A much-discussed research paper out of Oxford this month concluded that millions of tweets about the presidential election are generated by highly automated Twitter accounts. According to the authors’ analysis, about a third of pro-Trump traffic, and one fifth of pro-Clinton tweets, is “driven by bots and highly automated accounts.”
The Oxford study pegged Twitter accounts as highly automated if they posted at least 50 times a day using any one of a group of election hashtags—such as #MAGA, #TrumpTrain, #ImWithHer, and #StrongerTogether—over a three-day period.
The paper conceded that “extremely active” humans might post 50 or more times per day on one of the 52 hashtags they selected, “especially if they are simply retweeting the content they find in their social media feed.”
At the Electome, a project of the Media Lab at MIT, we use complex machine learning algorithms to analyze the election conversation on Twitter. The Oxford paper made us curious about the possibility of spotting bots in the dashboard we recently built for journalists covering the election.
Read more: How Mexican Twitter Bots Shut Down Dissent
Bot detection can be challenging, partly because they come in different varieties. Some are purely automated accounts, while others layer some manual curation on top of automated tweets.
Last week, we noticed a spike while searching our Twitter data on the keyword “rigged.”
In early September, the “rigged” discussion on Twitter, which previously had revolved around a variety of issues including economic inequality and the electoral process, shifted suddenly toward immigration—that is, tweets containing the word “rigged” also used terms connected to immigration.
Digging into the data, we found one verbatim tweet showing up across a dozen or so handles, each of which posted the same message over and over each day: “Immigration Policy is RIGGED against American Workers #Trump2016 #FeelTheBern.”
Beyond using identical phrasing—including idiosyncratic capitalization—the tweets coming from these accounts all linked to the same video, which compares statements by Donald Trump and Bernie Sanders about immigration policy. Each video, in turn, linked to the same anti-Clinton Twitter account.
Although the accounts don’t have the telltale bot profile image—the egg—based on their characteristics and activity, including breakneck output of strikingly similar content, these are clearly spam handles, and apparently at least somewhat automated.
Wading in further, we found that each account puts out a stream of photos and GIFs on a given theme, on top of a common rotation of anti-Clinton videos and memes.
The bots follow the same playbook: Publicly they tweet the same innocuous content fitting their theme, while simultaneously flooding the replies of public figures and media outlets—essentially piggybacking on famous tweets to influence users who see those tweets' replies—with campaign-driven videos and memes.
One apparent bot account has pumped out more than 27,000 tweets since its creation in March, with content that tends to mix videos of Clinton advisor John Podesta with memes from the 1970s film A Clockwork Orange:
A zombie-themed account boasts 30,000 tweets since April: Podesta mingled with the undead:
Then there’s the seeming food porn handle that has put out 21,000 tweets since March: Podesta plus photogenic snacks:
In the last few days, these three accounts have tweeted thousands of times, sometimes hundreds of posts in a single hour. Most went entirely dark on October 30, for some reason, then geared up early on October 31 to put out hundreds more by noon.
Other apparently automated accounts pay homage to burgers, the Doge meme, geese, Hydrox cookies, knights, pigs, pulp science fiction, Putin, trains, and Transformers. They vary in frequency of activity, but each circulates the same videos with identical accompanying text.
Spambots like these have been spotted at other points in this election. In April, a conservative activist noticed a few hundred accounts frantically tweeting an identical call to file federal complaints against Ted Cruz for robocalls.
In June, a reporter for New York magazine mined the feeds of three pro-Trump, alt-right accounts, noting that they consistently replied to Trump’s tweets within mere seconds and with memes attached. Like the accounts we’ve identified here, many of their replies lacked any connection to the subject of Trump’s original tweet.
Last week, one of those three accounts circulated a hoax image of immigration officers arresting Hispanic voters, according to ProPublica’s Electionland.
Difficult as it is to track down accounts like these or gauge their prevalence, it’s even harder to discern how they might affect the overall Twitter discussion about the election. Whether or not the Oxford analysis proves accurate, its authors performed a service merely by raising public awareness of election bots.
Shawn Musgrave is a writer for the Electome project at the MIT Media Lab.