How Pro-Trump Twitter Bots Spread Fake News

President-elect Donald Trump has credited the strength of his political movement, in part, to his immense reach on social-media platforms.

And it’s true, he does have a ton of followers on Facebook and Twitter. But not all of those followers are human. Throughout the 2016 presidential campaign, automated networks of social-media bots spread erroneous information to potential voters—often to the benefit of Trump.

According to a new memo compiling data from the election by a team of researchers including Oxford University Professor Philip Howard, automated pro-Trump activity outnumbered automated pro-Hillary Clinton activity by a 5:1 ratio by Election Day. And many of those auto-Trumpkins were busy spewing lies and fake news: that Democrats could vote on a different day than Republicans; that Clinton had a stroke during the final week of the election; and that an FBI agent associated with her email investigation was involved in a murder-suicide.

“The use of automated accounts was deliberate and strategic throughout the election, most clearly with pro-Trump campaigners and programmers who carefully adjusted the timing of content production during the debates, strategically colonized pro-Clinton hashtags, and then disabled automated activities after Election Day,” wrote Howard; Bence Kollanyi, a Ph.D. candidate at Corvinus University of Budapest; and Samuel Woolley, a Ph.D. candidate at the University of Washington.

The team of researchers used specific hashtags, including #CrookedHillary, #TrumpTrain, and #ImWithHer, to track the frequency with which messages and links to stories were being spread throughout the cycle. For the most recent memo, they focused particularly on the final week leading up to the election (Nov. 1-9).

Of the approximately 18.9 million tweets generated during the last week of the campaign that used some combination of the hashtags in the study, 55.1 percent were pro-Trump and 19.1 percent were pro-Clinton.

Many were automated, which the researchers defined as follows:

“These accounts are often bots that see occasional human curation, or they are actively maintained by people who employ scheduling algorithms and other applications for automating social-media communication. We define a high level of automation as accounts that post at least 50 times a day using one of these election-related hashtags, meaning 450 or more tweets on at least one of these hashtags during the data collection period.”

And as soon as the election ended, the researchers discovered a steep decline in content. Some nodes in networks began to disappear; others simply retweeted what Trump said.

“The pace of automated political campaigning dropped off after Election Day—a reminder that ultimately the campaigners and programmers behind such accounts are humans who could disable their automation on victory,” the memo read.

During the period after the third and final presidential debate, when most public polling indicated that Clinton had a steady lead, the hashtags generated by the bot accounts pointed to a bit of a different story.

“In an important way, Twitter activity reflects the shift in popular sentiment that was largely undetected by traditional polling,” the memo read. “In the last debate, 30.8 percent of the traffic about the debates was using relatively neutral hashtags, but was cut in half by Election Day. In the leadup to voting, only 15.2 percent of the traffic was using neutral hashtags, pro-Trump traffic grew from 46.7 to 55.1 percent, and pro-Clinton traffic grew from 10.4 to 19.1 percent.”

In a conversation with The Daily Beast, Howard described how the pro-Trump bot networks began to use pro-Clinton hashtags, injecting memes, links, and political messages into pro-Clinton circles. Like a virus, they essentially co-opted the opponent’s messaging and infiltrated her supporters. Using pro-Clinton hashtags like #ImWithHer and #uniteblue, memes describing Clinton as corrupt ricocheted across both blue and red feeds.

“Programmers made some strategic points about how to use the automation to maximum effect,” Howard said.

Nowhere was this more obvious than in the final week of the campaign. For example, Howard explained that the bots aggressively pushed the news of FBI Director James Comey looking into additional emails pertaining to Clinton’s investigation on the computer of her aide’s estranged husband, Anthony Weiner. Days later, when Clinton once again was not charged, the networks were only sharing the news of the investigation and not the absolvement.

Besides selective story-sharing, the networks pushed some outright false reports.

“There was a story from the ‘Denver Guardian’ about the FBI agent who was suspected in the Hillary email leaks who was found dead in a murder-suicide,” Howard said of one example.

He’s referring to a story from a nonexistent publication that spread so widely that The Denver Post, a real publication, had to write a thorough debunking of the report. As The New York Times reported, the fake story “claimed that an FBI agent connected to Hillary Clinton’s email disclosures had murdered his wife and shot himself.”

Howard said Trump’s simple turns of phrase, like the infamous “Crooked Hillary” line, made for naturally great hashtags that could be exploited by these accounts. In the final days leading up to the election, they even went so far as to employ pithy attempts at voter suppression.

“We tracked a fair amount of bot activity on Sunday [Nov. 6] on voter-suppression messaging,” Howard said. “There was a note that came over the Clinton hashtags that said voting would be postponed until Thursday.” Another claimed that “Democrat voters could SMS their vote in and there was another one that Clinton had a stroke over the weekend.”

All of these were means to trick unsuspecting users into not turning out on Election Day to vote for Clinton.

Howard provided numerous examples of pro-Trump accounts incorporating pro-Clinton hashtags as a means of spreading the message to a wider audience.

There were some accounts that were even personalized with human-seeming names, spewing out the same tweet verbatim. As Howard’s fellow researcher Samuel Woolley described one instance: “Pro-Trump bots that looked like Latino voters were launched after Trump’s primary win there and disappeared the next week.” This seemingly also happened with less frequency with pro-Clinton bots.

Howard described a similar social-media campaign that took place in the leadup to Brexit, a referendum with which Trump tried to associate his surprise win. As Howard and his team analyzed the data, he determined that the pro-Trump accounts were more successful than the pro-Clinton ones due to the negative nature of their messaging and the “offensive pictures with strange captions” that can so easily get passed around.

He, however, did not think that was the work of the campaign itself.

“I’m not sure the Trump campaign had enough of a strategy,” Howard said. “I’m not sure they had the brain or interest to do this. The PACs around him and the individual passionate Trump supporters would have done this.” (The Daily Beast previously reported on one such organization funded by Oculus founder Palmer Luckey.)

When asked whether the proliferation of pro-Trump content affected the electorate, Howard didn’t draw any direct conclusion but said he believed “they had an impact.”

Trump’s social-media activity was always a topic of discussion during the election and in particular instances, including his sharing of an unseemly image of Ted Cruz’s wife, drew sharp criticism from other candidates and the media. At the time, individuals affiliated with the Trump campaign, including Brad Parscale, who went on to assume a bigger role in Trump’s data operation, denied that anyone in the campaign had any involvement in the generation of these memes.

“I don’t have any other Twitter accounts,” Parscale told The Daily Beast at the time. “If there’s something else the campaign is doing, I don’t know about it.”

The persistence of the accounts during the election, according to Samuel Woolley, were meant to give the appearance that Trump may have had a bigger initial following than expected.

“Some of the botnets that supported Trump were more than likely purpose-built to create an illusion of massive online political traction for Trump,” Woolley said in an email. “These sorts of bots work to create a bandwagon effect among voters who are considering a candidate, or are focused on a specific issue. They also generate a spiral of silence among voters who might not agree with a candidate or issue but who experience a barrage of hugely enhanced content from the Trump bot network. These purpose-build bots and botnets often disappear right after a political campaign, some are even created for a specific issue within a campaign and go offline after working to manipulate public opinion around that one issue.”

And now, having done their duty, the remaining active networks are amplifying the president-elect’s message even louder with tens of thousands of retweets at a time.