One account features a photo of a middle-aged woman, and a bio that reads “Patriot, self employed, loving mother and grandmother.”
Another has a photo of a younger woman in sunglasses, described in the bio as a “NonProfit Exec born to LEGAL Immigrants who owned laundromat for 30 yrs to earn our #AmericanDream. #PresidentTrump #ProIsrael🇮🇱 #ThankAVet #BackTheBlue #MAGA.”
Both Twitter accounts frequently tweet or retweet in support of US president Donald Trump and in opposition to everything from immigrants, to the National Football League, to CNN. They’ve both had accounts on Twitter since 2012—and they both appear to be bots.
They were identified by a new bot created by Quartz, @probabot_, which searches Twitter for accounts that tweet about politics and scores them using Botometer, a classification tool that applies machine learning to determine how likely a given account is to be a bot.
Twitter has faced increased scrutiny over its ability to limit political influence campaigns on its platform since last month, when executives from the company were called in to a Senate committee hearing about Russian interference in the 2016 election, and presented findings that one committee member said were “inadequate.” The executives provided the committee with little more than 22 Twitter accounts they said were linked to the Russian political advertising campaign that had previously taken place on Facebook. (Twitter on Tuesday said it would begin adding labels to election-related ads and disclose who is behind them.)
In a blog post following the briefings, Twitter provided information about how it handles spam accounts and other bots in general. Its automated bot-detection systems “catch more than 3.2 million suspicious accounts globally per week,” it said, and the company estimates that “false or spam accounts” make up less than 5% of its monthly active users.
However, the total proportion of bots on the platform leaves out some important details, according to Emilio Ferrara, a computer science professor at University of Southern California who contributed to much of the research behind Botometer.
“They reiterate the concept that overall the platform hosts about 5% of bots,” he said in an email. “This is not a very informative figure.”
Bot activity on Twitter is concentrated in conversations about topics like politics, Ferrara said, and it would be better to know what proportion of those conversations include bot interactions. Even if bots and spam accounts make up a small portion of overall accounts, they’re capable of generating far more content than their human counterparts. One of the accounts identified by @probabot_, for example, has tweeted or retweeted 1.14 million times since April 2016–roughly 2,000 times a day.
“We found that approximately 15% of the users active in the political conversation one month prior to the 2016 election were likely bots,” Ferrara said, “and they were responsible to about one in five such political tweets (nearly 20%).”
Ferrara also found that humans retweeted those bot-generated tweets at about the same rate that they did tweets posted by other humans.
So far, @probabot_ has identified accounts that have followings ranging from hundreds of users to thousands, and see varying levels of engagement with other users. Some of the accounts present themselves as news feeds, which explains their bot-like rate of tweeting, while others present themselves as humans, like the accounts mentioned above.
Twitter did not to respond to our request for comment about the prevalence of these kinds of bots on the platform.
Looking more closely at the human-looking accounts, the “self employed, loving mother and grandmother” mostly retweets tweets from the far right, while the “NonProfit Exec born to LEGAL Immigrants” mostly tweets headlines of questionable news articles and randomly thanks other users for follows, without actually using their handles (“Hey WorldBibleInstitute thanks for the follow!”) A simple reverse-image search of that account’s avatar shows that the woman in sunglasses is not, in fact, a nonprofit exec, but is the Australian model Miranda Kerr.
Follow @probabot_ on Twitter, and look out for its tweets as it attempts to separate humans from machines. Each account it identifies will automatically be added to a Twitter list you can find here. As we monitor the accounts the bot finds, we’ll continue to improve the bot, and tweak its parameters to produce the best results.