While fake news has always been around (e.g. urban legends), the 2016 U.S. presidential election invigorated the topic. According to a joint NYU and Stanford fake news study, fake pro-Trump stories were shared at least 30 million times on social media, and fake pro-Clinton stories were shared at least 7.6 million times.
Countless more Americans saw —and even believed — these shared fake news stories.
Artificial intelligence (AI) won’t be a panacea, but it can play a major part in fighting fake news.
The key is for AI to augment human intelligence. Computers are really fast at consuming and analyzing large amounts of data, which means they can do a lot of the grunt work to determine if it’s likely or unlikely for news to be fake.
The latest AI advances would allow consumers and human fact checkers at news and sharing platforms to more quickly discern fake news from what’s real.
How a fake news AI would work
There are three major considerations that a human would go through to decide if a piece of news is fake or not that can be improved with AI:
Credibility of the site
Analyzing the domain name provides insight into if the site is credible. Does it have a common .com top-level domain, or is it something more suspicious, like .zip? Has the site been previously flagged for fake news?
A technical approach can simulate all of these actions. A supervised machine learning model can be trained on a labeled data set of domains that are legitimate versus contain fake news. Features inputted into the model can include domain length, characters used, the top-level domain used, and the age of the domain.
The model then can quickly classify if a site seems credible.
Credibility of the author
Judging the author’s credibility is another important step. First, verify there’s a person backing this article and find out if the person is real, and credible, based on their bio and background.
Do a “Search Google for Image,” to see if the person’s picture is actually found online attributed to a completely different person. If so, chances are that what you’re looking at is a stolen picture for a new identity.
The technical approach would mirror this but would involve less AI and more database /third-party lookups.
Content matches other legitimate sources
Humans are better at fact-checking than computers, but computers are starting to understand the underlying meaning behind language. This is important, since the same fact can be written numerous ways in the English language. This is where state-of-the-art AI and natural language processing (NLP) comes in.
While creating a fully automated model to determine if an article’s content is supported or contradicted by other sources is still difficult, it’s possible to create an AI-assisted approach.
The Fake News Challenge (FNC) believes that “Stance Detection” is the best first step. Stance Detection is the act of labeling the relationship between a headline and a text snippet as one of four things: agree, disagree, discuss, or unrelated.
This would help fact checkers gather viewpoints for a particular topic quickly, checking articles labeled “agree”, “disagree” or “discuss” with credible sites, and save hours of human time crawling the Web.
Putting it all together
Ultimately, these are all signals that a human would take in, and make a decision about whether they believe the news is fake or not. Similarly, a final machine learning ensemble model would take all of these inputs as features, and then classify whether the article is fake or not.
There won’t just be one method for drastically reducing the prevalence of fake news. Google and Facebook preventing ad revenues from fake news sites is a major force. It also helps that there is an increase in general awareness of fake news.
Ultimately, we as individuals still have to be responsible for what we read and believe, but AI will be a great tool used for this battle.