and quote tweets.
When it comes to something like Botometer, Gallwitz and Kreil point out that as of 2018, the service misclassified a significant number of U.S. Congress members, media accounts (such as Reuters), and similar groups and individuals as “bots.”
The entire study is definitely worth a close read.