During the 2016 U.S. presidential election, inaccurate and misleading articles burned through social networks. Since then, tech companies—from behemoths like Facebook and Google to scrappy startups—have built tools to fight misinformation (including what many call “fake news,” though that term is highly politicized). Most companies have turned to artificial intelligence (AI) in hopes that fast and automated computer systems can deal with a problem that’s seemingly as big as the Internet.
“They’re all using AI because they need to scale,” says Claire Wardle, who leads the misinformation-fighting project First Draft, based in Harvard University’s John F. Kennedy School of Government. AI can speed up time-consuming steps, she says, such as going through the vast amount of content published online every day and flagging material that might be false.
But Wardle says AI can’t make the final judgment calls. “For machines, how do you code for ‘misleading’?” she says. “Even humans struggle with defining it. Life is messy and complicated and nuanced, and AI is still a long way from understanding that.”
Facebook, which was widely criticized for failing to take action against false content in 2016, says it will use AI to do better in the U.S. midterm elections this November—and in other elections around the world. Jim Kleban, a Facebook product manager who works on reducing misinformation in the site’s news feed, explains that Facebook now uses AI to augment human intelligence. The AI goes through the millions of links shared on Facebook every day to identify suspect content, which is then sent to human fact-checkers. “For the foreseeable future, all these systems will require hybrid solutions,” he says.
When fact-checkers rate a piece of content as false, Facebook places it lower in users’ news feeds. Kleban says this method reduces future views of that content by 80 percent.
Facebook’s AI is trained via machine learning, a technique in which an AI system takes in a vast data set of labeled material and independently finds patterns. For example, an image-sorting AI might look at millions of photos labeled either “cat” or “dog” and learn the distinguishing characteristics of felines and canines. But training an AI to recognize false content is much trickier.
Kleban says the Facebook AI uses a variety of signals to pick out articles that contain misinformation, starting with the source of the content: “Knowing that a page or a website has shared false content in the past is a good predictor that it will happen again,” he says. There may also be a discernable pattern in how false content propagates across the Web; Kleban says that’s an active area of research. As for the text itself, the AI isn’t equipped to evaluate statements for their truthfulness, but it can find signals, such as expressions of disbelief in the comment section.[shortcode ieee-pullquote quote=""Knowing that a page or a website has shared false content in the past is a good predictor that it will happen again."" float="right" expand=1]
The London-based startup Factmata, whose high-profile investors include Twitter cofounder Biz Stone and Craigslist founder Craig Newmark, is developing an AI system with a different approach. It specifically does not look at the content’s publisher or their reputation, says Factmata founder Dhruv Ghulati. “We want to judge content based on content itself,” he says.
Factmata’s system is also a hybrid of human and machine, though in a different configuration: The humans are experts who label content used for the AI’s training. “Things like fake news and propaganda are inherently nuanced and subjective,” Ghulati says. “It does require expertise to understand the nature of the content and tag it appropriately.” With that proprietary data set, Factmata is training its AI to recognize politically biased content, false content, and hate speech.
The company is currently working on the “back end” of the Internet, helping advertising exchanges avoid placing ads on problematic content, and it may work with social networks in the future. Factmata’s system flags suspect content and explains what makes it suspicious, but the company leaves it up to the client to decide whether to steer clear of or moderate that material.
Some companies that began with other journalistic purposes have joined the fray. NewsWhip, based in Dublin, sells an AI-powered tool to news organizations that finds hot content and predicts its spread, enabling news teams to jump on stories that are going viral. In recent elections in France, the United Kingdom, and Germany, journalists used its tool to spot and debunk false stories that were gaining traction on social networks.[shortcode ieee-pullquote quote=""The risk is that you try to get the perfect definition of fake news and never reach an answer."" float="right" expand=1]
Krzana, a London-based company, helps journalists find breaking news with a customized real-time news feed. Reporters use Krzana’s AI-enabled tool to discover content across four languages (with more languages to come) based on keywords and search terms they’ve selected. In recent elections in Mexico, a media coalition used Krzana to quickly find stories that might contain misinformation.
“Rather than waiting for these stories to be shared by a lot of people, journalists were among the first to read them,” says Krzana cofounder Toby Abel. “If they were fake, they could be countered very quickly.”
Abel says an AI misinformation detector can’t yet be reliable on its own, and he agrees that there needs to be a “human in the loop.” He cites an example from the 2018 Mexican election, in which a political candidate responded to accusations of Russian ties with a playful stunt: He went down to the docks and pronounced that he was waiting for his Russian submarine. “If someone had read that without outside context and understanding, it would sound like fake news—but it’s not,” Abel says.
Satire is one of the toughest problems for AI systems that try to identify false content. Companies are also grappling with misinformation in images, videos, graphs, and other nontextual content. The possibilities for deception seem endless: A photograph might be legitimate, for example, but its caption may be misleading.
Full Fact, a nonprofit fact-checking organization based in London, is trying to steer clear of the gray areas. It’s using machine learning to improve a tool that scans text and video transcripts, looking for factual claims on topics such as economic trends and legal proceedings that can be verified by human fact-checkers. Mevan Babakar, Full Fact’s head of automated fact-checking, says its tool also clusters together similar claims from many different news sources. “So at the beginning of each editorial day, I can say to my fact-checkers, ‘Here are the top five [claims] that are spreading like wildfire.’”
Today’s AI systems may not be ready to parse complicated claims independently or to make sophisticated decisions about truth, says Ghulati of Factmata, but that doesn’t mean they shouldn’t be deployed now. “The risk is that you try to get the perfect definition of fake news and never reach an answer,” he says. “The important thing is to build something.”
An abridged version of this article appears in the September 2018 print issue.