Bot or not? A.I. looks at Twitter behavior to sort real accounts from fake

For the past several years, there’s been heightened concern about the impact of so-called bots on platforms like Twitter. A bot in this context is a fake account synonymous with helping to spread fake news or misinformation online. But how exactly do you tell the difference between an actual human user and a bot? While clues such as the use of the basic default “egg” avatar, a username consisting of long strings of numbers, and a penchant for tweeting about certain topics might provide a few pointers, that’s hardly conclusive evidence.

That’s the challenge a recent project from a pair of researchers at the University of Southern California and University of London set out to solve. They have created an A.I. that’s designed to sort fake Twitter accounts from the real deal, based on their patterns of online behavior.

“Detecting bots can be very challenging as they continuously evolve and become more sophisticated,” Emilio Ferrara, research assistant professor in the USC Department of Computer Science, told Digital Trends. “Existing tools that work well with older and simpler types of bots are not as accurate to predict more complex ones. So it’s always exciting to identify new, previously unknown characteristics of the behavior of human users that are not yet exhibited by bots. This could [be used to help] improve the accuracy of detection tools.”

The researchers leveraged various datasets of hand-labeled examples of both fake and real Twitter account messages, produced by other researchers in the community. In total, they trained their system on 8.4 million tweets from 3,500 human accounts and an additional 3.4 million tweets from 5,000 bots. This helped them to uncover a variety of insights into tweeting patterns. For instance, human users are up to five times more likely to reply to messages. They also get increasingly interactive with other users over the course of a long Twitter session, while the length of an average tweet decreases during this same time frame. Bots, on the other hand, show no such changes.

But don’t expect this work to be the definitive lasting word in this field. Just like the cat-and-mouse game between software companies and hackers, whereby one group tries to close vulnerabilities and the other works to find new ones, the field of bot discovery will continue to develop.

“These findings will inform future bot detection tools,” Ferrara said. “However, we expect that bot-making tools will see these findings as well, so it will be interesting to see when — or how soon — some of the open-source bot-making tools that are available online will adjust to reflect the human behavioral trends that we discovered.”

A paper describing the work was recently published in the journal Frontiers in Physics.

Editors' Recommendations