Think back to September 29, 2016. If someone had told you that, one year later, Twitter would be briefing the Senate and House intelligence committees on how Russia used a network of hundreds of automated bots to influence the U.S. election, would you have believed them?
Emilio Ferrara probably would have. Ferrara has studied bots for years, today as research leader of the Machine Intelligence and Data Science group at University of Southern California. His research has proven amazingly prescient. Just a day before the election, he published a study covered here on Co.Design showing that 20% of all political tweets during the televised presidential debates between Hillary Clinton and Donald Trump were made by bots. Today, automated Twitter users are still influencing the conversation around everything from the healthcare debate to the NFL’s #TakeAKnee controversy.
To find out, they began a long process of designing and building their own bot network, as they explain in a study published in PLOS One this week. Their “army” consisted of 39 Twitter bots, each with its own convincingly human profile, all located in the Bay Area. With a specific set of instructions, the bot net worked together to gain a whopping 25,000 followers, many of which followed multiple bots. Their mission? Spread a series of positive messages, or “memes for social good,” through tweets using a list of hashtags. For instance, there were health-focused messages like #GetYourFluShot and #HowManyPushups. Other sentiments were simply positive, like #SomethingGood and #HighFiveAStranger. Then, the bot net would systematically retweet and favorite other tweets that used the hashtags,
If those hashtags sound almost implausibly cheerful to you, you aren’t alone. Ferrara recalls that they weren’t sure if Twitter would care about their glorified PSAs. They were wrong. “It was surprising how successful some of these bots became in a very short time,” he says. In just two or three months over 2014, some of the bots garnered thousands of followers and carried out their tasks amazingly well, retweeting, favoriting, and following other Twitter users who engaged with their topics. The bots were actually pretty good at Twitter, even though they weren’t spreading negative messages.
I high-fived a stranger today and it was awesome! #highfiveastranger http://t.co/AhlkAg34x0
— James Met (@meetjamesmet) November 18, 2014
The purpose of the experiment wasn’t just to see if bots could spread positivity, though. It also concerned a complex–and increasingly crucial–topic: how information spreads on Twitter. While it’s pretty easy to model the way a virus travels in the real world, modeling the spread of online information is difficult. Do memes travel like viruses, which operate on a “simple contagion” model where each exposure to the “contagion” can independently “infect” a person? Or is it more complicated than that?
“[If] simple bots can be effective to that extent, then what happens if you have a very sophisticated bot that’s fueled by complex, advanced artificially intelligent algorithms?” he asks. These more advanced bots would be almost impossible to identify as artificial, and they could operate at speeds that we humans can’t match. “And they can actually overwhelmingly control the conversation if they’re used for negative manipulative purposes.”