Fake content is a big problem. It’s not just fake news ” it’s fake websites, social media profiles, and ads. There are plenty of uses for this stuff, ranging from subversive to downright malicious intent. At its core, fake content threatens the foundation of trust in society, a foundation necessary for societies to function properly, particularly democratic societies. Russia knows this, and it’s the reason why the Kremlin has an army of trolls disseminating fake content. The ultimate goal is to undermine American and European democracies.
In American homes, families are becoming increasingly concerned about fake news. A study by Panda Security found that fake news is right up there on the list of parents’ concerns next to sexual predators. This concern affects parents’ views towards media sites: nearly 50 percent view alt-right site Breitbart as unsafe for children and 20 percent feel the same about CNN. Nearly 6 percent of the parents surveyed block Facebook, which has been accused of unwittingly propagating fake news that may have influenced the 2016 presidential election. In comparison, only 2.5 percent of those parents block PornHub. Sexual predators oftentimes set up fake profiles on Facebook through which they lure kids.
Fake news isn’t the only issue. There’s another kind of fake content that is more immediately harmful. Senior citizens regularly fall prey to phishing scams, in which scammers send them emails directing them to go to fake websites where they’re prompted to enter their information. Perpetrators then steal their identity and empty their bank accounts, or they obtain their credit card numbers. Fake sites for phishing scams are hard to bust and take down because they’re temporary.
The sheer number of fake sites, fake pieces of content, and fake profiles makes it hard for humans tasked with determining what’s fake and what isn’t. Artificial intelligence, with its ability to analyze massive amounts of data quickly, would seem like a good candidate for detecting fake content. It would seem that determining whether a piece of content is fake would be a matter of logical inference based on a set of protocols.
Facebook, AI, and Fake Content
Mark Zuckerberg is counting on AI for the future of Facebook. The Verge reports Zuckerberg told congress that Facebook’s AI will be sophisticated enough to deal with linguistic nuance in 5 to 10 years. Count The Verge‘s Sarah Jeong among the doubters. AI might never be capable of dealing with certain categories of content, like fake news, she says. To her, Zuckerberg’s appeal to AI is a dodge. To MIT’s Jackie Snow, it might be the beginning of an automated arms race.
As companies like Facebook and Google work on improving their algorithms to detect fake content, hackers are working on AI that will make fake content harder to detect. There are some advancements in detection software that look promising, but because of the nature of fake content, there’s no telling exactly how effective they are.
AdVerif.ai is one of the software solutions on the market. Ms. Snow reports that AdVerif.ai was able to tell that the story, Evidence points to Bitcoin being an NSA-engineered psyop to roll out one-world digital currency is fake, but it wasn’t able to say the same about a story titled, NFL Player Photographed Burning an American Flag in Locker Room!
AdVerif.ai has a database of real and fake stories with which it compares new entries. Users can manually update the database to help the software learn. The software checks for veracity, malware, and other things the user wants it to look for, such as nudity. It classified Breitbart stories as unreliable, right, political, bias. It also was able to tell that The Onion is a satire site. The software looks for anomalies in content, such as too many capitalized letters or a headline that doesn’t match the story. Because it’s still relying on humans to help it learn what’s fake and what’s not, AdVerif.ai is still subject to human error and bad data.
Encouragingly, an AI program from Cisco’s cyber security division, Talos Intelligence, was able to get stance detection right 82 percent of the time in the Fake News Challenge (FNC). It beat 49 other entrants to earn the top prize. Stance detection is part of what AdVerif.ai does, it’s a comparison of an article’s title to its body. Contestants in the FNC received a dataset that had been vetted by journalists. AI had to determine whether an article’s headline was related or unrelated to the body of the article, and what relationship the body’s text has to the headline, specifically, whether it agrees, disagrees, discusses, or is unrelated.
The idea is to provide human fact checkers with information that takes away some of the legwork of fact checking. Interestingly, the FNC stance detection is more viable for AI than truth detection at this stage in the game because truth detection is a difficult business, even for humans. The FNC says, Any dataset containing claims with associated truth’ labels is going to be contested as biased. This gets at the heart of the issue. Determining truth starts with humans and is inherently subjective. For AI to be able to determine truth, it’s going to have to be as intelligent and capable of understanding linguistic nuances as a real human being. Even then, whether you’re AI or you’re human, your data is only as good as your sources.
The Future of Fake
A Gartner study predicts that, By 2022, the majority of individuals in mature economies will consume more false information than true information. This is due in large part to advancements in AI. The study says that by 2020 AI’s ability to create fake content will outpace AI’s ability to detect it. The deluge of fake content will include video, memes, articles–no realm of the digital world is safe.
If that’s the case, the world needs more experts in machine learning, deep learning, and other big data-related careers who can step up and work on AI programs to combat a fake content future. Together with teams of human investigators, the top AI programs will be indispensable and integral to a future in which democracy can thrive.
