None of that should have been a surprise for Musk, who tweeted that he was pausing the deal “pending details supporting calculation that spam/fake accounts do indeed represent less than 5% of users.” (He later said he was still committed to the $44 billion takeover, and some investors said they thought Musk was angling for a lower price that would not weigh as heavily on the Tesla shares he has pledged as loan collateral.)
Musk was referring to a Twitter regulatory filing this month that said false or spam accounts constituted fewer than 5 percent of its 229 million daily active users.
The number is hardly new: Twitter has been giving the same estimate for years, even though critics and experts have said they believe the company is lowballing the actual number of such accounts.
“That 5 percent is a very opportune and chosen metric,” said a former employee who spoke on the condition of anonymity tonot alienate a former employer. “They didn’t want it to be big, but also not small, because then they could get caught in a lie.”
Twitter declined to comment for this story. A person familiar with the acquisition negotiations, who spoke on the condition of anonymity to describe sensitive matters, said the negotiations were proceeding as usual, despite Musk’s claims of a hold. The person said requests to learn more about spam and fake accounts were routine for a potential acquirer of a social media business.
Twitter’s history with spam goes as far back as its 2013 public offering, when it disclosed the risk of automated accounts — a problem faced by all social media companies. (Facebook has also estimated that fake profiles account for about 5 percent of its user base.) For years, people wanting to manipulate public opinion could buy hundreds of fake accounts to pump up a celebrity or a product’s standing.
But the problem took a grave turn in 2016, when Russian operatives from the Internet Research Agency sowed disinformation about the election to millions of people in favor of then-presidential candidate Donald Trump, on Twitter, Facebook, YouTube and other platforms.
The Russia controversy, which culminated with congressional hearings in 2017, prompted Twitter to crack down. By 2018, the company had launched an initiative called Healthy Conversations and was culling more than a million fake accounts a day from its platform, The Washington Post reported at the time.
To tackle the problem internally, Twitter engineers launched an internal initiative called Operation Megaphone, in which they bought hundreds of fake accounts and studied their behavior.
“You grab a species and find others that behave like that species,” said a person familiar with the internal effort, speaking on the condition of anonymity to freely describe it. The person said they thought the 5 percent was probably an underestimate. “You’re making predications based on what you’ve observed, but you don’t know what you don’t know.”
Critics have argued that Twitter has an incentive to downplay the number of fake accounts on its platform and that the bot problem is far worse than the company admits. The company also allows some automation of accounts, such as news aggregators that pass along articles about specific topics or weather reports at set times or postings of photos every hour.
Twitter does not include automated accounts in its calculations of daily active users because those accounts do not view advertising, and it argues that all social media services have some amount of spam and fake accounts.
But the 5 percent number has long raised eyebrows among outside researchers who conduct deep studies of behavior on the platform around critical issues including public health and politics.
“Whether it was covid, or many elections studies in the U.S. and other countries, or around various movies, we see way more than that number of bots,” said Kathleen Carley, a computer science professor at Carnegie Mellon who directs the university’s Center for the Computational Analysis of Social and Organizational Systems.
“In all of the different studies we have done collectively, the number of bots ranges: We have seen as low as 5 percent, and we have seen as high as 35 percent.”
Carley said the proportion of bots tends to be much higher on topics where there is a clear financial goal, such as promoting a product or a stock, or a clear political goal, such as electing a candidate or encouraging distrust and division.
There are also very different types of bots, including basic promotional spam, nation-state accounts and amplifiers for commercial hire.
Rapidly developing technology allows geopolitical forces to seem more human, peppering their comments with personal asides, and to try to manipulate the flow of group conversations and opinions.
As an example, Carley said some pro-Ukraine bots were engaging in dialogue with groups normally focused on other issues to try to build coalitions supporting Ukrainian goals. “The number of bot technologies has gone up, and the cost of creating a bot has gone down,” she said.
Outsiders said it was very difficult for them to produce a good estimate of bot traffic with the limited help Twitter provides to research efforts.
“When we use our Botometer tool to evaluate a group of accounts, the result is a spectrum ranging from very humanlike to very bot-like,” said Kaicheng Yang, a doctoral student at Indiana University.
“In between are the so-called cyborgs controlled both by humans and software. We will always mistake bots for humans and humans for bots, no matter where we draw the line.”
Twitter gives some researchers access to a giant number of tweets, known inside the company as the “fire hose” for its immense volume and speed. But even that does not have the clues that would make identifying bots easier, such as the email addresses and phone numbers associated with the accounts behind each tweet.
“Pretty much every effort outside of Twitter to detect `botness’ is fatally flawed,” said Alex Stamos, the former Facebook security chief who leads the Stanford Internet Observatory.
Twitter itself does not do nearly as much as it could to hunt down and eliminate bots, two former employees told The Post. But two other former employees told The Post that after 2018, the company acted far more aggressively.
Some of the people speculated that financial incentives motivate Twitter to not find them. If the company identifies more bots and removes them, the number of “monetizable daily average users” would go down, the amount it could charge for advertising would also decline and the stock price would follow, as it did after Twitter confirmed a big cull to The Post in 2018.
The company uses a number of programs to seek out and block automated commercial accounts, but they are most effective at catching the obvious spammers, such as those that register hundreds of new accounts on the same day from the same device, the former employees said.
To produce its quarterly bot estimate, the company looks at a sample of millions of tweets.
But that is a tiny percentage of the total, and they are from a wide spectrum — not the hot-button issues that draw the most spam and the most viewer impressions.
“They honestly don’t know,” the former employee said. “There was significant resistance to doing any meaningful quantification.”
Twitter has protected itself legally with a disclaimer in its quarterly reports saying it could be off by a lot.
“We applied significant judgment, so our estimation of false or spam accounts may not accurately represent the actual number of such accounts, and the actual number of false or spam accounts could be higher than we have estimated,” Twitter said in its latest quarterly report.