Continued Bot Infiltration of Trump’s Facebook Pages

Researched and written by C.E. Carey

This past Thursday, Facebook released a white paper which “does not contradict” the U.S. intelligence community’s assessment that its platform was used to carry out influence operations leading up to the 2016 election. Though much has been written about the role of bots and influence campaigns — both foreign and domestic — in the election, it’s unclear what role — if any — such campaigns may play post-election. Have the bots stuck around to continue pushing their messages? Have they disappeared? Have they moved on to other causes and upcoming elections? Have new “bot armies” taken their place?

Our group previously provided evidence of a coordinated social media influence campaign during the 2016 election season marked by sudden, simultaneous shifts in language across multiple platforms, including Facebook. For those analyses, we identified nearly 30,000 accounts posting duplicate content (ex: multi-word comments each identical to at least 4 others) on Donald Trump’s official Facebook page and a fan-created page during the campaign. We considered this pattern of posting behavior to be consistent with that of automated accounts, or bots. When repeating this process using comments from both before and after the election and inauguration (July 14, 2015, to April 9, 2017), we once again identified roughly 30,000 likely bots — 32,003, to be exact.

These updated analyses indicate that many campaign-era bots remain in “maintenance mode” today, continuing to post somewhat regularly and ready to pounce whenever a particular administration action piques their interest. New bot networks have also been deployed in response to specific events since the election. For the time being, bots continue to be actively involved in the social media ecosystem and, without intervention on the part of the platforms, show no signs of letting up anytime soon.

(Throughout this article, we refer to any account which posted at least one 10+ token comment identical to 4+ others in our dataset as a “bot.” However, this method of bot identification is far from perfect; for example, any human-operated accounts copying and pasting “viral” messages would be classified as a bot for our purposes, and more sophisticated automated accounts that generate unique content for each post would not be caught by our somewhat crude classification system.)

The Election

Bots were most active on Trump’s Facebook pages right before the election, with nearly 15,000 active in the 30 days leading up to November 8, 2016. Though bots made up only 3% of active accounts, they posted 14% of all comments during this time.

This pre-election spike in bot activity mirrored an overall increase in posting, even among non-bot accounts. However, the jump in bot activity happened earlier and was more sustained, whereas non-bot activity increased rapidly in the days just prior to the election. Both peaked on November 9th, the day after the election.

Immediately after the election, posting levels dropped across-the-board, for both bot and non-bot accounts. Of the over 25,000 bot accounts active before the election, 44% continued posting after November 9, 2016.

The First 100 Days

During the first 100 days of Donald Trump’s presidency (although technically, only the first 80 days were included in our dataset), 32% of bots active pre-election continued to post — conversely, 63% of bots active post-inauguration were active pre-election. Bots generally remained active at post-election, pre-inauguration levels, posting roughly 12% of comments per day. Non-bot activity, on the other hand, trended downward over time.

Relative daily posting activity tracked similarly across both groups, likely a result of all comments being posted in response to page posts. For example, if a page produces 5 posts one day, and only 1 post another day, the day with 5 posts would present more opportunity for bots and non-bots alike to comment.

However, several notable differences in activity occurred. For example, during the first 2 weeks of the Trump presidency, 3 events provoked divergent responses from bot vs. non-bot accounts:

Non-bot accounts showed a dramatic spike in posting behavior on January 20, 2017 — the day of President Trump’s inauguration. Though bots also increased activity relative to the preceding days, the spike was nowhere near as drastic. Conversely, a spike in bot activity occurred on January 27, 2017 — the day President Trump signed an executive order halting immigration and refugee resettlement from 7 Muslim-majority nations. Whereas non-bot accounts posted substantially more comments on Inauguration Day than on the day of the so-called “Muslim ban,” for bots this pattern was reversed. Finally, a spike in non-bot commenting on January 31, 2017 — the day President Trump nominated (then-) Judge Neil Gorsuch to the US Supreme Court — was not accompanied by as large an increase in bot activity.

These different patterns of posting suggest that bots and non-bots have different “interests:” bots were less interested in the pomp and circumstance of the 58th US Presidential Inaugural Ceremonies and nomination of a traditionally conservative judge to the Supreme Court, while relatively more interested in an executive action affecting foreign relations, than non-bots.

