This AI can help spot fake news

April 23, 2021
AI might soon be helping human editors figure out real news from fake. (AP Photo/Steven Senne)

AI might soon be helping human editors figure out real news from fake. (AP Photo/Steven Senne)

Most people don't have the time or expertise to check whether everything they see online is true. But a new method of detecting fake news could someday help users, news publications and government agencies shut down the falsehoods that flood our social media feeds. 

In a patent application published March 25, FakeNewsNet is described as a new way of determining which stories are untrustworthy that uses a comprehensive set of data on both real and fake news stories to inform a machine-learning model. The system's lead inventor, Kai Shu, told The Academic Times that the innovation is designed to work alongside people on the frontlines of the information war. "It's a collaboration between the AI we're building and human experts," he said. "We're not replacing [people], but enabling them to better do their work."  

An assistant professor of computer science at the Illinois Institute of Technology, Shu said current methods of flagging fake news — algorithms that detect bots and human fact-checking — fall well short of addressing the scale of the problem. In 2020, nearly one in five engagements with the top 100 news sources on social media came from unreliable websites — more than twice as many as in 2019 — according to the misinformation-tracking company NewsGuard. 

Fact checking, one of journalism's main tools to fight fake news, is time-consuming and requires experts to comb through an overwhelming amount of online content. By the time fact checkers have published a debunk-style piece, the story containing false or misleading information could have been viewed by thousands or even millions of people. "At this stage, there is no tool or technology that is satisfactory in detecting fake news," Shu said. 

Shu started working on FakeNewsNet in 2017. He began by examining the existing literature on the subject. "I realized there wasn't much computational research going on to think about how machine learning can help solve this problem," he said. "At the time, most people were talking about the common patterns of fake news — [telling people] you should pay attention to the URL of the website, or the headline; maybe it's clickbait. These are not deterministic features to help detect fake news. I realized there was a pressing need."

Building a comprehensive data set was the next and most critical step. FakeNewsNet's detection system is supported by three types of information: news content that has been flagged as false or misleading by professional journalists; the level of engagement with social media posts containing problematic information; and spatiotemporal information on the networks of profiles that spread fake news. 

The researchers looked for potentially telltale signs of fake news, including the type of engagement a post is receiving. For example, the comments sections of fake news stories tend to be adversarial and polarized, while people who comment on professionally produced news stories agree with each other more often. "This is all very important information we can leverage to understand what is happening in terms of fake news and true news spreading," said Shu. 

In practice, the tool performs well for false or misleading news stories about politics, but it's limited in terms of adapting to emerging events such as the COVID pandemic, Shu said. Early in 2020, his team scrambled to add data on the news coverage and misinformation related to the ongoing public health crisis. "Does the [political] model work well in this new domain? Not necessarily," he said. "The topics are totally different. When we use it to predict topics in some new domain, it does not work well. So, I think it's really important to develop domain-adaptive detectors that can really work for newly emerging fake news." 

Not to mention, false or misleading content is now being machine-generated in ways that make it nearly indistinguishable from human-written disinformation. "How can we win this game of detecting AI-created fake news? This is a very new problem we're actively working on," Shu said. And there is far more to be done: He's continually improving the algorithm behind FakeNewsNet, which he hopes to translate into a commercial product, and has discussed launching a startup with his colleagues. But he doesn't anticipate that social media platforms themselves will be interested in his invention. 

"Social media companies have their own business models," he said. "It's all about traffic. If they want revenue, they need users to engage with their platform. … Maybe fake news gets more traffic, but we really don't know because these tech giants are not very transparent. Do they really want to stop this problem? I highly doubt it."  

The application for the patent, "Method and apparatus for collecting, detecting and visualizing fake news," was filed Sept. 11, 2020 with the U.S. Patent and Trademark Office. It was published March 25, 2021, with the application number 17/018877. The inventors of the pending patent are Kai Shu, Deepak Mahudeswaran and Huan Liu. The assignee is the Arizona Board of Regents on Behalf of Arizona State University.

Parola Analytics provided technical research for this story.

Saving
We use cookies to improve your experience on our site and to show you relevant advertising.