It may have been the first bit of fake news in the history of the Internet: in 1984, someone posted on Usenet that the Soviet Union was joining the network. It was a harmless April’s Fools Day prank, a far cry from today’s weaponized disinformation campaigns and unscrupulous fabrications designed to turn a quick profit. In 2017, misleading and maliciously false online content is so prolific that we humans have little hope of digging ourselves out of the mire. Instead, it looks increasingly likely that the machines will have to save us.
One algorithm meant to shine a light in the darkness is AdVerif.ai, which is run by a startup of the same name. The artificially intelligent software is built to detect phony stories, nudity, malware, and a host of other types of problematic content. AdVerif.ai, which launched a beta version in November, currently works with content platforms and advertising networks in the United States and Europe that don’t want to be associated with false or potentially offensive stories.
The company saw an opportunity in focusing on a product for companies as opposed to something for an average user, according to Or Levi, AdVerif.ai’s founder. While individual consumers might not worry about the veracity of each story they are clicking on, advertisers and content platforms have something to lose by hosting or advertising bad content. And if they make changes to their services, they can be effective in cutting off revenue streams for people who earn money creating fake news. “It would be a big step in fighting this type of content,” Levi says.
AdVerif.ai scans content to spot telltale signs that something is amiss—like headlines not matching the body, for example, or too many capital letters in a headline. It also cross-checks each story with its database of thousands of legitimate and fake stories, which is updated weekly. The clients see a report for each piece the system considered, with scores that assess the likelihood that something is fake news, carries malware, or contains anything else they’ve ask the system to look out for, like nudity. Eventually, Levi says he plans to add the ability to spot manipulated images and have a browser plugin.
Testing a demo version of the AdVerif.ai, the AI recognized the Onion as satire (which has fooled many people in the past). Breitbart stories were classified as “unreliable, right, political, bias,” while Cosmopolitan was considered “left.” It could tell when a Twitter account was using a logo but the links weren’t associated with the brand it was portraying. AdVerif.ai not only found that a story on Natural News with the headline “Evidence points to Bitcoin being an NSA-engineered psyop to roll out one-world digital currency” was from a blacklisted site, but identified it as a fake news story popping up on other blacklisted sites without any references in legitimate news organizations.
Some dubious stories still get through. On a site called Action News 3, a post headlined “NFL Player Photographed Burning an American Flag in Locker Room!” wasn’t caught, though it’s been proved to be a fabrication. To help the system learn as it goes, its blacklist of fake stories can be updated manually on a story-by-story basis.
AdVerif.ai isn’t the only startup that sees an opportunity in providing an AI-powered truth serum for online companies. Cybersecurity firms in particular have been quick to add bot- and fake news-spotting operations to their repertoire, pointing out how similar a lot of the methods look to hacking. Facebook is tweaking its algorithms to deemphasize fake news in its newsfeeds, and Google partnered with a fact-checking site—so far with uneven results. The Fake News Challenge, a competition run by volunteers in the AI community, launched at the end of last year with the goal of encouraging the development of tools that could help combat bad-faith reporting.
Delip Rao, one of its organizers and the founder of Joostware, a company that creates machine-learning systems, said spotting fake news has so many facets that the challenge is actually going to be done in multiple steps. The first step is “stance detection,” or taking one story and figuring out what other news sites have to say about the topic. This would allow human fact checkers to rely on stories to validate other stories, and spend less time checking individual pieces.
The Fake News Challenge released data sets for teams to use, with 50 teams submitting entries. Talos Intelligence, a cybersecurity division of Cisco, won the challenge with an algorithm that got more than 80 percent correct—not quite ready for prime time, but still an encouraging result. The next challenge might take on images with overlay text (think memes, but with fake news), a format that is often promoted on social media, since its format is harder for algorithms to break down and understand.
“We want to basically build the best tools for the fact checkers so they can work very quickly,” Rao said. “Like fact checkers on steroids.”
Even if a system is developed that is effective in beating back the tide of fake content, though, it’s unlikely to be the end of the story. Artificial-intelligence systems are already able to create fake text, as well as incredibly convincing images and video. (see “Real or Fake? AI Is Making It Very Hard to Know”). Perhaps because of this, a recent Gartner study predicted that by 2022, the majority of people in advanced economies will see more false than true information. The same report found that even before that happens, faked content will outpace AI’s ability to detect it, changing how we trust digital information.
What AdVerif.ai and others represent, then, looks less like the final word in the war on fake content than the opening round of an arms race, in which fake content creators get their own AI that can outmaneuver the “good” AIs (see “AI Could Set Us Back 100 Years When It Comes to How We Consume News”). As a society, we may yet have to reevaluate how we get our information.
A Roomba recorded a woman on the toilet. How did screenshots end up on Facebook?
Robot vacuum companies say your images are safe, but a sprawling global supply chain for data from our devices creates risk.
The viral AI avatar app Lensa undressed me—without my consent
My avatars were cartoonishly pornified, while my male colleagues got to be astronauts, explorers, and inventors.
Roomba testers feel misled after intimate images ended up on Facebook
An MIT Technology Review investigation recently revealed how images of a minor and a tester on the toilet ended up on social media. iRobot said it had consent to collect this kind of data from inside homes—but participants say otherwise.
How to spot AI-generated text
The internet is increasingly awash with text written by AI software. We need new tools to detect it.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.