The Syria Strike Spike

Post-election, April 7, 2017 — the date of the Syria strike — was the most active day for bots, with just under 2500 bot accounts posting over 6500 messages. On that day alone, bots made up 12% of active users and nearly one in four comments. Though posting by non-bots increased as well, it was not nearly to the same degree as the increase in bot posts.

Looking into the content of the bot posts, almost all were, predictably, about the Syria strike. Interestingly, more than half of the active bot accounts posted one (or more) of 3 almost identical messages, which in total made up just under half of all bot-posted messages that day:

These blows increase our determination, strength and will, and you are still condemning your great lie by calling it a revolution. Israel has increased its frustration with the strength of the Syrian Arab army, which is fighting on all fronts at home and this border is Syria, with its strength and successful leadership, the people are growing in love with its army and leadership. We resist and resist because we are right owners no matter how you and all the traitors and those who call themselves kings We will not surrender here the Syrian Arab Army. Here are the protectors of the homeland. [Posted 1500+ times]
We Syrian people denounce the attack on Our air base, which has always played a major role in the fight against terrorism. [Posted 1000+ times]
Trump you exceeded your limits in this strike American Zionist Saudi Arabia. We will tell you that the Syrian Arab Army is the myth and master of this time are the protectors of the homeland and its immune system. The reply is coming. Here is Syria Assad. [Posted 350+ times]

Of the bot accounts which posted those messages, only 8 (less than 1%) had posted prior to the election. Only 2% had posted prior to April 7, 2017. In contrast, of the bot accounts which posted other messages, 1 in 5 were active pre-election, and two-thirds were active prior to April 7. This suggests that additional bots may have been deployed to push pro-Assad messages in response to the Syria strikes. It’s worth noting that even when discounting messages from these “new” accounts, the remaining posts — greater than 3500 in all — still represent one of the highest single-day post volumes post-election.

Dueling Botnets

Nearly 1 in 3 bot accounts active during the early days of the Trump administration had not posted prior to the election. Some appear to have been activated immediately post-election, while yet another group of roughly equal size was deployed post-inauguration. Both of these groups have maintained a stable presence since then, with the exception of the day of the Syria strike. As mentioned earlier, the surge in active bot accounts that day likely represented an independent “bot attack,” but for visualization purposes, they are counted as part of the “post-inauguration” bot group.

This observation raises the possibility that separate actors may be controlling independent “bot armies,” each with specific — and perhaps conflicting — interests and intentions. And even as pre-election bots drop off, more may step in to take their place.

What’s Next?

Civic engagement with our elected officials, including the president, as well as other interested citizens is a critical component of democracy and US government. The Internet and social media were supposed to usher in an era of “democratization of knowledge,” wherein everyday citizens — not just elected officials and those with special access — could have unprecedented access to information about the inner workings of government and be able to discuss policy with others across the country — and the world.

Then came the 2016 election, in which Donald Trump surprisingly (at least to most political pundits, pollsters, and legacy media outlets) overtook frontrunner Hillary Clinton in the Electoral College to become the 45th President of the United States. Countless election “postmortems” have focused on the role of “fake newsand social media in influencing public opinion. In these less rosy takes, the “democratization of knowledge” has also increased public access to misinformation, often pushed by “users” who don’t have America’s best interests in mind.

Moreover, the massive prevalence of automated content and content posted by disingenuous actors (see also sockpuppets and trolls) violates an implicit compact users make when they submit their personal information to social media platforms: that they will be engaging with other actual people. When Facebook users visit an official or fan-created page of their president, they expect to receive updates about his actions and interact with other users — positively or negatively, civilly or less-than-civilly — in response. If over 1 in 10 comments on a typical day are the work of automated accounts, however, the opportunity for discourse collapses.

In the months since the election, Facebook has taken multiple steps to combat fake news and police fake accounts. On the fake news front, they’ve begun flagging “disputed” news stories, posted tips on how to spot fake news, and even taken out full-page print ads in newspapers in France and Germany. Regarding fake accounts, they recently shut down 30,000 accounts in France in advance of the presidential election and disrupted an international spam operation. Notably, these Facebook-initiated anti-bot measures took place directly after the period covered in our analyses of Trump’s Facebook pages. It remains to be seen whether they will have any impact site-wide, or whether fake accounts are so embedded in the social media economy that their continued use is all-but-inevitable.

Nonetheless, one thing’s for certain: they’re not going away on their own.