Twitter’s never-ending fight against spam accounts is now a problem for its new owner: Elon Musk, who in April vowed to defeat the bot plague or “die trying.”
He later cited bots as a reason for backing out of the deal to acquire the social network. Now that the billionaire has completed the purchase, he faces the task of making good on his promise to expunge the fake profiles that have plagued him and plagued Twitter since long before he expressed interest in buying it.
The stakes are high in the challenge. The number of bots matters, because advertisers—Twitter’s main source of revenue—want to know roughly how many real human beings their messages are reaching when they buy ads. It’s also important in preventing ill-intentioned individuals from amassing an army of accounts in order to amplify false information or harass political opponents.
“From my point of view, the big picture is: How do we make Twitter a better place for everyone,” said Emilio Ferrara, an expert on counting bots who over the summer has been working on investigating the problem for Musk. He mentioned the “value of the platform as a social experience, as a collective place to have civil discussions and speak freely without interference from nefarious accounts” and without scams, spam, pornography, or harassment.
To find out how bad the bot problem was, Musk hired Ferrara and other data scientists to investigate. At the time, he sought to show that Twitter was misleading the public when he said that less than 5% of its daily active users were fake or spam accounts. If Twitter lied or withheld crucial information about the bot count, Musk could argue that he was justified in terminating the $44 billion deal.
Ferrara, an associate professor of computer science and communications at the University of Southern California, said he had no real interest in whether Musk ended up owning the social network.
Instead, he was hopeful that “any findings could help improve the platform,” Ferrara told The Associated Press, speaking for the first time about his intended role as Musk’s expert witness in a trial.
The question now is what Musk will do with that information. Ferrara’s presentation — some 350 pages of analysis and supporting documents — is locked away in confidential court files, and he indicated that he cannot release his conclusions.
Former Twitter executives and their lawyers said Musk wildly exaggerated the issue because he had buyer’s remorse. Accurate counts are “almost impossible” to achieve because any bot estimate is based on assumptions that can lead to bias, said Filippo Menczer, a researcher who did not work for either side in the dispute.
“Nobody knows exactly how bad the problem is,” said Menczer, director of the Social Media Observatory at Indiana University. “I would guess it’s not as bad as Musk said or as good as Twitter claimed.”
Analytics firm CounterAction, which worked with Ferrara, said it concluded in a July 18 report submitted to the court that Twitter’s spam rate by monetizable accounts — those with value to advertisers — was at least 10%. and that could reach 14.2%, depending on how the rate is measured.
Trevor Davis, the firm’s founder and CEO, said that analysis was based on internal data Twitter gave Musk, but the company declined to provide additional data requested by Musk’s team.
“We expect that access to the withheld data will reveal an even higher actual spam rate,” Davis said in a prepared statement.
Musk has long worried about Twitter spambots promoting cryptocurrency, in part because as a famous user with more than 110 million followers, he sees a lot of them. Some scammers have set up accounts mimicking Musk’s name and likeness to try to make people think he’s endorsing something.
Not all bots are bad. Twitter encourages the use of automated accounts that report weather and earthquakes, or post humor or lines from literary classics. Twitter also allows anonymity, which protects freedom of expression and privacy, especially in authoritarian regions, but this practice can make it difficult to root out malicious fake accounts.
Ferrara first came to Twitter’s attention following revelations that Russia used social media to meddle in the 2016 US presidential election, when he led a research group that estimated that between 9% and 15% of active Twitter accounts in English were bots.
In a blog post shortly afterward, Twitter complained that such external investigations are “often inaccurate and methodologically flawed.” The company has repeatedly reported the figure below 5% in its quarterly reports to the Securities and Exchange Commission, although it also warns that it could be higher.
Before Musk’s takeover, Twitter said it removed 1 million spam accounts every day. To calculate how many accounts are malicious spam, Twitter reviews thousands of random accounts, using public and private data such as IP addresses, phone numbers, geolocation, and how the account behaves when it’s active.
But in recent months, Musk and Twitter have argued over the methodology. Twitter uses a metric it calls mDAU, which stands for Monetizable Daily Active Usage.
That “is literally a metric invented by them,” Ferrara said. “You can’t compare that metric to any other service.”
When Musk began publicly raising doubts about the bot numbers after agreeing to buy the company, another firm said it had the answer.
“That slippery number you’re looking for…we got it. It’s 13.7%,” Israel-based Cyabra tweeted on May 17, swooning Musk to get his attention.
Cyabra’s machine learning technology scans a large number of social media profiles to track behavior patterns, trying to identify those behaving like humans. These guesses may fail, but the tweet caught the attention of people close to Musk and perhaps the billionaire himself.
Cyabra CEO Dan Brahmy said the company began working with Musk’s team in late May. Regardless of what the true number is, he said it’s not going to be an easy problem to solve.
“Some bots are definitely obnoxious,” Brahmy said. “The trade-off is between being extremely stringent with registration standards and information security or being extremely open” in a way that fosters freedom of expression and creativity.