On 3 September, as German Chancellor Angela Merkel and her main opponent Martin Schulz faced off in an election debate that many viewers panned as more of a duet than a duel, a far livelier effort was underway on social media. People on Twitter started using the hashtag #verräterduell, which translates as “duel of traitors” and mirrors the claim by the right-wing Alternative für Deutschland party that both Merkel’s mainstream Christian Democrats and Schulz’s Social Democrats have “betrayed” the country.
Yet much of the venom may not have been fueled by angry voters, researchers say. Instead it looks like the work of bots, or fake social media profiles that appear to be connected to human users, but are really driven by algorithms.
With Germans going to the polls on 24 September to elect their new parliament, experts are watching closely for signs of automated propaganda on social networks. So far, bots seem less active than they did in the recent presidential elections in France and the United States, where some commentators believe Russia was seeking to boost right-wing candidates. But researchers sensitized by past elections are making the German contest a laboratory for studies of how to recognize bots and trace their effects.
Most researchers concentrate on Twitter, which does not prohibit automated accounts. The platform also makes 1% of tweets freely available through a programming interface—and, for a fee, it opens up 10%. After analyzing tweets from 14 million users worldwide, Emilio Ferrara, a computer scientist at the University of Southern California’s Information Sciences Institute in Los Angeles, estimated that up to 15% of Twitter profiles—a whopping 50 million—are bots. And most are creatures of politics. “Among the few topics that bots focus on,” Ferrara says, “politics is certainly one of the most prominent, if not the most prominent.”
Bots can inflate a topic’s importance or tarnish reputations by flooding social networks with fake news and by manipulating the currency of Twitter: likes and shares, follows and retweets. Just how that translates into votes is unclear, says Simon Hegelich, a political scientist at the Bavarian School of Public Policy at the Technical University of Munich in Germany. Bots are unlikely to change voters’ preferences, he believes, but they might influence decisions on whether to vote at all. “It’s hard to test this in a scientifically rigorous fashion,” he says.
Germany seemed a good place to try. The German parliament’s network was hacked in 2015—Russia is said to be the prime suspect—leading to worries that stolen emails might be published strategically to affect the election. (In France’s presidential election this spring, bots drew attention to stolen, as well as faked, documents.) Last October, Merkel urged political parties to refrain from using social bots; all major parties except Alternative für Deutschland agreed.
Now, research groups are trawling tens of millions of tweets related to the German elections for signs that bots are exerting influence. Lisa-Maria Neudert of the Computational Propaganda Project at the University of Oxford in the United Kingdom is comparing current bot activity to patterns seen during Germany’s presidential elections last February. In that election, in which a political body called the Federal Assembly votes rather than the public, bots accounted for a small fraction of political tweets, Neudert says. She expects more bot activity in the upcoming election, where public opinion is at stake.
Bot-spotting is one of the biggest challenges in the burgeoning field. Neudert’s metric is crude, she acknowledges: She labels any account that posts more than 50 tweets a day using certain political hashtags as a bot. “That’s wrong in both directions,” Hegelich says. Some human users post more, and some bots post far less. But Neudert says that method has been surprisingly good at spotting bots.
Earlier versions of social bots were easy to identify because many posted continuously day and night, but in the arms race between botmakers and bot-detectors they have become harder to identify. (There are signs that botmakers have adapted to Neudert’s rule, staying just below 50 tweets.) “You can never be 100% sure whether a profile is a bot,” Hegelich says. To detect the fingerprints of bots during the Merkel-Schulz debate, scientists in a project called PropStop relied on other measures of behavior. They found that accounts using the #verräterduell hashtag tended to be newer profiles and retweeted existing messages more often than other accounts.
Many researchers are turning to machine-learning techniques to distinguish real and fake users. For instance, Ferrara arrived at his estimate of bots using an algorithm that he trained on millions of tweets from verified human users and bots. It tracks hundreds of features, including an account’s age and use of emoticons. (Bot-generated content tends to be emotionally charged.) Hegelich, who is probing for correlations between voter turnout in the upcoming election and bot activity, examines factors such as the distribution of exclamation marks to pinpoint bots. Humans are inconsistent, he says. “Most bots either use a lot of exclamation marks or never.” But even the most sophisticated models probably miss many bots, Ferrara says. “We do a very good job at detecting simple bots, but for the more complex and advanced ones based on [artificial intelligence] we only have few examples, and we probably miss most of them these days.”
Perhaps the most urgent question is who is behind the bots. Ferrara tracked bots that were deployed in last year’s U.S. presidential election. After Donald Trump’s victory, “these accounts sort of went dark,” Ferrara says. Some roared back to life in April, on the eve of the French election, pushing Marine Le Pen, the far-right candidate in France’s election, he says. A number even switched to French language.
Ferrara is now investigating whether the same bots are active in Germany. If that’s the case, a handful of bad actors may be leading a veritable army of social media bots, seeking to tip elections in country after country.