In a video published on YouTube, Game of Thrones actor Kit Harington—in character as Jon Snow—apologized for several mistakes made during the controversial eighth season of the hit HBO show.
There’s just one problem: Harington never uttered the words that appear to come from his mouth. The now-viral video is a so-called deepfake, the term used for videos that have been digitally altered with the intent to deceive viewers.
As deepfake technology gets better, and is available to more people, the Internet could be inundated with believable fake videos that will make it harder for people to discern the truth, experts say.
“Let there be no question that this is a race. The better the manipulators get, the better the detectors need to be. And there are certainly orders of magnitude more manipulators in the race than detectors,” said David Doermann, director of the Artificial Intelligence Institute at the University of Buffalo, during a House Intelligence Committee hearing last Thursday.
While the Jon Snow video was intended as a parody, malicious examples of fake videos have been published online in recent weeks. Last month, a video of House speaker Nancy Pelosi, 79, was altered to make the highest-ranking woman in U.S. government sound as though she was slurring her words.
“PELOSI STAMMERS THROUGH NEWS CONFERENCE” pic.twitter.com/1OyCyqRTuk
— Donald J. Trump (@realDonaldTrump) May 24, 2019
Attention to the video grew as it spread on social networks; it received an even larger audience after broadcast television networks showed the clip and President Donald Trump tweeted it. (Facebook declined to remove the video, stating that it did not violate the social network’s community guidelines.)
Facebook policy was once again put to the test last week when someone created a deepfake video of co-founder and CEO Mark Zuckerberg declaring in a robotic sounding voice: “Whoever controls the data, controls the future.” The Menlo Park, Calif. company, which endured withering criticism this year over its data privacy practices, stuck to its policy and declined to remove the video. (A Facebook spokesperson told Fortune last week that steps would be taken to slow the spread of “inauthentic” content on the network if a third-party fact checker determined it to be false. The content itself wouldn’t be wired from the platform.)
“It’s a real problem,” says Falon Fatemi, CEO of Node, an AI-powered discovery engine. “This is not a case where we can put the genie back in the bottle.”
https://www.instagram.com/p/ByaVigGFP2U/
Deepfakes are getting frighteningly easy to make, even for people who don’t have major technical skills. A simple Google search for “how to create a deepfake” also offers guides for people looking to make their own fake videos. And last week, researchers from Stanford University, the Max Planck Institute for Informatics, Princeton University, and Adobe Research showed how they were able to develop software that allows people to easily change what someone is saying in a video. It’s as easy as editing a Microsoft Word document.
The government is trying to get ahead of the problem, but there’s so much more work to do. MediFor, a program run by DARPA, is working with researchers to create tools that can detect and localize manipulated videos and images.
Generative adversarial networks, called GANs, could also be another tool businesses may invest in to fight deepfakes. GANs are two neural networks—translation: artificial intelligence—that learn from each other.
“One gets trained on how to create a Picasso. Another gets trained on how to detect a fake Picasso. They basically argue with each other to train themselves,” Fatemi says. “It just keeps going. And what you end up with are two neural networks: one that is good at creating fakes and one that is good at detecting them.”
Yet for all of the technical tools being developed to catch up to the manipulators, experts say deepfakes are a problem that must be addressed at a broader level. Facebook makes its money by running advertisements against user activity, rather than charging users a subscription fee for the service. That business model incentivizes Facebook to keep popular content on the platform, Fatemi says, even if it’s not real.
“From an ethical perspective, the fact Facebook is a platform for distributing information at scale, they probably need to take a more serious stance,” she says. “What is always going to be challenging about any business with content on their platform is the financial aspect. Their business model has to change to encourage those ethics as well.”
Look no further than this fact as proof: False news headlines and and video clips continue to spread on social media well after they’ve been debunked.
The last line of defense, then, may be education. But the race is on. Fatemi says future deepfakes may be highly personalized and tailored to target specific groups—and it won’t stop at video. Fake stories will be produced to mislead specific groups at an even higher volume than today. Discerning the truth will prove to be increasingly difficult.
“If it were highly personalized, it would become a lot harder to track and prevent. This is where it gets particularly scary,” Fatemi says. “The lack of education [and] critical thinking at scale are the problems.”
More must-read stories from Fortune:
—Phishing hackers can now bypass two-factor authentication
—Apple’s sign-in feature is a “shot across the bow” at tech giant rivals
—Uber’s CEO has absorbed the COO role for more control
—Google is changing its search results. Here’s what to expect
—Listen to our new audio briefing, Fortune 500 Daily
Catch up with Data Sheet, Fortune‘s daily digest on the business of tech.
