Seeing may not be believing: AI deepfakes and trust in media
by Karl D. Stephan | October 15, 2018
The 1994 movie Forrest Gump featured authentic-looking newsreel footage of the 1960s in which President Kennedy allegedly appeared with Tom Hanks’ fictional Gump character. Trick photography is as old as cinema, and the only remarkable thing about such scenes were the technical care with which they were produced. At the time, these effects were state-of-the-art and took substantial resources of a major studio.
What only Hollywood could do back in the 1990s is soon coming to the average hacker everywhere, thanks to advanced artificial-intelligence (AI) deep-learning algorithms that have recently made it possible to create extremely realistic-looking audio and video clips that are, basically, lies.
An article in October’s Scientific American describes both the progress that AI experts have made in reducing the amount of labour and expertise needed to create fake recordings, and the implications that wider availability of such technology poses for the already eroded public trust in media. Fakes made with advanced deep-learning AI (called “deepfakes”) can be so good that even people who personally know the subject in question—President Obama, in one example—couldn’t tell it was fake.
The critical issue posed in the article is “what will happen if a deepfake with significant social or political implications goes viral”? Such fakes could be especially harmful if released just before a major election. It takes time and expertise to determine whether a video or audio record has been faked, and as technology progresses, that difficulty will only increase. By the time a faked video that influences an election has been revealed as a fake, the election could be over. We faced something similar to this in 2016, as it has been conclusively shown that Russian-based hackers spread disinformation of many kinds during the run-up to the Presidential election.
Some voters will believe anything they see, especially if it fits in with their prejudices. But the firmer a voter is embedded in one camp or the other, the less likely they are to change their vote based on a single fake video. The people who can actually change the outcome of an election are those who are undecided going into the final stretch of the campaign. If they are taken in by a fake video, then real harm has been done to the process.
On the other hand, the public as a collective body is not always as stupid as experts naively think. If a deepfake ever manages to be widely believed at a critical moment, and the fakery is later revealed publicly, the more thoughtful among us will keep in mind the possibility of fakery whenever we watch a video or listen to audio in the future. This can be likened to an immune-system response. The first invasion of a new pathogen into one’s body may do considerable damage, but a healthy immune system creates antibodies that fight off both the current infection and also any future attempts at invasion by the same pathogen.
If deepfakes begin to affect the public conversation significantly, we will all get used to the fact that any audio or video, no matter how genuine-looking, could be the concoction of some hacker’s imagination.
Low-tech versions of this sort of thing happen all the time, but with lower stakes. When I’m not writing this blog, I find time to do some lightning research, and a few years ago someone forwarded me a YouTube clip purporting to be a security-camera video of a guy who got struck by lightning, not once, but twice, and survived both times. I watched the grainy monochrome recording of a man walking toward the camera on a sidewalk. Suddenly there was a bright full-screen flash, and he’s down on the pavement, apparently dead. Then he raises his head, shakes himself, and groggily rises to his feet, only to have a second flash knock him down again. I heard from another lightning expert about this video that it was definitely fake.
Some people want so desperately to achieve viral fame, that they will go to the trouble of setting up an elaborate fraud like this one just on the hopes that their production will be kooky enough to get shared widely. And in this case, they succeeded.
Speaking theologically for a change, some (including myself) trace the origin of lies back to the father of lies himself, the devil, and attribute lying to the only Christian doctrine for which there is abundant empirical evidence: original sin. No amount of high-tech defence is going to stop some people from lying, and if they can bend deep-learning AI to nefarious purposes such as creating inflammatory deepfake videos, they will. The best defence from such scurrilous behaviour is not necessarily just working harder to make fake-video-detection technology better, although that is a good thing. It is to bear in mind that people will lie sometimes, and to use the time-honoured rules of evidence to seek the truth in any situation. And to bear in mind something that is often forgotten these days, that there is such a thing as objective truth.
I think a more serious problem than deepfake videos is the fact that in pursuit of the online dollar, social media companies have trained millions of their customers to react to online information with their lizard brains, going for the thing that is most titillating and most conforming to one’s existing prejudices regardless of the likelihood that it’s true. They have created an invisible mob eager and willing to be set off like a dry forest at the touch of a match. And once the forest fire is going, it doesn’t matter if the match was real or fake.
Karl D. Stephan is a professor of electrical engineering at Texas State University in San Marcos, Texas. This article has been republished, with permission, from his blog Engineering Ethics, which is a MercatorNet partner site. His ebook Ethical and Otherwise: Engineering In the Headlines is available in Kindle format and also in the iTunes store.
Sources: Brooke Borel’s article “Clicks, Lies, and Videotape” appeared on pp. 38-43 of the October 2018 issue of Scientific American.